[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Next Article in Journal
Current Capabilities and Challenges of Remote Sensing in Monitoring Freshwater Cyanobacterial Blooms: A Scoping Review
Previous Article in Journal
Remote Sensing Small Object Detection Network Based on Multi-Scale Feature Extraction and Information Fusion
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 5
10.3390/rs17050915
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Classifying Rocky Land Cover Using Random Forest Modeling: Lessons Learned and Potential Applications in Washington, USA

by
Joe V. Celebrezze
1,2,*,
Okikiola M. Alegbeleye
1,
Doug A. Glavich
3,
Lisa A. Shipley
1 and
Arjan J. H. Meddens
1
1
School of the Environment, Washington State University, Pullman, WA 99164, USA
2
Department of Ecology, Evolution, and Marine Biology, University of California, Santa Barbara, CA 93106, USA
3
Region 6 Ecology Program, United States Forest Service, 3200 SW Jefferson Way, Corvallis, OR 97331, USA
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(5), 915; https://doi.org/10.3390/rs17050915
Submission received: 20 December 2024 / Revised: 24 January 2025 / Accepted: 26 February 2025 / Published: 5 March 2025
Browse Figures
Graphical abstract
"> Figure 1
<p>A map of the study area, including delineations for the Cascades (green) and Columbia (blue) regions and their associated sub-regions (italicized, with labels colored according to their region) in Washington, USA. Base map: Bing VirtualEarth.</p> "> Figure 2
<p>Conceptual diagrams conveying our reasoning for formulating multi-temporal predictors based on Sentinel-2 data. (<b>A</b>) The seasonal difference in NDVI was calculated to distinguish rocky land cover from grasses or drought-deciduous shrubs, while (<b>B</b>) NBR interannual metrics were used to distinguish rocky land cover, expected to have low NBR interannual variability, from disturbed landscapes (typically, burned or clearcut forests in the Cascades region), expected to have higher interannual variability in NBR.</p> "> Figure 3
<p>A workflow diagram exhibiting the point selection, attribution, classification, and iterative process utilized to classify rocky land cover for each sub-region.</p> "> Figure 4
<p>Land cover classified as rocky (black) and non-rocky (green). Full random forest models (40 predictors) for sub-regions in the Cascades and Columbia regions of Washington, USA, yielded land cover classification apart from the Portland sub-region, which had limited rocky land cover points (see <a href="#remotesensing-17-00915-t002" class="html-table">Table 2</a>). Base map: Bing VirtualEarth.</p> "> Figure 5
<p>(<b>A</b>) Variable importance of 40 predictors for the Cascades region (large black points) and associated sub-regions (smaller gray points), with boxes colored by predictor group. (<b>B</b>) Variable importance for 40 predictors for the Columbia region (large black points) and associated sub-regions (smaller gray points), with boxes colored by predictor group.</p> "> Figure 6
<p>Segmented regression results for overall accuracy, producer’s accuracy, and user’s accuracy statistics for (<b>A</b>) the Cascades and (<b>B</b>) the Columbia regions. For both regions, the mean breakpoint (vertical dashed line) was between six and seven predictors; thus, the optimized models utilize seven predictors for both regions.</p> "> Figure 7
<p>For optimized models, predictor values and their distributions vary between rocky and non-rocky land cover classifications for the Cascades and Columbia regions. Violin plots display seven ‘optimal’ predictors for each region, ordered by variable importance (see <a href="#app1-remotesensing-17-00915" class="html-app">Figure S2</a> for predictors’ variable importance in optimized models; see <a href="#app1-remotesensing-17-00915" class="html-app">Figure S3</a> for violin plots for all predictors).</p> "> Figure 8
<p>Case Study 1: A rocky patch in the southeast corner of the <span class="html-italic">Forbidden</span> sub-region shown using (<b>A</b>) NAIP imagery via Google Earth, (<b>B</b>) classified rocky land cover using 10 m full models and (<b>C</b>) 1 m resolution NAIP- and topography-based models, and (<b>D</b>) the overlap (purple) between 10 m (red) and 1 m resolution (blue) rocky land cover classification.</p> "> Figure 9
<p>Case Study 2: classification of non-rocky and rocky land covers and various unstable classes, determined by a (<b>A</b>) time series stability analysis for the <span class="html-italic">Snoqualmie</span> sub-region and (<b>B</b>–<b>D</b>) a zoomed-in area in the sub-region. The zoomed-in area shows (<b>B</b>) the stability classes, (<b>C</b>) the natural color image, and (<b>D</b>) the rocky land cover classification without stability classes included.</p> "> Figure 10
<p>Applications of Case Study 2: comparing sub-regions by investigating (<b>A</b>) proportions of classified area fitting in rocky and unstable land cover classifications and (<b>B</b>) their stability ratios, showing that sub-regions in the Cascades region have more stable classifications of rocky land cover than those in the Columbia region. Proportions of non-rocky land cover are not displayed, as their proportions were significantly higher than those of rocky or unstable classifications.</p> "> Figure 11
<p>Case Study 3: comparing rocky habitat classification to WDFW priority habitat maps for talus slopes and cliffs. Maps show the spatial extent of (<b>A</b>) WDFW priority habitat polygons and (<b>B</b>) the rocky land cover classification derived from our random forest modeling approach for the full Cascades region. Selected zoomed-in areas (shown in (<b>A</b>,<b>B</b>) with (<b>C</b>–<b>F</b>) blue and (<b>G</b>–<b>J</b>) red rectangles) exhibit differences in spatial extent and precision of rocky land cover classification, showing (<b>C</b>,<b>G</b>) natural color images, (<b>D</b>,<b>H</b>) WDFW priority habitat polygons, (<b>E</b>,<b>I</b>) our rocky habitat classification, and (<b>F</b>,<b>J</b>) how the two directly compare.</p> ">
Versions Notes

Abstract

:
Rocky land cover provides vital habitat for many different species, including endemic, vulnerable, or threatened plants and animals; thus, various land management organizations prioritize the conservation of rocky habitat. Despite its importance, land cover classification maps rarely classify rocky land cover explicitly, and if they do, they are limited in spatial resolution or extent. Consequently, we used random forest models in Google Earth Engine (GEE) to classify rocky land cover at a high spatial resolution across a broad spatial extent in the Cascade Mountains and Columbia River Gorge in Washington, USA. The spectral indices derived from Sentinel-2 satellite data and NAIP aerial imagery, the specialized multi-temporal predictors formulated using time series of normalized burn ratio (NBR) and normalized difference in vegetation index (NDVI), and topographical predictors were especially important to include in the rocky land cover classification models; however, the predictors’ relative variable importance differed regionally. Beyond evaluating random forest models and developing classification maps of rocky land cover, we conducted three case studies to highlight potential avenues for future work and form connections to land management organizations’ needs. Our replicable approach relies on open-source data and software (GEE), aligns with the goals of land management organizations, and has the potential to be elaborated upon by future research investigating rocky habitats or other rare habitat types.

Graphical Abstract">
Graphical Abstract

1. Introduction

Globally, more than 2000 vertebrate species rely on rocky habitats and approximately 18% of these species have been identified as conservation concerns [1]. Rocky habitat takes many forms, including broken rock or boulders on steep terrain (talus slopes), rock outcroppings, cliffs, block fields, and caves. Broken-rock habitats—accumulations of rock debris or fragments which provide function(s) to at least one species—act as thermal and hydric refuges against extreme cold, heat, or drought [2,3,4]; hiding places that provide concealment from predators; and vantage points (i.e., promontories) that provide opportunities for predator surveillance [5,6]. Additionally, broken-rock habitats harbor important foods, such as diverse forage plants in arid Australia [7] and energy-rich army cutworms (Euxoa auxiliaris) sought after by grizzly bears (Ursus arctos) in the northern Rocky Mountains [8]. Cliffs provide escape terrain for large ungulates like klipspringers (Oreotragus oreotragus; [9]) and Rocky Mountain bighorn sheep (Ovis canadensis; [10]). Furthermore, rocky habitats are critical to federally listed plant species such as Gorman’s aster (Aster gormanii), Howell’s reedgrass (Calamagrostis howellii), and tree clubmoss (Dendrolycopodium dendroideum; [11]). Federal and state agencies recognize the numerous ecological benefits associated with rocky habitats and the importance of conserving them. The United States Forest Service (USFS) identifies rocky habitats as special habitats [12], and the Washington Department of Fish and Wildlife (WDFW) lists talus slopes, cliffs, herbaceous balds, and caves as priority habitats [13].
Despite the recognized value of rocky habitats, efforts to map land cover often fail to explicitly classify rocky land cover, making it challenging to locate patches of rocky habitat, identify their spatial extent, and determine habitat connectivity (i.e., through their spatial distribution). For example, the National Land Cover Dataset (NLCD; [14]) and the ESA WorldCover dataset [15] group rocky land cover with other land cover types (e.g., as ‘barren land’ in NLCD and ‘bare land or sparse vegetation’ in WorldCover). Although the Coordination of Information on the Environment (CORINE) program (European Commission Land Monitoring Service) explicitly classifies ‘bare rock’ in their land cover maps, these data are limited to Europe and only available at 100 m resolution. Limited spatial resolution (also, 30 m in NLCD; [14]) can lead to the imprecise delineation of boundaries of rocky habitat patches and increase the likelihood of a ‘mixed pixel problem’, which decreases the accuracy of resulting land cover classification products [16,17,18]. The limitations of popular products that classify and map land cover—broad land cover categorization (e.g., NLCD and WorldCover) and limited spatial resolution (e.g., NLCD and CORINE)—present key barriers for ecologists or land managers tasked with identifying and studying rocky habitats.
Contrasting with the large spatial extent, relatively low spatial resolution, and broad characterization associated with popular products that classify and map land cover, geomorphologists attempt to classify rocky habitats and emphasize high-resolution, precise classification and in-depth characterization of various rocky habitat classes [19,20]. For example, Schneevoigt et al. [21] investigated a 17 km2 sub-catchment in the Bavarian Alps and categorized the land cover as talus sheets or cones, rockfall (with and without vegetation), alluvial fans, floodplains, debris cones, avalanche deposits, or moraine using remotely sensed imagery. Geomorphology studies offer detailed depictions of rocky landforms; however, modeling efforts are limited in their spatial extent because they are typically case-study-specific and the characterization schemas these studies employ are often too detailed for use by those without a similar background [22]. After reviewing existing products that could potentially be used to identify rocky land cover, various land management organizations (e.g., USFS, WDFW, and National Parks Service) have called for improved methods to classify and map rocky habitats. Ideally, new methods could bridge gaps between geomorphology studies and broad-scale land cover mapping efforts by offering a high-resolution classification of rocky land cover over a large spatial extent.
Land cover classification studies have used different satellite or aerial imagery sources and employed different modeling approaches to reach a similar goal. Data sources include Landsat [14,23], Sentinel-2 multispectral data [24,25], National Agriculture Imagery Program (NAIP) high-resolution imagery [20,26], and Sentinel-1 synthetic aperture radar (SAR) data [27,28]. Frequently, studies classifying land cover pair data from satellite or aerial imagery with topographical data, deriving predictors such as elevation, slope, and aspect from digital elevation models [20,28]. Previous studies have compared the capabilities of different data sources to classify certain land cover types or to identify land use change [29]; however, it remains unclear which data sources and predictors should be included in rocky land cover classification models. Additionally, different modeling approaches have been deployed to classify land cover or model habitat suitability, including maximum entropy modeling (MaxEnt; [30]), support vector machines (SVM; [31,32]), k-nearest neighbor [33], decision trees or boosted decision trees [34,35], deep learning [36,37,38], and random forest modeling [39,40,41,42]. Each modeling approach has strengths and weaknesses; therefore, no universal ‘best’ modeling approach has been adopted [23,43]. In this study, we opted to rely on random forest modeling primarily because this approach is simple and approachable compared to other techniques, such as deep learning models [37], and it has been widely used in other ecological studies [41].
In response to the limitations of current land cover classification products and the calls of various land management organizations, we sought to classify rocky land cover at a high resolution over a broad spatial extent using random forest models in Google Earth Engine (GEE; [44]). Our approach relied on open-source and widely accessible data and software (GEE [44]), so future work can use our approach to identify rocky land cover in different regions—particularly, where different species are identified as conservation concerns. In this study, we ask the following: (1) How do random forest models perform as a classification technique for rocky land cover in a model study area, the Cascade Mountains and the Columbia River Gorge in Washington? (2) Which data sources and predictors are most important to include to ensure accurate classification of rocky land cover? (3) How can rocky land cover maps be applied in future ecological studies or used to identify important habitats and inform conservation or land management practices?

