Mapping Banana Plants from High Spatial Resolution Orthophotos to Facilitate Plant Health Assessment

3.1. Image and Field Data

Ten orthophoto tiles each covering 1 km × 1 km were selected for this research project from a total of 3600 tiles covering the whole of the Sunshine Coast Region. The 10 tiles, selected to represent the majority of landscape variation on the Sunshine Coast, were captured between 10 and 16 September 2011 and consisted of four multi-spectral bands with 0.10 m pixels located in the blue, green, red and near infrared (NIR) part of the spectrum. The 10 tiles all had a large number of banana plants or clusters of banana plants, occurring within different types of urban, peri-urban and semi-rural environments typical for the Sunshine Coast Region. In addition, the 10 tiles were also selected because BBTV had been detected within these tiles and there was an interest in identifying other banana plants in the vicinity of those infected banana plants.

GPS recordings of banana plant locations were collected on the 16 October 2012 through an initial inspection of several properties in the Sunshine Coast Region, including one of the 10 image tiles, and used for calibration purposes for the development of the mapping approach. Forty-two sites with banana plants were visited for calibration purposes. A two-day field campaign was carried out on 25 and 26 March 2013 to collect independent validation data to assess if the banana plants identified in the remaining nine orthophotos were banana plants, and if the method used for mapping banana plants had omitted any banana plants. This was done once the maps of delineated banana plants had been completed. To provide an understanding of why some banana plants were omitted with the existing mapping approach, omission errors were assessed by systematic door-knocking to inspect all properties along the streets within four of the 10 selected image tiles (two tiles in Kin Kin and two tiles in Buderim). In addition to the four image tiles systematically assessed, 97 out of 134 of those banana plant sites delineated in the five remaining orthophotos used for validation were inspected to assess the presence of banana plants. A total of 164 sites with banana plants were visited for validation purposes within the nine image tiles. One hundred and forty-six out of 183 sites (80% of sites) mapped with banana plants were checked in the field. Access constraints prevented 37 mapped banana plant sites from being checked in the field. A GPS receiver was used to record a single point based location of all 164 identified sites with banana plants to an accuracy within 3 m. The extent of banana plants at each site varied from approximately 1 m² (single dwarf plant) to 3000 m² (small plantation). In the majority of cases, a cluster of less than 20 banana plants were found at each location, with the exception of three small plantations and 11 clusters of banana plants varying in sizes from approximately 200–500 m². Banana plants that were identified in the field, but could not be visually identified in the orthophotos and could not be determined to have an age of more than 1.5 years, were omitted from the accuracy assessment because of the 1.5 year time gap between the capture date of the orthophotos and the field validation campaign. Classification accuracies were calculated based on the collected independent field data.

3.2. Initial Evaluation of Banana Plant Detection and Delineation Approaches

Most tree crown detection and delineation algorithms are based on two distinct spectral properties of tree crowns and their representation in high spatial resolution image data: (1) the association of a tree apex with a local maximum image brightness value; and (2) delineation of the crown boundary by local minima brightness values [7]. One of the most fundamental assumptions of crown delineation algorithms is that the centre of a crown appears radiometrically brighter than the edge [8,9] due to the more direct illumination of the crown peak [10]. The valley-following approach (local minima approach) was developed in Canada [11] based on delineation of tree crown boundaries by finding the shaded parts between the tree crowns, which appear to have low brightness values [12].

Quackenbush et al. [13] found canopy illumination and background effects to significantly reduce mapping accuracies of conventional pixel-based classifications of individual trees, especially in highly heterogeneous landscapes such as urban areas [14]. Hence, object-based image analysis is now considered more appropriate for single tree classification because of the added information available using objects [15]. Bunting and Lucas [16] used an object-based image analysis approach to delineate tree crowns of mixed species and achieved accuracies between 48% and 88%. Their approach did also look for object maxima as seeds for tree crown identification with subsequent expansion from the seed location to the edges of the tree crowns and clusters.

However, these tree crown delineation and object-based approaches were not found suitable for the delineation of individual and clusters of banana plants because the banana plants’ structure, composition and leaf shape, which prevent the detection of a sunlit apex. Leaves are spirally arranged and may grow up to 3 m long and about 0.60 m wide (Figure 3a). As leaves are easily torn by the wind, the leaves often appear with a frond look (Figure 3b). This causes some parts of the leaves to be sunlit and other parts to be shaded (Figure 3a,b), and hence reducing the ability to use spectral information for the detection of banana plants. Hence the main focus of the detection of banana plants was on the leaf shape and length and the contrast between the leaves/leaf shadows and the ground. As most banana plants are <7 m tall, distinct shadows cast by the leaves onto the ground were in most cases “in touch” with the banana plant leaves from an aerial viewpoint (Figure 4).