2. Materials and Methods

2.1. Defining Rocky Land Cover

In this study, we broadly defined rocky land cover as an open surface area (e.g., above-ground with no canopy overtopping) dominated by rock. Our classification schema primarily relied on aerial imagery sources (NAIP imagery) to attribute points and run random forest models; therefore, we focused on exposed rocky land cover. We excluded bare soil, gravel, asphalt, and sparsely vegetation vegetated land cover. Additionally, land cover dominated by sand or small pebbles was classified as non-rocky land cover because it often appeared indistinguishable from bare soil in aerial imagery.

2.2. Study Area

To demonstrate the applicability and accuracy of our methods and to investigate our research questions, we classified rocky land cover in (1) the Cascade Mountain Range (Cascades region) and (2) the Columbia River Gorge (Columbia region) in Washington, USA (Figure 1). The Cascades span the center of Washington from north to south and play a large role in climate and species distribution. To the west of the Cascades, precipitation and cloud cover are prevalent, which contrasts with the arid or semi-arid landscapes to the east [45]. This ‘rain shadow’ directly affects species distribution and community composition [46,47]. Likewise, the Columbia River substantially influences species distribution and shapes the surrounding landscape. Winding from the northeastern corner of the state to its delta in the southwestern corner, the Columbia flows through different ecosystems and showcases cliffs, huge sloping valley walls, and tremendous rocky outcroppings [48,49]. Cutting through the Cascades, the Columbia River Gorge is a key part of the landscape in Washington, and its unique geomorphology has been the focus of intense study [50]. Our study area incorporates varied land ownership, mostly consisting of federal lands including national parks (e.g., Mt. Rainier and North Cascades National Parks) and forests (e.g., Gifford-Pinchot, Okanogan-Wenatchee, and Mt. Baker-Snoqualmie National Forests); however, private, state, and tribal lands (e.g., Colville, Snoqualmie, and Yakama reservations) are also present, interspersed between national forests, wilderness, monuments, and parks.
These regions encompass vital rocky habitats for species identified as conservation concerns by management organizations such as the hoary marmot (Marmota caligata), American pika (Ochotona princeps), and Larch Mountain salamander (Plethodon larselli). Studies on American pikas [4], hoary marmots [51], and Larch Mountain salamanders [52] have been conducted in the study area, highlighting potential applications to conservation efforts regarding these species of interest among many others.

2.3. Data Sources and Predictors

To classify rocky land cover using random forest modeling, we used 40 predictors from various data sources (Table 1; see Table A1 for more information, including how we calculated each index). All data are open-source and available through GEE [44]. Pre-processing varied for different data sources; however, we consistently resampled predictors to 10 m resolution, as needed. For data available at resolutions finer than 10 m (e.g., NAIP aerial imagery), we used a nearest neighbor approach. For data with coarser resolutions (e.g., topographical data and certain Sentinel-2 bands; see Table A1), we used bilinear interpolation. Additionally, to reduce image noise, median values were used for Sentinel-2, NAIP, and Sentinel-1 predictors across varying timeframes (see below and Table A1).
From Sentinel-2 satellite imagery, we used 12 spectral reflectance bands and the following spectral indices derived from those bands: normalized difference vegetation index (NDVI), normalized difference burn ratio (NBR), normalized difference built-up index (NBI), modified soil adjusted vegetation index 2 (MSAVI2), enhanced vegetation index (EVI), normalized difference water index (NDWI), automatic water extraction index—shadow (AWEIshadow), and automatic water extraction index—no shadow (AWEIno shadow) (Table 1). We pre-processed the Sentinel-2 data by filtering images containing less than 20% cloud cover and clipping images to our study area. The values for the spectral reflectance bands and associated indices were calculated using median spectral values from 1 July 2023 to 30 September 2023.
From the NAIP aerial imagery, we used four spectral reflectance bands, calculated NDVI, and used NDVI to calculate two texture metrics (Table 1). The texture metrics, NDVImedian, and NDVImaximum, were calculated using median and maximum NDVI values in the center cell within a moving 3 × 3 cell focal window [53]. Previous studies have used texture metrics such as these to capture fine-scale variability across the landscape while reducing photographic contrast variability [53,54]. Because high-resolution NAIP imagery is typically captured biannually, making only one image composite for an area available each year, we incorporated a wider timeframe than Sentinel-2 data, calculating median band values from available images between 1 July 2017 and 30 September 2023.
From Sentinel-1 synthetic aperture radar (SAR), we used the vertical–vertical (VV) and vertical–horizontal (VH) radar retrieval bands. VV transmits and receives vertically polarized signals, and VH transmits vertically polarized signals and receives horizontally polarized signals [55]. Previous studies used Sentinel-1 bands for different applications including characterizing mangrove forest structure [56] and assessing wildfires [57]; however, their connections to rocky land cover are largely unknown. We pre-processed the Sentinel-1 SAR VV and VH bands by filtering images for interferometric wide swath (IW) instrument mode and the ‘descending’ orbit direction and clipping the images to our study area. As a result of preliminary predictor testing, we incorporated more wet months into the SAR images to improve the separation of wetland and dryland; therefore, we filtered images between 1 May 2023 and 30 November 2023.
We used six topographical predictor bands from multiple data sources. We used elevation and slope from the US Geological Survey digital elevation model (DEM). From the Geomorpho90m dataset, we used the topographic ruggedness index (TRI) and the compound topographic index (CTI), a topographic wetness index [58]. From the Conservation Science Partners/Ecologically Relevant Geomorphology (CSP/ERGo) dataset (270 m resolution), we used the continuous heat-insolation load index (CHILI) and the topographic position index (TPI). Other than resampling to 10 m resolution, no pre-processing was necessary for the topographical data.
Aside from the standard remote sensing predictors described above, we formulated new multi-temporal predictors using NDVI and NBR time series to effectively differentiate rocky land cover from non-rocky land cover types commonly misclassified as rocky land cover (Figure 2). First, we used seasonal differences in NDVI (Figure 2A) to better differentiate between rocky habitats and grasses or drought-deciduous shrubs. To isolate rocky habitats, which remain a similar color throughout the year, from vegetation (e.g., grasses or drought-deciduous shrubs) that green up in the early spring, we calculated the difference between the median NDVI in the early fall (1 September to 30 October 2023) and late spring (1 May to 30 June 2023). Second, we used NBR interannual metrics to isolate areas affected by severe disturbances such as wildfires or clearcutting (Figure 2B). To calculate NBR interannual metrics, we relied on a time series of yearly NBR values, determined using Sentinel-2 data collected from the growing season (1 June to 30 October) each year from 2016 to 2023. For NBR maximum variance, we calculated the variance for each growing season and selected the maximum value of variance under the assumption that, for disturbance-impacted pixels, this maximum variance would be high during the year of the disturbance. For the NBR variance of medians and NBR variance of 25th percentiles, we calculated the median and 25th percentile of NBR for each growing season and calculated their variance from 2016 to 2023 under the assumption that disturbance-impacted pixels would have higher interannual variation in NBR (Figure 2B).

2.4. Point Selection and Attribution

To train random forest models, we used a supervised classification approach that entails selecting location reference points for each sub-region in the study area and attributing those points as ‘rocky’ or ‘non-rocky’ (Table 2) using QGIS (QGIS, Version 3.28). We divided the Cascades and Columbia regions into sub-regions to allow for manageable visual assessments of land cover maps and iterative point attribution. We used QGIS and GEE for point selection; attribution; classification; and, importantly, iterative training/testing point addition for each sub-region (Figure 3). To ensure that the selected points spanned the spatial extent of the study area, we randomly generated points. Then, we attributed the random points using imagery available within Google Earth (generally, National Agriculture Imagery Program (NAIP) imagery). In the few cases in which it was difficult or impossible to detect land cover for a randomly sampled point (e.g., in topographically shaded areas or areas covered by clouds), we either moved those points to nearby locations where attribution was possible or removed them entirely. Because rocky land cover is typically sparse and patchily distributed across the landscape, we supplemented random sampling with a selective point selection. In addition to adding rocky points, we attributed points in areas that may be incorrectly classified as rocky land cover (as identified through initial classification runs) as non-rocky. These areas include disturbed forests, developed land, roads, shorelines, and glaciers. Generally, rocky land cover is more prevalent in the montane Cascades region relative to the Columbia region; therefore, we categorized a higher proportion of points in this region as rocky (Table 2). For both the random and the selective point datasets, we ensured that points were at least 40 m away from each other, limiting spatial autocorrelation.
To ensure an accurate finalized classification map, we iteratively selected training and validation points to assess the random forest models’ classification performance visually and quantitatively multiple times (Figure 3). More specifically, after initial random and selective point generation, we ran a random forest model in GEE [44,59], which yielded a classified map for the sub-region (see below, in Section 2.5). Then, we evaluated the classification performance by visually assessing the mapping performance in QGIS and through confusion matrices [60]. Afterward, we added training points to misclassified areas or areas with low point density. To improve mapping performance, we ran at least three iterations of the random forest model, adding training points for each iteration. We stopped adding points and finalized the classification map when classification accuracy statistics stopped improving and visual assessments confirmed that rocky and non-rocky land cover classes were accurately mapped. In the Cascades region, we attributed 39,182 non-rocky and 17,228 rocky points, and in the Columbia region, we attributed 19,982 non-rocky and 4568 rocky points (Table 2).