3.3. Banana Plant Detection and Delineation

The objective of geographic object-based image analysis (GEOBIA) is to develop and apply theory, methods and tools for replicating and improving human interpretation of remotely sensed image data in an automated manner. GEOBIA consists of image segmentation, i.e., clustering of pixels into homogenous objects, and subsequent classification or labelling of the objects, and modelling based on the characteristics of objects. In high spatial resolution image data, individual features on the ground usually consist of multiple pixels, which may exhibit variable spectral reflectance characteristics [17]. GEOBIA overcomes problems of traditional pixel-based techniques of high spatial resolution image data, by firstly defining segments rather than pixels to classify, and allowing spectral reflectance variability to be used as an attribute for discriminating features in the segmentation approach. In addition, GEOBIA allows inclusion of additional information to guide the classification and modelling processes, by the use of: object average reflectance; object standard deviation; object maximum, minimum and median values; area and shape of objects; texture of objects; location of objects in relation to other objects and land cover classes in the landscape; relation of objects to the image scene characteristics; exploitation of existing geographic information from spatial data infrastructures, and many other contextual information properties not available in traditional pixel-based approaches [18,19]. Very high spatial resolution image data are required to enable detection of individual banana leaves to identify and map individual and small clusters of banana plants. Using orthophotos of 0.10 m pixels, GEOBIA was deemed to be the only suitable approach for mapping banana plants.

GEOBIA builds on the older segmentation, edge-detection, feature extraction and classification concepts of remote sensing image analysis and the idea of incorporating contextual information in the classification [20,21]. The increasing availability of high spatial resolution digital image data in the 2000s coincided with the release of the first commercially available object-based image analysis software, eCognition [22]. A variety of software packages focusing on GEOBIA have been developed since the launch of eCognition in 2000 [18]. However, the eCognition software is currently the most advanced and sophisticated GEOBIA software available. The eCognition Server allows batch processing of very large data sets by utilising multiple computers for the processing generally using a tiling and stitching approach. The tiling and stitching approach cuts an image into multiple subsets, sends each subset to separate computers for processing and upon completion of the processing, the subsets are stitched back together.

The eCognition Client version was used for the rule set development, while testing of developed rule sets were performed with the eCognition Server for faster processing. A total of 16 computers were used for the processing (4 × 2.66 GHz with 16GB RAM, 6 × 2.9 GHz with 64 GB RAM, 6 × 2.53 GHz with 24 GB RAM). The sites with banana plants and clusters of banana plants identified in the field for calibration purposes were used to train the rule set and select suitable algorithm parameter values for each process in the rule set. The individual steps in the processing approach are illustrated in Figures 5 and 8, using a subset of orthophoto g in Figure 2. This subset was selected because three clusters of banana plants of varying sizes were located in close proximity and next to different land-cover classes. The overall processing approach in the eCognition software included a number of main stages:

• Mapping of vegetation features (Table 1);

• Creation of base objects [23] (Table 2);

• Production of edge extraction bands (Table 2);

• Detection of bright and dark lines from the edge extraction bands (Table 2);

• Extraction of banana plant candidates based on lines and areas with a high density of edges (Table 3);

• Creation of buffers around banana plant candidates (Table 4);

• Manual/Visual acceptance or decline of banana plant candidate objects; and

• Export results to shapefiles.

In eCognition Architect an Action Library and an associated Architect Solution were produced to manually accept or decline each of the banana plant candidate buffers (Figure 8h). The visually assessed maps were subsequently exported as a shapefile. In ArcMap, the delineated banana plant objects in the shapefile were further assessed and assigned to be mapped with “high confidence” or “low confidence” as some banana plants and clusters of banana plants were easier than others to confidently select in the eCognition Architect.