2.5. Random Forest Modeling

To classify rocky land cover, we used random forest models. Although other classification techniques could have performed similarly to random forest models, we selected random forest models instead of newer techniques, such as deep learning models, primarily because random forest models are more accessible (i.e., they can easily be run using open-source data and on open-source platforms [GEE] with most modern computers) [37], and land management organizations have more experience working with and interpreting random forest model outputs. Furthermore, in preliminary trials, we compared MaxEnt [30] and random forest models, and by visually assessing rocky land cover classification maps produced by each technique, we opted to continue with random forest models. Random forest models involve fitting many classification trees to a dataset and maximizing homogeneity of the input variables between groups (as measured by the Gini index), and they can be applied for both classification and regression problems [39,40,41,42]. Random forest modeling allows for unbiased accuracy statistics and the quantification of variable importance—enabling a comparison of predictors from different data sources [40]. Additionally, random forest models are especially useful for ecological studies because of their robustness to imbalanced datasets, multicollinearity, and overfitting [41].
We classified each sub-region within the Cascades and Columbia regions separately and then built regional models for the Cascades and Columbia regions. We trained the models using a random selection of 80% of attributed points, stratified by land cover, and the remaining 20% was used for validation. The training and testing points were evenly dispersed across the landscape (Figure S1). We assessed model performance using validation metrics—overall accuracy and, for rocky and non-rocky classes, producer’s accuracies, user’s accuracies, and F-scores. We used 500 classification trees in each model, and the number of variables per split was determined by calculating the square root of the number of variables included in the model (i.e., for full models with 40 variables, approximately six variables were included per split) following established norms [39]. Variable importance was determined using training data and the default algorithm in GEE, which assesses how each variable contributes to information gain and sums the decreases in variables across the decision trees [61]. To visualize the differences in variable importance, we scaled and centered variable importance values for each sub-region and region and displayed these scaled values with boxplots for each region and its associated sub-regions.
To optimize the random forest models, predictors were iteratively removed based on their variable importance scores until the models could no longer adequately separate rocky and non-rocky land cover types. Beginning with full models (40 predictors), we iteratively removed the two predictors with the lowest variable importance until eight predictors remained, after which the least important variable was removed until classification accuracy statistics indicated that further predictor omissions would result in drastically reduced mapping accuracy. This iterative process was repeated for the Cascades and Columbia regions, and accuracy statistics were calculated for each iteration. To identify breakpoints for each accuracy statistic indicating when model performance significantly declined, we conducted segmented regressions using R (R Version 4.4.1; package ‘segmented’; [62]) to compare the accuracy statistics and the number of predictors included. To determine the optimal breakpoint, the package ‘segmented’ uses a maximum likelihood estimation process to minimize the residual sum of squares [62]. The predictors included in the model directly before the breakpoint occurred were then included in optimized random forest models for the Cascades and Columbia regions, respectively, and the sub-regions associated with each region. Lastly, we investigated how predictor values varied for rocky and non-rocky land cover types by extracting the values from each predictor for each attributed point. To compare the differences in predictor values, we visualized these data using violin plots, which present data distributions using mirrored kernel density plots, and statistically compared the predictor values across rocky and non-rocky classes with t-tests.

2.6. Case Studies

To showcase broader applications of our classification approach and demonstrate how future studies could potentially improve mapping performance, we conducted three case studies. The case studies involved (1) more precisely classifying boundaries of rocky land cover patches by using a higher spatial resolution, (2) analyzing the temporal stability of our classification results (as in [63]), and (3) comparing our rocky land cover classification maps to rocky priority habitat polygons delineated by the Washington Department of Fish and Wildlife (cliffs and talus slopes). Case studies often involve qualitative interpretation of classification maps rather than quantitative statistical testing.
Case Study 1 demonstrates the potential to improve classified maps’ spatial resolution by using 1 m NAIP data instead of 10 m Sentinel-2-based imagery, investigating a small subset of the study region. When improving the delineation of rocky land cover patches, we relied on 1 m resolution NAIP data, post-processed using a majority filter with a square kernel (9 × 9 pixels) to eliminate noise, and topographical data downscaled to 1 m resolution (see Table 1) in our classification approach.
In Case Study 2, we determined the classification stability [63] for rocky habitats using classification results from 2017, 2019, 2021, and 2023. To limit computation time for each model, 17 predictors were included, incorporating Sentinel-2 bands and indices and multi-temporal and topographical predictors, as determined using variable importance values from sub-regional and regional random forest models. Due to limitations in image acquisition dates and computing time, we did not include predictors calculated from NAIP data. If the rocky classification results were identical at a given pixel location for all four years, we denoted the classification as ‘stable’. Otherwise, we defined varying levels of ‘unstable’ classification when a pixel was classified as rocky land cover only for one, two, or three years. Unstable classifications could indicate misclassification or a change in land cover, due to landslides (from non-rocky to rocky) or revegetation (from rocky to non-rocky). Additionally, we calculated a stability ratio for each sub-region as the proportion of stable rocky area to unstable area. We interpreted sub-regions with lower stability ratio values as more sensitive to yearly variation in predictors included in the random forest modeling.
For Case Study 3, we visually compared our rocky land cover classification maps to polygon shapefiles for rocky Washington Department of Fish and Wildlife (WDFW) priority habitats – specifically talus slopes and cliffs. We juxtaposed WDFW priority habitat polygons to our classified rocky land cover maps and considered differences in spatial extent, precision, and accuracy of rocky land cover/rocky habitat delineation.

3. Results

3.1. Variable Importance

We classified rocky land cover in the Cascades and Columbia regions (Figure 4), assessed variable importance for 40 predictors, and evaluated how predictors’ variable importance differed between regions (Figure 5). For both regions, important predictors included NBR interannual metrics (see Figure 2B), elevation, slope, and topographic roughness. For the Cascades region specifically, the importance of NBR interannual metrics was more substantial than the Columbia region, as three of the four most important variables described NBR interannual trends (Figure 5A). Conversely, slope had a higher importance in the Columbia region relative to the Cascades region (Figure 5B). For both regions, many Sentinel-2 and NAIP spectral reflectance bands had low variable importance. For the Cascades region, Sentinel-1 predictors were less important than other predictor groups; however, the importance of Sentinel-1 predictors in the Columbia region was comparable to NAIP-derived and Sentinel-2-derived predictors.

3.2. Optimizing Random Forest Models

When optimizing the random forest models through the step-wise removal of predictor variables, we determined that classifying rocky land cover was possible with as few as seven spectral, topographical, and multi-temporal predictors, as indicated by the consistent breakpoints for the assessed accuracy statistics (Figure 6) and the predictors included in the optimized models (Figure 7). For both regions, the accuracy statistics remained relatively constant for models including 10–40 predictors, and models that incorporated less than 7 predictors performed increasingly worse upon further exclusion of predictors for both regions (Figure 6). Therefore, we used seven predictors in the optimized models for the Cascades and Columbia regions and their associated sub-regions (see Table 3). Elevation, slope, NBR variance of the 25th percentile, and NBR were incorporated in the optimized models for both regions, but the other three predictors varied between regions. The Cascades models incorporated modified soil adjusted vegetation index 2 (MSAVI2) and red edge 1 and blue Sentinel-2 bands (Figure 7; Figure S2a), while the Columbia models incorporated normalized buildup index (NBI) and near-infrared (NIR) and aerosols Sentinel-2 bands (Figure 7; Figure S2b). Predictor values significantly differed between rocky and non-rocky points (t-test; p < 0.001) for all predictors included in the optimized models for both regions (Figure 7), suggesting that each predictor plays a key role in classifying rocky vs. non-rocky land cover.

3.3. Accuracy Statistics

The overall accuracy was typically above 95% for full models on regional and sub-regional scales (Table 3; see Table S1 for more accuracy statistics). Optimized models had comparable accuracy statistics to full models; however, they tended to be slightly lower (Table 3; Table S1). Classification accuracy was higher for the Cascades region and associated sub-regions, compared to the Columbia region, for the full and optimized models. For example, there was generally a low incidence rate of accuracy statistics less than 0.9, but the Columbia models exhibited low accuracy statistics disproportionately more than Cascades models. For full models, 1.59% (1/63) of accuracy statistics were less than 0.9 for the Cascades models, whereas 28.6% (12/42) of accuracy statistics were less than 0.9 for the Columbia models (Table S1). The contrast was similar for optimized models (i.e., 3.17% of accuracy statistics were less than 0.9 for the Cascades models versus 38.1% for the Columbia models). The accuracy statistics derived from the training data were consistently higher than 0.99 (Table S2).

4. Case Studies

4.1. Case Study 1: Improving Delineation of Rocky Land Cover Patch Boundaries

Upon visually inspecting the 10 m resolution land cover classification map (Figure 4), it became evident that the boundary delineation of rocky land cover patches was imprecise at very fine spatial resolutions (Figure 8A,B). To remedy this, we attributed points at a higher density within a case study area in the southeastern corner of the Forbidden sub-region (Cascades region; see Figure 1), mapped rocky land cover at 1 m resolution, and subsequently post-processed the resulting classification map (Figure 8C). Although the classified rocky habitat for 10 m and 1 m resolution maps largely overlapped, the 1 m resolution maps appeared to improve patch boundary delineation and address both overestimates and underestimates of rocky land cover (Figure 8D; see Table S3 for performance metrics and variable importance). Despite evident improvements in mapping performance in this case study, we do not recommend 1 m resolution mapping for broad spatial applications or extents (i.e., for the entire Cascades region) due to the vastly increased computation time associated with classifying rocky land cover at such a fine scale and the increased time required to attribute a higher density of points. But, for applications which seek finer resolution classification of rocky habitat on relatively small spatial scales, this case study demonstrates that models reliant on NAIP aerial imagery and topographical predictors suffice. Additionally, we attempted to classify rocky land cover at this finer resolution using NAIP-derived predictors alone (i.e., without the addition of topographical predictors); however, mapping performance suffered dramatically (not presented here).

4.2. Case Study 2: Stability Analysis

By iteratively assessing classification maps and attributing training and testing points (Figure 3), we partially corrected for incorrect classification of non-rocky land cover (e.g., disturbed forests, developed land, or bare soil) as rocky. To further account for instances of incorrect classification, we utilized a stability analysis using data from 2017, 2019, 2021, and 2023 to determine and categorize stability of rocky land cover classification (Figure 9). This approach was previously established by Stahl et al. [63] for assessing crop rotations and temporal change in conservation areas. Unlike the 1 m resolution mapping employed in Case Study 1, mapping stability classes did not require further point attribution, nor did it greatly increase computation time. With the stability analysis, we found that classification of the edges of rocky patches was more unstable than the central core of rocky patches (Figure 9A,B).
We identified two primary applications of the stability analysis: (1) stable rocky land cover could be utilized to make conservative estimates of rocky habitat area over wide spatial extents and (2) classification stability could be quantified to compare how mapping performance could vary through time (i.e., stability ratios). To demonstrate these applications, we categorized rocky land cover classification stability, calculated proportions allocated to each category (Figure 10A; Table S4), and quantified stability ratios for each sub-region (Figure 10B). Columbia sub-regions exhibited lower stability ratios than Cascades sub-regions (Figure 10B), providing another line of evidence that rocky land cover classification maps were more accurate and reliable in the Cascades region relative to the Columbia region. Based on visual assessments of stability maps for sub-regions in the Columbia region, we inferred that the lower stability ratio was attributable to occasional misclassification of steep rocky cliffs and talus slopes; incorrect classification of sand, gravel, or grasses and deciduous shrubs as rocky landcover; and, to a lesser extent, annual variability of water height.