3.4. Map Validation

Upon completion of the maps of banana plants, nine out of the 10 image tiles were visited in the field to assess errors of commission and omission (the banana plant location information from the last image tile was used for calibration purposes). The orthophotos and the derived banana plant maps were brought on the field trip on A3 printouts and on a laptop. Areas that appeared to have changed (9% of sites visited), e.g., (1) if no banana plant-looking features were present during the field trip on 25 and 26 March 2013 while mapped in the orthophotos (six locations) or (2) if banana plants were present at the time of the field trip, but did not appear in the orthophotos (12 locations), were excluded from the accuracy assessment. Even though the three small banana plantations within the image tiles produced multiple objects (28 objects for the largest plantation) in the analysis because of the separation between some of the banana plants, these objects were only considered as a single location in the accuracy assessment to provide a more realistic measure of the detection rate.

4.1. Banana Plant Mapping Results

The eCognition processing of each image tile was performed in 15–20 min when cutting the 1 km × 1 km orthophoto tiles into four subsets for batch processing in the eCognition Server software. Generally the identification and delineation of clusters of banana plants worked very well using the developed rule set (Figures 9 and 10), whereas individual banana plants (16) were in some cases omitted. Those that were omitted generally had limited contrast against the ground (Figure 11), whereas those showing more contrast or creating distinct leaf shaped shadows on the ground were automatically identified (Figure 12).

An example of the final classification results is presented in Figure 13. Based on the field validation data, 94 of the 96 banana plant objects classified with high confidence (yellow outlines) were in fact banana plants. This yields a user’s accuracy of 98%. Of the 50 low confidence (orange outlines) objects inspected in the field, 35 of them were in fact objects with banana plants. This yields a user’s accuracy of 70%. The false positives (features being incorrectly outlined as banana plants) were a combination of palms, Heliconia sp. and Strelitzia sp. That means if all mapped objects (146 in total) were inspected in the field, there is over 88% probability that the object will in fact represent banana plants.

A total of 35 banana plants/clusters of banana plants were not identified by the mapping approach used. Nineteen objects of banana plants were found underneath trees and could not be identified in the orthophotos. Some of the other unidentified banana plants were either too small to identify (six banana plants) in the images or lacked contrast against the ground (10 banana plants). The dwarf banana plant varieties were in general more difficult to identify, because of their smaller size and shorter leaves, which made the automatic detection of dark and bright lines in the eCognition software difficult. The lack of contrast between banana plant leaves and leaf shadows against the ground resulted in the lack of edge and line detection in the rule set and hence the omission of these banana plants. After visual orthophoto reassessment of all the 35 missed objects with banana plants, it was concluded that only nine (three dwarf banana plants and six plants with limited contrast against the ground) of the 35 omitted banana plants could be confidently identified visually in the images. The nine locations of banana plants were missed as part of the automatic processing. In some cases, the lack of edge detection in the eCognition software caused the smoothing of the added bright and dark line images to produce brightness values that fell below the threshold set in the rule set for identification of banana plant candidates. While the automatic approach did select most of them, they were in some cases difficult to confidently accept in the final manual step using the eCognition Architect. The fact that not all banana plants can be identified from the orthophotos presents a challenge for field inspection teams searching for banana plants infected with BBTV. However, with a detection rate of 79% (129 out of 164 locations), field inspection teams will save significant amounts of time and be able to inspect a much larger number of sites with banana plants than previously.

The denser populated areas, e.g., Buderim, had many single banana plants or small clusters of banana plants, which made these areas the most challenging to map, especially when combined with the heterogeneous urban landscape with many other distinct edge and line features. In urban areas, other types of trees and buildings were often found adjacent to single or a small number of banana plants, which in some cases were partly obstructed or appeared partly shaded. This reduced the ability to detect banana plant leaves and their shadows as line features. The presence of many varieties of palms in urban areas increased the number of false positives.

The number of objects classified as banana plant candidates from the automatic rule set processing of the 10 tiles is presented in Table 5. It is obvious that the automatic processing method produced a large number of false positives, which emphasize the need for the final step using the eCognition Architect solution to visually accept or decline banana plant objects. The time spent on the visual assessment in the eCognition Architect for one image tile was between 5 and 31 min. This is significantly faster than visually assessing the whole image tile and manually delineating banana plants, which took 73 min on average for the 10 image tiles, without the use of the object-based image classification approach.

4.2. Large Area Mapping

The eCognition Server was found suitable for large area mapping. A total of 16 computers were used to process the 3600 orthophoto tiles of the Sunshine Coast. About 500–600 tiles were processed per 24 h. Given the vast extent of the Sunshine Coast, it would be beneficial to reduce the number of false positives, while still avoiding omitting objects representing banana plants. Some forested areas had in many cases a large number of false positives, as several features occurred with distinct edge and line characteristics fulfilling the set criteria used to map banana plants. LiDAR data collected with properly selected data acquisition specifications and at the same time as the orthophotos would be useful to prevent false positives in forested areas based on height information. Banana plants vary in height depending on the variety and growing conditions, but they rarely exceeds a height of 8 m. Most varieties are between 2 m and 6 m in height. As many palms and trees are taller than banana plants, the use of height information derived from LiDAR data can be used to reduce the number of false positives.

Along drainage lines and streams within forested areas, palms were in many cases present, which also resulted in false positives because of their compositional similarity to banana plants (Figure 14). Palms in general were the cause of many false positives in both peri-urban and urban areas (Figure 15). Especially the area around Buderim resulted in many false positives because of the large number of palms in this area. Man-made features and associated shadows appearing with distinct edges and line objects between 1 and 5 m in length also resulted in false positives. Examples included roof tops, foot paths and power poles and their shadows. While this was not a major issue for the 10 image tiles mapped in this research project, there were many of these examples from the processing of the whole Sunshine Coast data set, highlighting the need to implement additional mapping steps in the rule set for large areas to reduce the number of false positives and improve the detection of banana plants.

The main limitation of the automatic mapping approach was the inclusion of many false positives. While this was not a major issue because of the eCognition Architect solution developed as part of the rule set, which automatically goes through each outlined object and allow the image analyst to accept or decline the objects, it does increase the time spent on completing the mapping of banana plants, and it is impractical to implement for large area mapping. Based on a sample of 100 assessed image tiles, the number of false positives ranged from 5 to 2759, with an average of 497. The visual assessment step in the eCognition Architect was only applied to the sample of 100 image tiles. While banana plants occurred in 52 of the selected 100 image tiles, there were no obvious trends in their distribution. In addition, the final results rely on the expertise of the image interpreter conducting the final visual mapping step and the ability to visually discriminate banana plants from similar looking features also outlined in the automatic mapping approach. By reducing the number of false positives through the change of thresholds, there will be fewer objects to visually accept or decline, making this process faster, but that includes the risk of omitting banana plants, especially those with less contrast against the background. Hence the overall mapping approach is currently a compromise between accuracy and time spent per image tile. There may be other approaches to explore in order to optimise the eCognition rule set further for large area mapping, but that was outside the scope of this project.

Despite this compromise, the rule set included novel individual mapping approaches that will be useful for many mapping applications in the future. An example is the automatic creation of buffers around the banana plant candidates. This approach can be used for mapping other features of interest such as different types of trees and agricultural areas with distinct characteristics, e.g., damage, to clearly and neatly outline these features in an automated manner. This approach is particularly useful when mapping features where the entire part of the feature cannot be mapped. For example, in this research only some banana plant leaves were identified based on their edge and line characteristics, but the whole banana plant or cluster of plants was generally delineated automatically because of the buffering approach.

4.3. Accuracies, Errors, and Uncertainties

The main strength of the developed object-based image analysis approach was the high detection rate of banana plants. While the producer’s accuracy was 79%, over half of the omission errors were due to the obstruction of banana plants by tall trees, making it impossible to detect these banana plants. The main limitation of the mapping approach was the large number of false positives, which required the final visual assessment step to reduce the errors of commission. After the application of the final visual assessment step the errors of commission were reduced to 12%, which represented Heliconia sp. and Strelitzia sp. and some palms that looked very similar to some banana plants. Chemura et al. [24] reported similar accuracies for mapping individual oil palms and that palm age influenced the user’s and producer’s accuracies. The differences in banana plant varieties, their density in clusters of plants, their health, height, age, and location in the landscape contributed to variation in their appearance in the orthophotos, which caused some errors of commission to the final visual assessment step. Johansen et al. [25] also found that young, variably aged and damaged banana plants reduced the mapping accuracy of banana plantations in north Queensland, Australia. Although the final visual assessment step focussed on reducing errors of commission, this step was ultimately a compromise between user’s and producer’s accuracies.

The only uncertainties introduced in the accuracy assessment were attributed to the time gap of 1.5 years between the orthophoto and field data acquisition times. Twenty-six locations with banana plants identified in the field were not identified in the orthophotos, in most cases because of the obstruction of tall trees. Because of the 1.5 year time gap between the orthophoto and field data collections, some of these 26 banana plants/clusters of banana plants may not have been present at the time of the orthophoto acquisition. Hence, the producer’s accuracy may in fact be higher than the reported 79%.