4.3. Case Study 3: Comparing Rocky Habitat Classification to Priority Habitat Maps

Because of the conservation concerns for wildlife species that depend on rocky habitats such as the American pika [64,65] and hoary marmot [51], the Washington Department of Fish and Wildlife (WDFW) manually delineates priority habitats such as talus slopes and cliffs [13]. We overlaid our mapping approach with the delineated areas from WDFW to highlight both corresponding and diverting results (Figure 11). By visually assessing rocky WDFW priority habitat maps, we confirm that WDFW has yet to systematically map rocky habitats throughout Washington (Figure 11A) and their classified polygons have imprecise boundaries (Figure 11D,H). In contrast, our rocky habitat classification mapping offers improved spatial extent and spatial precision (Figure 11). As a result, our work will allow agencies and organizations to better tailor mapping products to their needs and the goals described by their priority habitat and species programs. Although a random forest-based approach offers some improvements over current WDFW maps, our classified maps fail to specify priority habitats such as talus slopes or cliffs, instead opting to broadly classify ‘rocky land cover’. To overcome this shortcoming and offer spatially continuous and precise mapping of priority habitats, we could either attribute rocky points to more specific classes (e.g., talus, rocky outcroppings, cliffs, and other rocky habitat) or incorporate additional information, such as lidar remote sensing, to distinguish between specific classes of rocky habitat (e.g., rock size or slope).

5. Discussion

To accurately classify rocky land cover over a large scale, we developed an approach using publicly available data to inform random forest modeling using the open and accessible GEE platform for the Cascades Mountains and Columbia River Gorge in Washington, USA. We showed that our approach was effective for detecting rocky land cover (e.g., accuracies generally >95%) across a large extent. Our mapping products can be used for assessing a variety of unique habitats (e.g., talus slopes), species distribution mapping (e.g., for species identified as conservation concerns such as the American pika and hoary marmot), and monitoring efforts. Furthermore, we showed that we can tailor our rocky land cover classification approach to meet the needs of land management organizations. Additionally, because our approach is simple, replicable, and accessible (e.g., in comparison to other classification techniques such as deep learning models [37]), it could be readily adopted in other regions and improved upon by increasing spatial resolution or accuracy of rocky patch delineation (as in Case Study 1). Lastly, the approach we present can be adopted for other unique land cover types, including other habitats identified as conservation priorities [12].

5.1. Predictors and Optimization

When classifying rocky land cover, multispectral, topographical, and multi-temporal predictors were important for both study regions (Figure 5 and Figure 7). Multispectral indices including NBR, NBI, and MSAVI2 were vital inclusions (Figure 7). These indices broadly represent vegetation greenness or wetness and likely played a substantial role in differentiating rocky land cover from green vegetation. However, they also could have played a role in separating rocky land cover from other land cover classes with similar spectral signatures such as burned forests and developed concrete structures. NBR and NBI both rely on near infrared and short-wave infrared spectral wavelengths and could separate rocky land cover from land cover impacted by wildfire (i.e., NBR was developed primarily to identify burned areas [66]) or developed land cover (i.e., NBI was developed to identify built-up or developed areas [67]). MSAVI2, on the other hand, was developed to improve the detection of sparse vegetation and may have contributed to effectively separating rocky land cover and land cover with high soil cover [68]. Topographical metrics such as elevation and slope were key indicators of rocky land cover, highlighted by their high variable importance (Figure 5) and inclusion in optimized models for both regions (Figure 7). Lastly, our formulated multi-temporal predictors were important inclusions, as indicated by their high variable importance values (Figure 5). Aligning with our conceptual frameworks of NBR interannual metrics (Figure 2B), certain non-rocky pixels had especially high variance of the 25th percentile of NBR, indicating that these pixels may have been impacted by a disturbance (Figure 7).
For certain predictors, we identified regional differences in their relative importance. In the Cascades region, the NBR and NBR interannual metrics were especially important (i.e., high variable importance; Figure 5), likely because large-scale and severe disturbances such as wildfires and clearcutting are more prevalent relative to the Columbia region ([69]; Figure 2B). Moreover, the higher relative importance of NBR as a predictor in Cascades models could have resulted from a greater contrast in greenness between rocky land cover and coniferous forests relative to rocky land cover and the shrub-steppe and grasslands, which commonly dominate the slopes of the Columbia River Gorge [69]. In the Columbia region, seasonal differences in NDVI were especially important (i.e., high variable importance; Figure 5) likely because of the seasonal browning of grasses and shrubs that dominate the valley walls in the Columbia River Gorge ([48]; see Figure 2A). For topographical predictors, elevation was a more important inclusion in the Cascades models (Figure 7), likely because rocky patches were commonly alpine broken-rock habitat, or talus slopes or exposed rocky outcroppings above the tree-line. On the other hand, slope was a better indicator of rocky land cover in the Columbia region (Figure 7), likely because rocky patches were most commonly along steep slopes beneath basalt outcroppings on either side of the Columbia River. By identifying regional differences in relative predictor importance, we emphasize the importance of considering the spatial distribution of rocky land cover patches and ecological dynamics in the surrounding landscape prior to selecting predictors for random forest models.
Future studies seeking to reduce time spent on data preparation and modeling could optimize their modeling effort. Our results show that optimized models including only seven predictors performed similarly to full models with 40 predictors at both the regional and sub-regional scales (e.g., for the regional models, the overall accuracies for the optimized models were 97.4% [Cascades] and 93.9% [Columbia], which were marginally lower than those for the full models, 97.9% [Cascades] and 95.2% [Columbia]; Table 3). We argue that predictors derived from Sentinel-1 data, automated water extraction indices, the compound topographic index, and the topographic position index could be excluded from future works classifying rocky land cover (see Figure 5). Synthetic aperture radar from Sentinel-1 [70] detects surface water and wetlands, the automated water extraction indices detect surface water [71], and the compound topographic index indicates potential soil wetness [72]. It is sensible, therefore, that these data are not strong predictors of rocky land cover. Otherwise, we found that TPI could be excluded with minimal or no effect on mapping performance because other topographical predictors (i.e., slope and elevation) were more important to include.
Although we successfully classified rocky land cover using only seven predictors, we suggest that future modeling efforts should test a variety of predictors before settling on a final optimized model. Because of regional differences in the predictors selected for each optimized model (Figure 7), the most important variables differed depending on the context of the surrounding landscape (Figure 5). Although many Sentinel-2 and NAIP bands had relatively low variable importance, it is necessary to prepare the data for Sentinel-2 and NAIP to calculate essential spectral indices, and certain bands were unexpectedly included in optimized models (e.g., aerosols and red edge 1 Sentinel-2 bands; Figure 7); therefore, we advocate for their inclusion. In sum, we advise any future rocky land cover classification efforts to include certain topographical predictors (e.g., slope and elevation), multispectral indices broadly representing vegetation greenness (e.g., NBR, NDVI, MSAVI2, and NBI), multi-temporal predictors (see Figure 2), and bands from Sentinel-2 and NAIP data.

5.2. Limitations and Next Steps

Our approach to classifying rocky land cover has some limitations; however, future research expanding upon our approach could work towards overcoming these limitations. Primarily, we were limited by what we could identify as rocky habitat with aerial imagery while attributing points. Mossy rocky land cover, rocky land cover beneath forest canopies, and caves were likely classified as non-rocky land cover using our methods because of our inability to detect their presence using aerial imagery. Consequently, certain species identified as conservation concerns that rely on shaded, moist, or vegetated rocky habitats could be overlooked. For example, Larch Mountain salamander is a key species that relies on rocky habitat, but because this lungless salamander requires moist rocky habitats often obscured by vegetation [3,52], we were unlikely to capture their habitat in our rocky land cover classification. Also, we were unable to classify certain WDFW priority rocky habitats (i.e., caves and mossy talus slopes) as rocky land cover because we could not identify their presence with aerial imagery. With further data inputs, such as lidar [73] or ground-penetrating radar [74], and investigation of the landscape (e.g., through field surveys), future studies could potentially overcome these challenges. Otherwise, we identified the availability and accessibility of data and software associated with our approach as a strength; however, NAIP or high-resolution aerial data may be unavailable in certain regions or for certain time frames [26]. We do not foresee this inhibiting future works from applying our approach elsewhere, as predictors derived from Sentinel-2 and topography data were generally more important in our random forest models than NAIP-derived predictors (Figure 5 and Figure 7).
In addition to overcoming the limitations of our approach, future work could build upon these classification maps in various ways. First, rocky land cover classification maps could be used in species distribution modeling [75,76] to identify which rocky habitat patches support species of interest such as the American pika or hoary marmot. By identifying species presence across the region and taking measurements on habitat characteristics, future research could combine field-based measurements with remote sensing data to improve our understanding of the species that rely on rocky habitats. Furthermore, by field-validating other data sources such as photogrammetry or lidar, rocky habitat patches could be characterized [77]; therefore, species distribution models could be refined, allowing for improved guidance for conservation efforts. Additionally, by altering the categorization of rocky land cover during the point selection and attribution phase—for example, into talus slopes, cliffs, rocky outcroppings, and other rocky land covers—future studies could expand upon our work and distinguish between specific rocky land cover classes. Lastly, future work could analyze rocky habitat mapping or species distribution maps and develop habitat connectivity maps [78]. Habitat connectivity is crucial for the conservation of rocky habitat obligate species, which underscores the need for mapping rocky habitat patches [79,80,81,82,83].

6. Conclusions

As the climate continues to change and concerns regarding vulnerable species that depend on rocky habitats continue to rise, it is critical to map and characterize rocky habitats with high resolution and precision across a broad spatial extent. We showcase a random forest modeling approach that fulfills this need and demonstrates its ability to accurately classify rocky land cover in the Cascades Mountains and Columbia River Gorge. Furthermore, through three case studies, we call attention to potential improvements or alterations to our classification approach that could be used in future studies and directly compare our approach with existing maps that agencies and organizations rely upon for the conservation and management of rocky priority habitats. Therefore, our study not only acts as a starting point for future research endeavors designed to conserve rocky habitats but also has direct connections to active conservation efforts such as those in Washington, USA.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/rs17050915/s1, Figure S1: Training and testing points for the Cascades and Columbia regions are evenly dispersed on the landscape; Figure S2: Variable importance boxplots for optimized random forest models; Figure S3: Violin plots for all 40 predictors; Table S1: Accuracy statistics for full and optimized regional and sub-regional random forest models using testing data; Table S2: Accuracy statistics for full and optimized regional and sub-regional random forest models using training data; Table S3: Case Study 1—accuracy statistics and variable importance for the top seven predictors; Table S4: Case Study 3—areas of stability classes for each region and sub-region.

Author Contributions

Conceptualization, A.J.H.M., L.A.S., D.A.G. and J.V.C.; methodology, J.V.C., D.A.G. and A.J.H.M.; software, J.V.C. and D.A.G.; formal analysis, J.V.C. and D.A.G.; data curation, J.V.C. and D.A.G.; writing—original draft preparation, J.V.C.; writing—review and editing, J.V.C., O.M.A., D.A.G., L.A.S. and A.J.H.M.; visualization, J.V.C. and O.M.A.; project administration, A.J.H.M. and L.A.S.; funding acquisition, A.J.H.M. and L.A.S. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by a NASA Biological Diversity & Ecological Conservation Program Award (80NSSC24K0234) and the USDA National Institute of Food and Agriculture (NIFA), McIntire Stennis Project 1019284.

Data Availability Statement

The data presented in this study as well as any code and figures are openly available at https://github.com/celebrezze/classifying-rocky-landcover, accessed on 1 March 2025.

Acknowledgments

We acknowledge Meghan Camp and Reid Camp’s roles in establishing the original ideas which were foundational to the NASA Rocky Habitat project; Keith Folkerts and Jeffrey M. Azerrad from the Washington Department of Fish and Wildlife for providing data and feedback for Case Study 3; and the rest of the NASA Rocky Habitat team (including Amanda T. Stahl, Benjamin Bright, Andrew Hudak, Donald Brown, Ana Torres-Ferreira, Allison M. Stift, Adam Duarte, Tara Chestnut, and Mason P. Mahacek) for comments on our approach and research questions at a work retreat in February 2024.

Conflicts of Interest

The authors declare no conflicts of interest.

Appendix A

Table A1. Descriptions of the 40 predictors included in random forest modeling; additionally, defining acronyms, resolutions, and formulas as needed.
Table A1. Descriptions of the 40 predictors included in random forest modeling; additionally, defining acronyms, resolutions, and formulas as needed.
Data SourceData TypePredictorDescription
Sentinel-2BandsAerosols (B1)Wavelength: 443.9 nm (S2A), 442.3 nm (S2B); Resolution: 60 m
Blue (B2)Wavelength: 496.6 nm (S2A), 492.1 nm (S2B); Resolution: 10 m
Green (B3)Wavelength: 560 nm (S2A), 559 (S2 B); Resolution: 10 m
Red (B4)Wavelength: 664.5 nm (S2A), 665 nm (S2B); Resolution: 10 m
Near Infrared (NIR; B8)Wavelength: 835.1 nm (S2A), 833 nm (S2B); Resolution: 10 m
Red Edge 1 (B5)Wavelength: 703.9 nm (S2A), 703.8 nm (S2B); Resolution: 20 m
Red Edge 2 (B6)Wavelength: 740.2 nm (S2A), 739.1 nm (S2B); Resolution: 20 m
Red Edge 3 (B7)Wavelength: 782.5 nm (S2A), 779.7 nm (S2B); Resolution: 20 m
Water Vapor (B9)Wavelength: 945 nm (S2A), 943.2 nm (S2B); Resolution: 60 m
Cirrus (B10)Wavelength: 1373.5 nm (S2A), 1376.9 nm (S2B); Resolution: 60 m
Short-wave Infrared 1 (SWIR 1; B11)Wavelength: 1613.7 nm (S2A), 1610.4 nm (S2B); Resolution: 20 m
SWIR 2 (B12)Wavelength: 2202.4 nm (S2A), 2185.7 nm (S2B); Resolution: 20 m
IndicesNormalized Difference in
Vegetation Index (NDVI)		Commonly used index of vegetation greenness, calculated as:
N I R − R e d N I R + R e d
Normalized Burn Ratio (NBR)Index intended to identify burned areas and provide a measure of burn severity, calculated as
N I R − S W I R 2 N I R + S W I R 2
Normalized Buildup Index (NBI)Index intended to identify developed areas calculated as
R e d + B l u e − G r e e n R e d + B l u e + G r e e n
Modified Soil-Adjusted
Vegetation Index 2 (MSAVI2; [68])		Index of vegetation greenness adjusted to address limitations of NDVI when applied to areas with exposed soil, calculated as
2 N I R + 1 − 2 × N I R + 1 2 − 8 ( N I R − R e d ) 2
Enhanced Vegetation Index (EVI; [84]) Index of vegetation greenness adjusted to increase sensitivity in densely forested areas, calculated as
2.5 × N I R − R e d N I R + 6 × R e d − 7.5 × B l u e + 1
Normalized Difference Water
Index (NDWI; [85])		Index which highlights water and moisture, calculated as
G r e e n − N I R G r e e n + N I R
Automated Water Extraction
Index, no shadows
(AWEIno shadows; [71])		Index intended to identify surface water in areas where shadows are not an issue, calculated as
4 G r e e n − S W I R 1 − ( 0.25 × N I R + 2.75 × S W I R 2 )
Automated Water Extraction
Index, shadows
(AWEIshadow; [71])		Index intended to identify surface area in areas where shadows might be an issue, calculated as
B l u e + 2.5 ( G r e e n ) − 1.5 N I R + S W I R 1 − 0.25 ( S W I R 2 )
Multi-
Temporal		Seasonal Difference in NDVIA predictor we formulated for this study using monthly values of NDVI (see Figure 2A)
NBR Maximum VarianceInterannual NBR predictors we formulated for this study using yearly values of NBR (see Figure 2B)
NBR Variance of Medians
NBR Variance of 25th Percentiles
NAIPBandsBlue (B)Wavelength range: 420–492 nm; Resolution: 60 cm−1 m
Green (G)Wavelength range: 533–587 nm; Resolution: 60 cm−1 m
Red (R)Wavelength range: 604–664 nm; Resolution: 60 cm−1 m
Near Infrared (NIR; N)Wavelength range: 604–664 nm; Resolution: 60 cm−1 m
IndexNDVISee above
TextureNDVImedian textureTexture derived from a moving 3 × 3 cell window, calculating median NDVI values from a NAIP 1 m cell
NDVImaximum textureTexture derived from a moving 3 × 3 cell window, calculating maximum NDVI values from a NAIP 1 m cell
Sentinel-1 (SAR)BandsVertical–Vertical (VV)Transmission of vertically polarized light, reception of vertically polarized light
Vertical–Horizontal (VH)Transmission of vertically polarized light, reception of horizontally polarized light
TopographyMetricsElevationExtracted from the 10 m 3DEP digital elevation model (USGS)
SlopeExtracted using ee.Terrain.slope from the national elevation dataset (NED, USGS)
IndicesTerrain Ruggedness Index (TRI)Calculated using the SRTM digital elevation model (USGS) by using a convolution with a 1-pixel-radius kernel square neighborhood; used bilinear interpolation to resample to 10 m resolution
Topographic Position Index (TPI)Extracted from the Ecologically Relevant Geomorphology (ERGo) Datasets (Conservation Science Partners) to parse valleys and ridgetops; 270 m native resolution, bilinear interpolation to 10 m
Continuous Heat-Insolation Load Index (CHILI)Extracted from the Ecologically Relevant Geomorphology (ERGo) Datasets (Conservation Science Partners) to quantify topographic shading and insolation and their impacts on transpiration; 90 m resolution, bilinear interpolation to 10 m
Topographic Roughness [58]Extracted from the Geomorpho90m Dataset to describe ruggedness and topographic complexity; 90 m native resolution, bilinear interpolation to 10 m
Compound Topographic Index (CTI; [58])Extracted from the Geomorpho90m Dataset to describe topographic relief; 90 m native resolution, bilinear interpolation to 10 m

References

  1. IUCN The IUCN Red List of Threatened Species. Version 2023-1. 2023. Available online: https://www.iucnredlist.org (accessed on 25 February 2025).
  2. García, M.B.; Domingo, D.; Pizarro, M.; Font, X.; Gómez, D.; Ehrlén, J. Rocky Habitats as Microclimatic Refuges for Biodiversity. A Close-up Thermal Approach. Environ. Exp. Bot. 2020, 170, 103886. [Google Scholar] [CrossRef]
  3. Herrington, R.E. Talus Use by Amphibians and Reptiles in the Pacific Northwest. In General Technical Report RM-166: Management of Amphibians, Reptiles, and Small Mammals in North America: Proceedings of the Symposium; USDA Forest Service Rocky Mountain Forest and Range Experimment Station: Fort Collins, CO, USA, 1988; pp. 216–221. [Google Scholar]
  4. Moyer-Horner, L.; Mathewson, P.D.; Jones, G.M.; Kearney, M.R.; Porter, W.P. Modeling Behavioral Thermoregulation in a Climate Change Sentinel. Ecol. Evol. 2015, 5, 5810–5822. [Google Scholar] [CrossRef] [PubMed]
  5. Mares, M.A.; Lacher, T.E. Ecological, Morphological, and Behavioral Convergence in Rock-Dwelling Mammals. In Current Mammalogy; Genoways, H.H., Ed.; Springer: Boston, MA, USA, 1987; pp. 307–348. ISBN 978-1-4757-9911-8. [Google Scholar]
  6. Tyser, R.W. Use of Substrate for Surveillance Behaviors in a Community of Talus Slope Mammals. Am. Midl. Nat. 1980, 104, 32. [Google Scholar] [CrossRef]
  7. Freeland, W.J.; Winter, J.W.; Raskin, S. Australian Rock-Mammals: A Phenomenon of the Seasonally Dry Tropics. Biotropica 1988, 20, 70. [Google Scholar] [CrossRef]
  8. White, D.; Kendall, K.C.; Picton, H.D. Grizzly Bear Feeding Activity at Alpine Army Cutworm Moth Aggregation Sites in Northwest Montana. Can. J. Zool. 1998, 76, 221–227. [Google Scholar] [CrossRef]
  9. Norton, P.M. The Habitat and Feeding Ecology of the Klipspringer Oreotragus Oreotragus (Zimmermann, 1783) in Two Areas of the Cape Province; University of Pretoria: Pretoria, South Africa, 1980. [Google Scholar]
  10. McKinney, T.; Boe, S.R.; de Vos, J.C. GIS-Based Evaluation of Escape Terrain and Desert Bighorn Sheep Populations in Arizona. Wildl. Soc. Bull. 2003, 31, 1229–1236. [Google Scholar]
  11. Fertig, W. Status of Federally Listed Plant Taxa in Washington State. In Washington Natural Heritage Program; Washington Department of Natural Resources: Olympia, WA, USA, 2021. [Google Scholar]
  12. Glavich, D.G.; Stratton, A. Fine-Scaled Special Habitat Mapping: Remote Sensing with Satellite Imagery and Photo-Interpreted Model Training Data (December 2024 Draft); USDA Forest Service Report in Preparation; Pacific Northwest Region; US Forest Service: Portland, OR, USA, in prep.
  13. Washington Department of Fish and Wildlife. Priority Habitat and Species List. Olympia, WA. 2008. Available online: https://wdfw.wa.gov/sites/default/files/publications/00165/wdfw00165.pdf (accessed on 25 February 2025).
  14. Homer, C.; Dewitz, J.; Yang, L.; Jin, S.; Danielson, P.; Coulston, J.; Herold, N.; Wickham, J.; Megown, K. Completion of the 2011 National Land Cover Database for the Conterminous United States–Representing a Decade of Land Cover Change Information. Photogramm. Eng. Remote Sens. 2015, 81, 346–354. [Google Scholar]
  15. Zanaga, D.; Van De Kerchove, R.; Daems, D.; De Keersmaecker, W.; Brockmann, C.; Kirches, G.; Wevers, J.; Cartus, O.; Santoro, M.; Fritz, S.; et al. ESA WorldCover 10m 2021 v200; 2022. Available online: https://worldcover2021.esa.int/documentation (accessed on 25 February 2025).
  16. Foody, G.M.; Cox, D.P. Sub-Pixel Land Cover Composition Estimation Using a Linear Mixture Model and Fuzzy Membership Functions. Int. J. Remote Sens. 1994, 15, 619–631. [Google Scholar] [CrossRef]
  17. Lillesand, T.M.; Kiefer, R.W.; Chipman, J.W. Remote Sensing and Image Interpretation, 7th ed.; John Wiley & Sons: Hoboken, NJ, USA, 2015. [Google Scholar]
  18. Foschi, P.G. A Geometric Approach to a Mixed Pixel Problem: Detecting Subpixel Woody Vegetation. Remote Sens. Environ. 1994, 50, 317–327. [Google Scholar] [CrossRef]
  19. Smith, M.J.; Pain, C.F. Applications of Remote Sensing in Geomorphology. Prog. Phys. Geogr. Earth Environ. 2009, 33, 568–582. [Google Scholar] [CrossRef]
  20. Smith, J.; Heinemeyer, K. Modeling Talus Habitat Using NAIP Imagery and Topographic Features in Portions of Idaho, Montana and Wyoming; Report to Idaho Department Fish and Game, Round River Conservation Studies: Bozeman, MT, USA, 2017. [Google Scholar]
  21. Schneevoigt, N.J.; Van Der Linden, S.; Thamm, H.-P.; Schrott, L. Detecting Alpine Landforms from Remotely Sensed Imagery. A Pilot Study in the Bavarian Alps. Geomorphology 2008, 93, 104–119. [Google Scholar] [CrossRef]
  22. Stead, D.; Donati, D.; Wolter, A.; Sturzenegger, M. Application of Remote Sensing to the Investigation of Rock Slopes: Experience Gained and Lessons Learned. ISPRS Int. J. Geo-Inf. 2019, 8, 296. [Google Scholar] [CrossRef]
  23. Gómez, C.; White, J.C.; Wulder, M.A. Optical Remotely Sensed Time Series Data for Land Cover Classification: A Review. ISPRS J. Photogramm. Remote Sens. 2016, 116, 55–72. [Google Scholar] [CrossRef]
  24. Phiri, D.; Simwanda, M.; Salekin, S.; Nyirenda, V.; Murayama, Y.; Ranagalage, M. Sentinel-2 Data for Land Cover/Use Mapping: A Review. Remote Sens. 2020, 12, 2291. [Google Scholar] [CrossRef]
  25. Thanh Noi, P.; Kappas, M. Comparison of Random Forest, k-Nearest Neighbor, and Support Vector Machine Classifiers for Land Cover Classification Using Sentinel-2 Imagery. Sensors 2017, 18, 18. [Google Scholar] [CrossRef]
  26. Maxwell, A.E.; Warner, T.A.; Vanderbilt, B.C.; Ramezan, C.A. Land Cover Classification and Feature Extraction from National Agriculture Imagery Program (NAIP) Orthoimagery: A Review. Photogramm. Eng. Remote Sens. 2017, 83, 737–747. [Google Scholar] [CrossRef]
  27. Abdikan, S.; Sanli, F.B.; Ustuner, M.; Calò, F. Land Cover Mapping Using Sentinel-1 SAR Data. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2016, XLI-B7, 757–761. [Google Scholar] [CrossRef]
  28. Balzter, H.; Cole, B.; Thiel, C.; Schmullius, C. Mapping CORINE Land Cover from Sentinel-1A SAR and SRTM Digital Elevation Model Data Using Random Forests. Remote Sens. 2015, 7, 14876–14898. [Google Scholar] [CrossRef]
  29. Holtgrave, A.-K.; Röder, N.; Ackermann, A.; Erasmi, S.; Kleinschmit, B. Comparing Sentinel-1 and -2 Data and Indices for Agricultural Land Use Monitoring. Remote Sens. 2020, 12, 2919. [Google Scholar] [CrossRef]
  30. Phillips, S.J.; Anderson, R.P.; Schapire, R.E. Maximum Entropy Modeling of Species Geographic Distributions. Ecol. Model. 2006, 190, 231–259. [Google Scholar] [CrossRef]
  31. Drake, J.M.; Randin, C.; Guisan, A. Modelling Ecological Niches with Support Vector Machines. J. Appl. Ecol. 2006, 43, 424–432. [Google Scholar] [CrossRef]
  32. Pouteau, R.; Meyer, J.-Y.; Taputuarai, R.; Stoll, B. Support Vector Machines to Map Rare and Endangered Native Plants in Pacific Islands Forests. Ecol. Inform. 2012, 9, 37–46. [Google Scholar] [CrossRef]
  33. Mucherino, A.; Papajorgji, P.J.; Pardalos, P.M. Chapter 4. k-Nearest Neighbor Classification (in Data Mining in Agriculture); Springer Optimization and Its Applications; Springer: New York, NY, USA, 2009; Volume 34, ISBN 978-0-387-88614-5. [Google Scholar]
  34. Gislason, P.O.; Benediktsson, J.A.; Sveinsson, J.R. Random Forests for Land Cover Classification. Pattern Recognit. Lett. 2006, 27, 294–300. [Google Scholar] [CrossRef]
  35. Barnard, D.M.; Germino, M.J.; Pilliod, D.S.; Arkle, R.S.; Applestein, C.; Davidson, B.E.; Fisk, M.R. Cannot See the Random Forest for the Decision Trees: Selecting Predictive Models for Restoration Ecology. Restor. Ecol. 2019, 27, 1053–1063. [Google Scholar] [CrossRef]
  36. Alhassan, V.; Henry, C.; Ramanna, S.; Storie, C. A Deep Learning Framework for Land-Use/Land-Cover Mapping and Analysis Using Multispectral Satellite Imagery. Neural Comput. Appl. 2020, 32, 8529–8544. [Google Scholar] [CrossRef]
  37. Petliak, H.; Cerovski-Darriau, C.; Zaliva, V.; Stock, J. Where’s the Rock: Using Convolutional Neural Networks to Improve Land Cover Classification. Remote Sens. 2019, 11, 2211. [Google Scholar] [CrossRef]
  38. Tejasree, G.; Agilandeeswari, L. Land Use/Land Cover (LULC) Classification Using Deep-LSTM for Hyperspectral Images. Egypt. J. Remote Sens. Space Sci. 2024, 27, 52–68. [Google Scholar] [CrossRef]
  39. Belgiu, M.; Drăguţ, L. Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  40. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  41. Cutler, D.R.; Edwards, T.C.; Beard, K.H.; Cutler, A.; Hess, K.T.; Gibson, J.; Lawler, J.J. Random Forests for Classification in Ecology. Ecology 2007, 88, 2783–2792. [Google Scholar] [CrossRef]
  42. Rodriguez-Galiano, V.F.; Ghimire, B.; Rogan, J.; Chica-Olmo, M.; Rigol-Sanchez, J.P. An Assessment of the Effectiveness of a Random Forest Classifier for Land-Cover Classification. ISPRS J. Photogramm. Remote Sens. 2012, 67, 93–104. [Google Scholar] [CrossRef]
  43. Maxwell, A.E.; Warner, T.A.; Fang, F. Implementation of Machine-Learning Classification in Remote Sensing: An Applied Review. Int. J. Remote Sens. 2018, 39, 2784–2817. [Google Scholar] [CrossRef]
  44. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-Scale Geospatial Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [Google Scholar] [CrossRef]
  45. Telker, M.J. The Cascade Mountains. J. Geogr. 1915, 13, 242–250. [Google Scholar] [CrossRef]
  46. Franklin, J.F. Vegetation and Soils in the Subalpine Forests of the Southern Washington Cascade Range. Ph.D. Thesis, Washington State University, Pullman, WA, USA, 1966. [Google Scholar]
  47. Siler, N.; Roe, G.; Durran, D. On the Dynamical Causes of Variability in the Rain-Shadow Effect: A Case Study of the Washington Cascades. J. Hydrometeorol. 2013, 14, 122–139. [Google Scholar] [CrossRef]
  48. Detling, L.E. Peculiarities of the Columbia River Gorge Flora. Madroño 1958, 14, 160–172. [Google Scholar]
  49. Halofsky, J.E.; Peterson, D.L.; Gravenmier, R.A. Climate Change Vulnerability and Adaptation in the Columbia River Gorge National Scenic Area, Mount Hood National Forest, and Willamette National Forest; U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station: Portland, OR, USA, 2022; p. PNW-GTR-1001. [Google Scholar]
  50. Klema, N.; Karlstrom, L.; Cannon, C.; Jiang, C.; O’Connor, J.; Wells, R.; Schmandt, B. The Magmatic Origin of the Columbia River Gorge, USA. Sci. Adv. 2023, 9, eadj3357. [Google Scholar] [CrossRef]
  51. Johnston, A.N.; Christophersen, R.G.; Beever, E.A.; Ransom, J.I. Freezing in a Warming Climate: Marked Declines of a Subnivean Hibernator after a Snow Drought. Ecol. Evol. 2021, 11, 1264–1279. [Google Scholar] [CrossRef]
  52. Herrington, R.E.; Larsen, J.H. Current Status, Habitat Requirements and Management of the Larch Mountain Salamander Plethodon Larselli Burns. Biol. Conserv. 1985, 34, 169–179. [Google Scholar] [CrossRef]
  53. Nielsen, E.M.; Noone, M.D. Tree Cover Mapping for Assessing Greater Sage-Grouse Habitat in Eastern Oregon; Institute for Natural Resources Publications: Corvalis, OR, USA, 2014; Volume 14. [Google Scholar]
  54. Henderson, E.B.; Bell, D.M.; Gregory, M.J. Vegetation Mapping to Support Greater Sage-grouse Habitat Monitoring and Management: Multi- or Univariate Approach? Ecosphere 2019, 10, e02838. [Google Scholar] [CrossRef]
  55. Huang, W.; DeVries, B.; Huang, C.; Lang, M.W.; Jones, J.W.; Creed, I.F.; Carroll, M.L. Automated Extraction of Surface Water Extent from Sentinel-1 Data. Remote Sens. 2018, 10, 797. [Google Scholar] [CrossRef]
  56. Castillo, J.A.A.; Apan, A.A.; Maraseni, T.N.; Salmo, S.G. Estimation and Mapping of Above-Ground Biomass of Mangrove Forests and Their Replacement Land Uses in the Philippines Using Sentinel Imagery. ISPRS J. Photogramm. Remote Sens. 2017, 134, 70–85. [Google Scholar] [CrossRef]
  57. Colson, D.; Petropoulos, G.P.; Ferentinos, K.P. Exploring the Potential of Sentinels-1 & 2 of the Copernicus Mission in Support of Rapid and Cost-Effective Wildfire Assessment. Int. J. Appl. Earth Obs. Geoinf. 2018, 73, 262–276. [Google Scholar] [CrossRef]
  58. Amatulli, G.; McInerney, D.; Sethi, T.; Strobl, P.; Domisch, S. Geomorpho90m, Empirical Evaluation and Accuracy Assessment of Global High-Resolution Geomorphometric Layers. Sci. Data 2020, 7, 162. [Google Scholar] [CrossRef]
  59. Phan, T.N.; Kuch, V.; Lehnert, L.W. Land Cover Classification Using Google Earth Engine and Random Forest Classifier—The Role of Image Composition. Remote Sens. 2020, 12, 2411. [Google Scholar] [CrossRef]
  60. Congalton, R.G. A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data. Remote Sens. Environ. 1991, 37, 35–46. [Google Scholar] [CrossRef]
  61. Haifeng, L. Random Forest Classification (Smile GitHub Repository). Available online: https://github.com/haifengl/smile/blob/master/core/src/main/java/smile/classification/RandomForest.java (accessed on 25 February 2025).
  62. Muggeo, V. Regression Models with Break-Points/Change-Points Estimation (with Possibly Random Effects). R J. 2024, 2.1-0, 49–56. [Google Scholar]
  63. Stahl, A.T.; Fremier, A.K.; Heinse, L. Cloud-Based Environmental Monitoring to Streamline Remote Sensing Analysis for Biologists. BioScience 2021, 71, 1249–1260. [Google Scholar] [CrossRef]
  64. Smith, A.T. Conservation Status of American Pikas (Ochotona princeps). J. Mammal. 2020, 101, 1466–1488. [Google Scholar] [CrossRef]
  65. Broadbooks, H.E. Ecology and Distribution of the Pikas of Washington and Alaska. Am. Midl. Nat. 1965, 73, 299. [Google Scholar] [CrossRef]
  66. Key, C.H.; Benson, N.C. Landscape Assessment: Sampling and Analysis Methods; USDA Forest Service General Technical Report; US Forest Service Rocky Mountain Research Station: Fort Collins, CO, USA, 2006; RMRS-GRT-164-CD. [Google Scholar]
  67. Zha, Y.; Gao, J.; Ni, S. Use of Normalized Difference Built-up Index in Automatically Mapping Urban Areas from TM Imagery. Int. J. Remote Sens. 2003, 24, 583–594. [Google Scholar] [CrossRef]
  68. Qi, J.; Chehbouni, A.; Huete, A.R.; Kerr, Y.H.; Sorooshian, S. A Modified Soil Adjusted Vegetation Index. Remote Sens. Environ. 1994, 48, 119–126. [Google Scholar] [CrossRef]
  69. Franklin, J.F.; Dyrness, C.T. Vegetation of Oregon & Washington; United States Forest Service, Pacific Northwest Forest and Range Experiment Station: Portland, OR, USA, 1969. [Google Scholar]
  70. Zhang, B.; Wdowinski, S.; Gann, D.; Hong, S.-H.; Sah, J. Spatiotemporal Variations of Wetland Backscatter: The Role of Water Depth and Vegetation Characteristics in Sentinel-1 Dual-Polarization SAR Observations. Remote Sens. Environ. 2022, 270, 112864. [Google Scholar] [CrossRef]
  71. Feyisa, G.L.; Meilby, H.; Fensholt, R.; Proud, S.R. Automated Water Extraction Index: A New Technique for Surface Water Mapping Using Landsat Imagery. Remote Sens. Environ. 2014, 140, 23–35. [Google Scholar] [CrossRef]
  72. Chaplot, V.; Walter, C. Subsurface Topography to Enhance the Prediction of the Spatial Distribution of Soil Wetness. Hydrol. Process. 2003, 17, 2567–2580. [Google Scholar] [CrossRef]
  73. Wehr, A.; Lohr, U. Airborne Laser Scanning—An Introduction and Overview. ISPRS J. Photogramm. Remote Sens. 1999, 54, 68–82. [Google Scholar] [CrossRef]
  74. Daniels, D.J.; Gunton, D.J.; Scott, H.F. Introduction to Subsurface Radar. IEE Proc. F Commun. Radar Signal Process. 1988, 135, 278. [Google Scholar] [CrossRef]
  75. Guisan, A.; Zimmermann, N.E. Predictive Habitat Distribution Models in Ecology. Ecol. Model. 2000, 135, 147–186. [Google Scholar] [CrossRef]
  76. Elith, J.; Leathwick, J.R. Species Distribution Models: Ecological Explanation and Prediction Across Space and Time. Annu. Rev. Ecol. Evol. Syst. 2009, 40, 677–697. [Google Scholar] [CrossRef]
  77. Westoby, M.J.; Brasington, J.; Glasser, N.F.; Hambrey, M.J.; Reynolds, J.M. ‘Structure-from-Motion’ Photogrammetry: A Low-Cost, Effective Tool for Geoscience Applications. Geomorphology 2012, 179, 300–314. [Google Scholar] [CrossRef]
  78. Rudnick, D.A.; Ryan, S.J.; Beier, P.; Cushman, S.A.; Dieffenbach, F.; Epps, C.W.; Gerber, L.R.; Hartter, J.; Jenness, J.S.; Kintsch, J.; et al. The Role of Landscape Connectivity in Planning and Implementing Conservation and Restoration Priorities. Issues Ecol. 2012, 16, 1–20. [Google Scholar]
  79. Kuipers, K.J.J.; Hilbers, J.P.; Garcia-Ulloa, J.; Graae, B.J.; May, R.; Verones, F.; Huijbregts, M.A.J.; Schipper, A.M. Habitat Fragmentation Amplifies Threats from Habitat Loss to Mammal Diversity across the World’s Terrestrial Ecoregions. One Earth 2021, 4, 1505–1513. [Google Scholar] [CrossRef]
  80. Fitzsimons, J.A.; Michael, D.R. Rocky Outcrops: A Hard Road in the Conservation of Critical Habitats. Biol. Conserv. 2017, 211, 36–44. [Google Scholar] [CrossRef]
  81. Michael, D.R.; Lindenmayer, D.B.; Cunningham, R.B. Managing Rock Outcrops to Improve Biodiversity Conservation in Australian Agricultural Landscapes. Ecol. Manag. Restor. 2010, 11, 43–50. [Google Scholar] [CrossRef]
  82. Stewart, J.A.E.; Wright, D.H.; Heckman, K.A. Apparent Climate-Mediated Loss and Fragmentation of Core Habitat of the American Pika in the Northern Sierra Nevada, California, USA. PLoS ONE 2017, 12, e0181834. [Google Scholar] [CrossRef]
  83. Peacock, M.M.; Smith, A.T. The Effect of Habitat Fragmentation on Dispersal Patterns, Mating Behavior, and Genetic Variation in a Pika (Ochotona princeps) Metapopulation. Oecologia 1997, 112, 524–533. [Google Scholar] [CrossRef]
  84. Ahmad, F. Spectral Vegetation Indices Performance Evaluated for Cholistan Desert. J. Geogr. Reg. Plann. 2012, 5, 165–172. [Google Scholar] [CrossRef]
  85. Xu, H. Modification of Normalised Difference Water Index (NDWI) to Enhance Open Water Features in Remotely Sensed Imagery. Int. J. Remote Sens. 2006, 27, 3025–3033. [Google Scholar] [CrossRef]
Remotesensing 17 00915 g001
Figure 1. A map of the study area, including delineations for the Cascades (green) and Columbia (blue) regions and their associated sub-regions (italicized, with labels colored according to their region) in Washington, USA. Base map: Bing VirtualEarth.
Figure 1. A map of the study area, including delineations for the Cascades (green) and Columbia (blue) regions and their associated sub-regions (italicized, with labels colored according to their region) in Washington, USA. Base map: Bing VirtualEarth.
Remotesensing 17 00915 g001
Remotesensing 17 00915 g002
Figure 2. Conceptual diagrams conveying our reasoning for formulating multi-temporal predictors based on Sentinel-2 data. (A) The seasonal difference in NDVI was calculated to distinguish rocky land cover from grasses or drought-deciduous shrubs, while (B) NBR interannual metrics were used to distinguish rocky land cover, expected to have low NBR interannual variability, from disturbed landscapes (typically, burned or clearcut forests in the Cascades region), expected to have higher interannual variability in NBR.
Figure 2. Conceptual diagrams conveying our reasoning for formulating multi-temporal predictors based on Sentinel-2 data. (A) The seasonal difference in NDVI was calculated to distinguish rocky land cover from grasses or drought-deciduous shrubs, while (B) NBR interannual metrics were used to distinguish rocky land cover, expected to have low NBR interannual variability, from disturbed landscapes (typically, burned or clearcut forests in the Cascades region), expected to have higher interannual variability in NBR.
Remotesensing 17 00915 g002
Remotesensing 17 00915 g003
Figure 3. A workflow diagram exhibiting the point selection, attribution, classification, and iterative process utilized to classify rocky land cover for each sub-region.
Figure 3. A workflow diagram exhibiting the point selection, attribution, classification, and iterative process utilized to classify rocky land cover for each sub-region.
Remotesensing 17 00915 g003
Remotesensing 17 00915 g004
Figure 4. Land cover classified as rocky (black) and non-rocky (green). Full random forest models (40 predictors) for sub-regions in the Cascades and Columbia regions of Washington, USA, yielded land cover classification apart from the Portland sub-region, which had limited rocky land cover points (see Table 2). Base map: Bing VirtualEarth.
Figure 4. Land cover classified as rocky (black) and non-rocky (green). Full random forest models (40 predictors) for sub-regions in the Cascades and Columbia regions of Washington, USA, yielded land cover classification apart from the Portland sub-region, which had limited rocky land cover points (see Table 2). Base map: Bing VirtualEarth.
Remotesensing 17 00915 g004
Remotesensing 17 00915 g005
Figure 5. (A) Variable importance of 40 predictors for the Cascades region (large black points) and associated sub-regions (smaller gray points), with boxes colored by predictor group. (B) Variable importance for 40 predictors for the Columbia region (large black points) and associated sub-regions (smaller gray points), with boxes colored by predictor group.
Figure 5. (A) Variable importance of 40 predictors for the Cascades region (large black points) and associated sub-regions (smaller gray points), with boxes colored by predictor group. (B) Variable importance for 40 predictors for the Columbia region (large black points) and associated sub-regions (smaller gray points), with boxes colored by predictor group.
Remotesensing 17 00915 g005
Remotesensing 17 00915 g006
Figure 6. Segmented regression results for overall accuracy, producer’s accuracy, and user’s accuracy statistics for (A) the Cascades and (B) the Columbia regions. For both regions, the mean breakpoint (vertical dashed line) was between six and seven predictors; thus, the optimized models utilize seven predictors for both regions.
Figure 6. Segmented regression results for overall accuracy, producer’s accuracy, and user’s accuracy statistics for (A) the Cascades and (B) the Columbia regions. For both regions, the mean breakpoint (vertical dashed line) was between six and seven predictors; thus, the optimized models utilize seven predictors for both regions.
Remotesensing 17 00915 g006
Remotesensing 17 00915 g007
Figure 7. For optimized models, predictor values and their distributions vary between rocky and non-rocky land cover classifications for the Cascades and Columbia regions. Violin plots display seven ‘optimal’ predictors for each region, ordered by variable importance (see Figure S2 for predictors’ variable importance in optimized models; see Figure S3 for violin plots for all predictors).
Figure 7. For optimized models, predictor values and their distributions vary between rocky and non-rocky land cover classifications for the Cascades and Columbia regions. Violin plots display seven ‘optimal’ predictors for each region, ordered by variable importance (see Figure S2 for predictors’ variable importance in optimized models; see Figure S3 for violin plots for all predictors).
Remotesensing 17 00915 g007
Remotesensing 17 00915 g008
Figure 8. Case Study 1: A rocky patch in the southeast corner of the Forbidden sub-region shown using (A) NAIP imagery via Google Earth, (B) classified rocky land cover using 10 m full models and (C) 1 m resolution NAIP- and topography-based models, and (D) the overlap (purple) between 10 m (red) and 1 m resolution (blue) rocky land cover classification.
Figure 8. Case Study 1: A rocky patch in the southeast corner of the Forbidden sub-region shown using (A) NAIP imagery via Google Earth, (B) classified rocky land cover using 10 m full models and (C) 1 m resolution NAIP- and topography-based models, and (D) the overlap (purple) between 10 m (red) and 1 m resolution (blue) rocky land cover classification.
Remotesensing 17 00915 g008
Remotesensing 17 00915 g009
Figure 9. Case Study 2: classification of non-rocky and rocky land covers and various unstable classes, determined by a (A) time series stability analysis for the Snoqualmie sub-region and (B–D) a zoomed-in area in the sub-region. The zoomed-in area shows (B) the stability classes, (C) the natural color image, and (D) the rocky land cover classification without stability classes included.
Figure 9. Case Study 2: classification of non-rocky and rocky land covers and various unstable classes, determined by a (A) time series stability analysis for the Snoqualmie sub-region and (B–D) a zoomed-in area in the sub-region. The zoomed-in area shows (B) the stability classes, (C) the natural color image, and (D) the rocky land cover classification without stability classes included.
Remotesensing 17 00915 g009
Remotesensing 17 00915 g010
Figure 10. Applications of Case Study 2: comparing sub-regions by investigating (A) proportions of classified area fitting in rocky and unstable land cover classifications and (B) their stability ratios, showing that sub-regions in the Cascades region have more stable classifications of rocky land cover than those in the Columbia region. Proportions of non-rocky land cover are not displayed, as their proportions were significantly higher than those of rocky or unstable classifications.
Figure 10. Applications of Case Study 2: comparing sub-regions by investigating (A) proportions of classified area fitting in rocky and unstable land cover classifications and (B) their stability ratios, showing that sub-regions in the Cascades region have more stable classifications of rocky land cover than those in the Columbia region. Proportions of non-rocky land cover are not displayed, as their proportions were significantly higher than those of rocky or unstable classifications.
Remotesensing 17 00915 g010
Remotesensing 17 00915 g011
Figure 11. Case Study 3: comparing rocky habitat classification to WDFW priority habitat maps for talus slopes and cliffs. Maps show the spatial extent of (A) WDFW priority habitat polygons and (B) the rocky land cover classification derived from our random forest modeling approach for the full Cascades region. Selected zoomed-in areas (shown in (A,B) with (C–F) blue and (G–J) red rectangles) exhibit differences in spatial extent and precision of rocky land cover classification, showing (C,G) natural color images, (D,H) WDFW priority habitat polygons, (E,I) our rocky habitat classification, and (F,J) how the two directly compare.
Figure 11. Case Study 3: comparing rocky habitat classification to WDFW priority habitat maps for talus slopes and cliffs. Maps show the spatial extent of (A) WDFW priority habitat polygons and (B) the rocky land cover classification derived from our random forest modeling approach for the full Cascades region. Selected zoomed-in areas (shown in (A,B) with (C–F) blue and (G–J) red rectangles) exhibit differences in spatial extent and precision of rocky land cover classification, showing (C,G) natural color images, (D,H) WDFW priority habitat polygons, (E,I) our rocky habitat classification, and (F,J) how the two directly compare.
Remotesensing 17 00915 g011
Table 1. Predictors used in the random forest modeling and their associated data types and sources; see Table A1 for more detailed descriptions of the predictors and definitions of the acronyms.
Table 1. Predictors used in the random forest modeling and their associated data types and sources; see Table A1 for more detailed descriptions of the predictors and definitions of the acronyms.
Data SourceData TypePredictors
Sentinel-2Spectral reflectanceAerosol, Blue, Green, Red, NIR, Red Edge 1-3, Water Vapor, Cirrus, SWIR 1-2
IndicesNDVI, NBR, NBI, MSAVI2, EVI, NDWI, AWEIno shadow, AWEIshadow
Multi-temporalSeasonal Difference in NDVI, NBR Maximum Variance, NBR Variance of Medians, NBR Variance of 25th Percentiles
NAIPSpectral reflectanceBlue, Green, Red, NIR
IndexNDVI
TextureNDVImedian, NDVImaximum
Sentinel-1RadarVertical–vertical (VV), Vertical–horizontal (VH)
TopographyMetricsElevation, Slope
IndicesTopographic Position Index (TPI), CHILI, Roughness, Compound Topographic Index (CTI), Terrain Ruggedness Index (TRI)
Table 2. The area of each sub-region/region, number of training/validation points attributed as rocky and non-rocky, and total number of training/validation points for the Cascades and Columbia regions and their associated sub-regions (italicized).
Table 2. The area of each sub-region/region, number of training/validation points attributed as rocky and non-rocky, and total number of training/validation points for the Cascades and Columbia regions and their associated sub-regions (italicized).
Sub-Region/RegionArea (km2)Number of Training/Validation Points
Non-RockyRockyTotal
N. Cascades8274280228185620
Winthrop14,800485214956347
Forbidden10,290465926417300
Baker17,6506465353510,000
Snoqualmie14,717495726607617
Rainier17,261584319197762
Adams17,416483916136452
St. Helens16,25347655475312
Cascades116,66239,18217,22856,410
Colville766131506343784
Chelan3971348510844569
Quincy211320077132720
Tri-Cities10,69429122163128
CRGNSA2424719519139108
Portland3978123381241
Columbia30,86219,982456824,550
Table 3. Overall accuracy and F-scores, as assessed using testing data for full (all 40 predictors; see Table 1) and optimized (7 predictors; see Figure 5 and Figure 6) models for the Cascades and Columbia regions and their associated sub-regions (italicized). For the Portland sub-region, there were too few rocky areas to classify the area; however, it was included in the Columbia region. For user’s and producer’s accuracies for non-rocky and rocky land cover, see Table S1.
Table 3. Overall accuracy and F-scores, as assessed using testing data for full (all 40 predictors; see Table 1) and optimized (7 predictors; see Figure 5 and Figure 6) models for the Cascades and Columbia regions and their associated sub-regions (italicized). For the Portland sub-region, there were too few rocky areas to classify the area; however, it was included in the Columbia region. For user’s and producer’s accuracies for non-rocky and rocky land cover, see Table S1.
Sub-Region/
Region		Full ModelOptimized Model
Overall
Accuracy (%)		F-ScoreOverall Accuracy (%)F-Score
Non-RockyRockyNon-RockyRocky
N. Cascades95.70.9560.95795.60.9550.956
Winthrop97.70.9850.95396.90.9800.938
Forbidden98.10.9860.97698.10.9860.972
Baker98.00.9850.97197.30.9800.961
Snoqualmie98.00.9850.97396.70.9740.955
Rainier98.40.9890.97098.10.9870.964
Adams97.50.9840.94897.30.9830.944
St. Helens97.70.9870.90597.60.9860.896
Cascades97.90.9850.96597.40.9810.954
Colville97.00.9840.78494.60.9680.820
Chelan95.30.9710.88492.20.9510.809
Quincy93.60.9570.87993.10.9540.864
Tri-Cities97.50.9860.86898.10.9900.870
CRGNSA96.50.9790.90895.70.9730.888
Columbia95.20.9710.86493.90.9630.825
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Celebrezze, J.V.; Alegbeleye, O.M.; Glavich, D.A.; Shipley, L.A.; Meddens, A.J.H. Classifying Rocky Land Cover Using Random Forest Modeling: Lessons Learned and Potential Applications in Washington, USA. Remote Sens. 2025, 17, 915. https://doi.org/10.3390/rs17050915

AMA Style

Celebrezze JV, Alegbeleye OM, Glavich DA, Shipley LA, Meddens AJH. Classifying Rocky Land Cover Using Random Forest Modeling: Lessons Learned and Potential Applications in Washington, USA. Remote Sensing. 2025; 17(5):915. https://doi.org/10.3390/rs17050915

Chicago/Turabian Style

Celebrezze, Joe V., Okikiola M. Alegbeleye, Doug A. Glavich, Lisa A. Shipley, and Arjan J. H. Meddens. 2025. "Classifying Rocky Land Cover Using Random Forest Modeling: Lessons Learned and Potential Applications in Washington, USA" Remote Sensing 17, no. 5: 915. https://doi.org/10.3390/rs17050915

APA Style

Celebrezze, J. V., Alegbeleye, O. M., Glavich, D. A., Shipley, L. A., & Meddens, A. J. H. (2025). Classifying Rocky Land Cover Using Random Forest Modeling: Lessons Learned and Potential Applications in Washington, USA. Remote Sensing, 17(5), 915. https://doi.org/10.3390/rs17050915

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop