[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Next Article in Journal
An InSAR Deformation Phase Retrieval Method Combined with Reference Phase in Mining Areas
Next Article in Special Issue
Enhancing Object Detection in Remote Sensing: A Hybrid YOLOv7 and Transformer Approach with Automatic Model Selection
Previous Article in Journal
Probing Dust and Water in Martian Atmosphere with Far-Infrared Frequency Spacecraft Occultation
Previous Article in Special Issue
C2S-RoadNet: Road Extraction Model with Depth-Wise Separable Convolution and Self-Attention
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 15
Issue 18
10.3390/rs15184572
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 thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Efficient Wheat Lodging Detection Using UAV Remote Sensing Images and an Innovative Multi-Branch Classification Framework

by
Kai Zhang
1,†,
Rundong Zhang
2,†,
Ziqian Yang
2,
Jie Deng
1,2,*,
Ahsan Abdullah
2,
Congying Zhou
2,
Xuan Lv
2,
Rui Wang
3 and
Zhanhong Ma
2
1
Jiangjin Meteorological Bureau, China Meteorological Administration Key Open Laboratory of Transforming Climate Resources to Economy, Chongqing 402260, China
2
Department of Plant Pathology, College of Plant Protection, China Agricultural University, Beijing 100193, China
3
Kaifeng Experimental Station, China Agricultural University, Kaifeng 475000, China
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Remote Sens. 2023, 15(18), 4572; https://doi.org/10.3390/rs15184572
Submission received: 31 July 2023 / Revised: 4 September 2023 / Accepted: 15 September 2023 / Published: 17 September 2023
(This article belongs to the Special Issue Recent Advances in High Resolution Remote Sensing Image Processing and Analysis: Methodology and Application)
Browse Figures
Figure 1
<p>(<b>a</b>) Map of China; (<b>b</b>) Xiangcheng District, Hubei Province, China, Revision number: GS(2016)2923; (<b>c</b>) map and sample image projection of the study area in Oumiao Town: WGS84 UTM ZONE 50N. Data produced by the authors.</p> "> Figure 2
<p>Detailed overview of labeled samples.</p> "> Figure 3
<p>Adapted Deeplabv3 + CNN architecture for lodging area segmentation [<a href="#B22-remotesensing-15-04572" class="html-bibr">22</a>].</p> "> Figure 4
<p>RGB imagery and multispectral imagery.</p> "> Figure 5
<p>The framework used for single-branch multi-classification models and multi-branch binary classification models.</p> "> Figure 6
<p>Prediction results of different algorithm models.</p> "> Figure 7
<p>Comparison of the detection results of different types of remote sensing data.</p> "> Figure 8
<p>Comparison of detection results under different frameworks.</p> "> Figure 9
<p>Examples of model misclassification.</p> ">
Versions Notes

Abstract

:
Wheat lodging has a significant impact on yields and quality, necessitating the accurate acquisition of lodging information for effective disaster assessment and damage evaluation. This study presents a novel approach for wheat lodging detection in large and heterogeneous fields using UAV remote sensing images. A comprehensive dataset spanning an area of 2.3117 km2 was meticulously collected and labeled, constituting a valuable resource for this study. Through a comprehensive comparison of algorithmic models, remote sensing data types, and model frameworks, this study demonstrates that the Deeplabv3+ model outperforms various other models, including U-net, Bisenetv2, FastSCN, RTFormer, Bisenetv2, and HRNet, achieving a noteworthy F1 score of 90.22% for detecting wheat lodging. Intriguingly, by leveraging RGB image data alone, the current model achieves high-accuracy rates in wheat lodging detection compared to models trained with multispectral datasets at the same resolution. Moreover, we introduce an innovative multi-branch binary classification framework that surpasses the traditional single-branch multi-classification framework. The proposed framework yielded an outstanding F1 score of 90.30% for detecting wheat lodging and an accuracy of 86.94% for area extraction of wheat lodging, surpassing the single-branch multi-classification framework by an improvement of 7.22%. Significantly, the present comprehensive experimental results showcase the capacity of UAVs and deep learning to detect wheat lodging in expansive areas, demonstrating high efficiency and cost-effectiveness under heterogeneous field conditions. This study offers valuable insights for leveraging UAV remote sensing technology to identify post-disaster damage areas and assess the extent of the damage.

1. Introduction

As one of the world’s three major food crops, wheat plays a crucial role in providing phytochemicals that are essential for human health, such as vitamins, starch, protein, and dietary fiber. In 2022, China’s wheat production reached approximately 1.38 × 109 tons, accounting for about 20% of the country’s total grain production and highlighting the strategic importance of securing wheat production for food security [1]. However, the phenomenon of wheat lodging, which refers to the bending or breaking of wheat stems due to adverse weather conditions or improper farming practices, poses a significant challenge to wheat cultivation [2]. Wheat lodging adversely affects water and nutrient transport and photosynthesis and increases susceptibility to pests and diseases, hindering seed filling and having a substantial impact on wheat yield. The degree of lodging intensifies if it occurs during the late-ripening and late-filling stages of wheat growth, resulting in yield reductions ranging from 10% to 20%. If lodging happens before or after wheat flowering, the yield reduction can exceed 50%, leading to complete yield loss [3]. Given this situation, leveraging computer technology to obtain timely and accurate regional wheat lodging information is vital for predicting total wheat yields and promoting global grain production regulation. Such efforts are crucial to supporting agricultural development and presenting food security strategies. Numerous strategies have been pursued within the realm of investigations concerning expansive canopy systems. These encompass approaches that dissect individual crops as distinct entities, utilizing visualization methodologies in tandem with wind dynamics [4,5,6]. However, the methods centered on comprehensive measurements at a significant scale stand out as preeminent among these strategies. Currently, the most widely used methods for wheat lodging detection include manual field measurements and high-throughput remote sensing measurements [7]. Manual measurements are subject to environmental variability and lack objectivity and uniform standards, resulting in inefficiencies and low accuracy in wheat lodging detection [8,9].
In recent years, remote sensing technology has emerged as a critical tool in detecting crop failure. Scholars have employed remote sensing techniques, including unmanned aerial vehicles (UAVs) and satellite imagery, to analyze the spatial structures and color differences between areas with and without wheat lodging. They have utilized deep learning and classical machine learning methods to advance wheat lodging detection research [10,11]. Despite its limitations regarding spatial and temporal resolution, satellite remote sensing has been used to monitor wheat lodging information. However, with the rapid development of UAV technology and data processing software, UAV remote sensing has gained popularity in agriculture due to its cost-effectiveness, operability, and high spatial and temporal resolutions [12,13]. Scholars have made significant progress in detecting wheat lodging using UAV RGB remote sensing images. For instance, Li Guang et al. achieved winter wheat lodging detection with an overall accuracy of 86.44% using textual methods, support vector machines, neural networks, and maximum likelihood methods [14].
Zhang et al. extracted features from RGB images acquired using UAVs and evaluated three classification methods: random forest, neural networks, and support vector machines. They then incorporated the robust convolutional neural network, GoogLeNet, achieving a final accuracy of 93% [15]. While classical machine learning methods have been widely used in these studies, they rely heavily on traditional feature selection methods and lack model robustness. With the advancement of computing power and the development of deep learning network architecture, deep learning techniques have shown remarkable results in agricultural disaster assessment and other areas. Based on UAV remote sensing images capturing wheat at five developmental stages, Yu et al. incorporated the attention module CBAM into the PSPNet model and employed the Tversky loss function, resulting in an approximate overall accuracy of 95% [16]. In parallel, Zhang et al. presented a novel approach that integrates transfer learning and the Deeplabv3+ network to extract the lodging area of wheat during various growth stages [17]. Their findings demonstrated an achieved dice coefficient of around 90%.
In practical applications, the wheat planting area data exhibit evident heterogeneity. The aforementioned deep learning methods rely solely on a singular type of remote sensing image data and are limited by a small data range, with most studies having data ranges of less than 0.32 hectares [15,16,17], constraining their applicability to a restricted set of practical scenarios. Therefore, it is crucial to conduct a meticulous comparative inquiry concerning the identification of wheat lodging, imbuing the use of remote sensing image data with conspicuous heterogeneity. This imperative arises because of the evolving landscape in practical applications, whereby wheat cultivation regions exhibit overt disparities. Notably, our study bears the distinction of being the first to achieve exceptional efficacy across an expansive and heterogeneous field dataset.
This study utilized a comprehensive dataset spanning 2.3117 km2 of unmanned aerial vehicle (UAV) multispectral and RGB remote sensing images, comprising a total of six images acquired from three distinct districts and counties within Xiangyang City. The key objectives addressed in this research are as follows: (1) selecting the optimal model for accurately segmenting wheat lodging in complex field conditions; (2) conducting a comparative analysis of various remote sensing data types to identify the model with superior performance; and (3) developing a novel classification framework to address land affiliation variations and achieve the pixel-level classification of UAV images.
The rest of the paper is structured as follows: Section 2 outlines the study area, data collection, and processing; Section 3 presents the methodology; Section 4 presents the comparative results of the models and inputs; Section 5 contains the discussion; and Section 6 concludes and offers future directions.

2. Materials

2.1. Description of Study Area

The study was conducted in Liangjiazhuang, Oumiao Town, Xiangcheng District, Hubei Province (112°09′12″E, 31°51′13″N) (Figure 1), a region characterized by a humid subtropical monsoon climate featuring cold and dry winters, hot and rainy summers, and simultaneous precipitation and heat. The region experiences an average annual temperature ranging from 15.2 to 16.0 °C, with an average annual sunshine duration of 1622 to 1841 h and a frost-free period of approximately 250 days. These favorable climatic conditions provide an optimal environment for the robust growth of wheat. However, the area frequently experiences wheat lodging due to the adverse impact of severe weather conditions, including strong winds and heavy rainfall, coupled with suboptimal farming practices during the middle and late stages of wheat growth [18].

2.2. Data Acquisition

The DJI M300 RTK multi-rotor UAV offers numerous advantages, including high operational efficiency, flight stability, altitude maneuverability, versatility in capturing various types of images, and minimal constraints on takeoff and landing, making it highly suitable for conducting rapid aerial photography operations in rural areas. Consequently, the M300 RTK UAV manufactured by Shenzhen DJI Innovation Technology Co. (Shenzhen, China) was utilized in this study. The UAV has a total weight of 6.3 kg (including the battery and rotors), a wheelbase of 895 mm, a maximum flight speed of 23 m/s, a maximum takeoff altitude of 7000 m, and exceptional hovering accuracy in the RTK mode: vertical ± 0.1 m; horizontal ± 0.1 m.
Data collection was completed from 23 April to 29 April 2022, under clear weather conditions with adequate illumination. The UAV was equipped with a Zenmuse H20 camera (DJI Technology Co., Shenzhen, China) and a RedEdge MX Dual dual-camera multi-spectrometer (MicaSense, Seattle, WA, USA). DJI Pilot software was utilized to plan the flight route, ensuring a flight altitude of 100 m with a heading overlap rate of 75% and a collateral overlap rate of 80%. This setup allowed for the acquisition of RGB remote sensing images with ground resolutions of 1.8 cm, as well as multispectral remote sensing images with an 8 cm resolution. The resulting imagery depicted instances of wheat lodging occurring during the middle and late stages of growth (Figure 2).

2.3. DataSet Construction and Annotation

The images captured by the UAV underwent a series of processing steps using Agisoft Metashape software to generate four images. These images were subjected to essential operations, including standardization, the construction of a dense cloud, network construction, texture layer generation, and ortho-morph construction, using specialized photogrammetry software. To ensure consistency, all image layers were projected onto the WGS 1984 UTM Zone 50N projection coordinate system in the GeoTIFF format, based on the geographic location of the image area.
For this study, semantic segmentation techniques in deep learning were employed, requiring the data to be provided in the form of masks for the efficient labeling of classified regions. Manual labeling was conducted using ArcGIS Pro 2.5.2, with experienced interpreters conducting visual interpretations. The images were categorized into “Others” (including soil, weeds, canola, houses, etc.), “Health” (representing healthy wheat), and “Lodging” (indicating wheat lodging). The detection of wheat lodging served as a supervised classification task with three classes (Figure 2). The visual interpretation of lodgings was cross-checked by two interpreters, with areas of uncertainty excluded from the subsequent classification. Furthermore, field visits were conducted to verify the accuracy of the visual interpretation.
To facilitate comparisons between different types of remote sensing data when modeling, the nearest resampling method in ArcGIS 10.7 was employed to resample RGB remote sensing images, multispectral remote sensing images, and their corresponding reference masks to an 8 cm resolution. This ensured consistency in evaluating the differences across various remote sensing data types.

2.4. Data Processing

Regarding the exported images, ensuring compliance with the criteria set forth by the semantic segmentation module in deep learning necessitated the utilization of a sliding cut approach. This approach effectively divided the images into non-repetitive segments with dimensions of 512 × 512, ensuring optimal compatibility with the module while maintaining a low repetition rate of 0.1.
The balance of the dataset plays a pivotal role in the performance of the deep learning model [19]. This study employed a dataset balancing method, based on the pixel value ratio, to select the most suitable model. Specifically, the proportion of each label value within each image was initially computed. If the combined proportion of healthy or background label values exceeded 60% and the total proportion of lodging label values was less than 1%, the image was removed. By using this approach, the dataset was balanced across the three label types, leading to improved accuracy in evaluating the model performance. Simultaneously, in order to retain the realism of the original dataset’s label proportions, no balancing treatment was applied during the comparison of different frames and different types of remote sensing images. Following these principles, the dataset was divided into four distinct groups, with detailed information provided in Table 1.
The four image blocks were randomly partitioned into training, validation, and test datasets, with a ratio of 7:2:1. The training and validation sets were utilized during the model training process, while the test set served as an independent dataset for evaluating the performance of the trained deep learning model.
Furthermore, to enhance the generalization capability and training efficiency during model training, a real-time data augmentation technique was employed in this study [20]. As a crucial component within the PaddleRS framework’s data preprocessing pipeline, this technique encompassed several enhancements, including data normalization to the range [−1, 1] and random horizontal flipping with a probability of 50%. These augmentations aimed to improve the model’s ability to generalize to unseen data and expedite the training process.

3. Method

3.1. Training of Deep Neural Networks

Semantic segmentation is a comprehensive technique that integrates image classification, target detection, and image segmentation, aiming to partition an image into distinct regions with specific spatial extents while identifying the semantic class of each region. Compared to traditional methods, convolutional neural network (CNN)-based semantic segmentation enables end-to-end training, exhibits superior adaptability and scalability, and significantly enhances the accuracy of semantic segmentation models [21].
Deeplabv3+ is a widely used semantic segmentation model that is employed extensively in various domains. Its overarching architecture encompasses two integral components: the encoder and the decoder [22] (Figure 3). Deeplabv3+ extends the Deeplabv3 model by introducing a novel encoder–decoder network structure. The encoder module retains the core features of Deeplabv3 and leverages atrous convolution to enhance model detection capabilities for small targets, which is particularly beneficial for detecting small lodging areas. Notably, the encoder incorporates atrous spatial pyramid pooling (ASPP), a critical component that performs convolutions with various dilation rates, enabling the extraction of feature representations with diverse perceptual fields. ASPP effectively exploits multi-scale feature information to achieve superior object boundary segmentation. The decoder involves the upsampling and fusion of feature maps, combining the advantages of both methods to handle objects of different sizes and produce a robust model.
In addition to Deeplabv3+, we employed several classical model architectures that have been used in the field of semantic segmentation. These include U-net, Bisenetv2, HRNet, FastSCN, and RTFormer. The U-net model, characterized by an encoder–decoder structure, enables the precise recovery of edge information in the segmentation map via feature concatenation during upsampling [23]. Bisenetv2 constructs a bilateral segmentation network with a two-way encoder that combines a lightweight network structure with a densely connected residual network structure, achieving a balance between computational speed and final accuracy [24]. HRNet employs a high-resolution feature pyramid structure, leveraging multi-layer feature pyramids to handle objects at different scales and effectively improve model performance [25]. The FastSCN model adopts a lightweight network structure that utilizes spatial context to enhance segmentation results [26]. RTFormer, based on the Transformer architecture, utilizes the self-attentive mechanism to capture global contextual information while preserving spatial details, surpassing traditional convolutional neural networks (CNN) in capturing contexts [27].
During the exploration of the optimal segmentation model, the six aforementioned models were trained on Dataset A, a balanced ultra-high-resolution image dataset, to maximize the model’s performance and generalization capabilities. Subsequently, a comparative analysis was conducted to select the most suitable semantic segmentation model for detecting wheat lodging.

3.2. Application of Multispectral Datasets

Multispectral image classification, an important application of hyperspectral technology, aims to classify various features based on the differences in reflectance across different wavelengths of light. Compared to traditional RGB images, multispectral data contain a richer set of waveband information (Figure 4), enabling more detailed feature classification. In recent years, UAV-based multispectral imaging has been extensively used in agricultural disaster detection [28,29,30,31].
In the context of semantic segmentation, the dataset plays a pivotal role in training and evaluating the model’s performance. In this study, PaddleRS, an intelligent interpretation development kit for remote sensing images, was employed to optimize the classification model’s parameters, adjust the number of input bands, and utilize the pixel values from each band of the multispectral images as inputs. These inputs were further processed by a neural network model to identify the occurrence of wheat lodging. Unbalanced Datasets C and D were employed to compare the RGB data with multispectral data, leveraging the abundant spectral information present in multispectral data. Subsequently, the Deeplabv3+ model was employed to model the wheat lodging datasets, allowing for a comprehensive comparison between RGB and multispectral data.

3.3. Multi-Branch Binary Classification Framework

In most classification problems, the conventional approach involves using a single model for multi-classification, where the model extracts and transforms features from input data using neural networks or other machine learning algorithms. The output layer of this approach consists of nodes representing different classes, and the class of the input data is determined based on the scores of these nodes.
However, traditional multi-classification methods encounter a significant limitation when dealing with the special case of wheat lodging occurring only within wheat regions, specifically related to class affiliation. This situation can lead to the incorrect classification of other non-target regions such as wheat lodging, resulting in inaccurate classification outcomes. Such inaccuracies can significantly impact the detection of wheat lodging areas within the target region, necessitating a more refined classification method to address this issue.
Hence, this study constructed an innovative multi-branch binary classification framework [32]. In this framework, an additional branch was added to the existing single branch, transforming the problem into a binary classification task. One branch focused on distinguishing wheat areas (including healthy wheat and wheat lodging) from other areas, while the other branch focused solely on identifying wheat lodging areas among other areas. Subsequently, the logic depicted in Figure 5 was applied to the results obtained from the two branches, ensuring that wheat lodging was exclusively included within the wheat region.
During the actual training process, this study utilized Dataset A and Dataset B for image segmentation using the Deeplabv3+ model. The effectiveness of the multi-branch binary classification framework under different dataset balancing scenarios was compared to explore its performance across various situations.

3.4. Model Training

In this experiment, uniform hyperparameter settings were applied to all models. The experiments were conducted using PaddlePaddle 2.4.1 and a CUDA-compatible NVIDIA GPU (GeForce GTX 1080 Ti) with a CUDA11.7 library. As some datasets exhibited imbalanced class distributions, the lodging class contained a limited number of samples, thereby restricting the optimization effect of a single loss function. Hence, a combination of two loss functions, Dice and Cross Entropy, was employed for training to address this issue.
Regarding optimizer selection, the Momentum algorithm was utilized in this experiment. As for the learning rate scheduler, the OneCycleLR method was adopted to linearly increase the learning rate from a lower value to a higher value and subsequently linearly decrease it to a value close to 0. This approach facilitates faster model convergence and mitigates overfitting risks [33]. The initial learning rate was set to 0.01 and gradually increased to 0.1 within the first 30% of the training cycles, followed by a gradual decrease to 0.0001 for the remaining 70% of the cycles. A total of 100 training cycles were conducted. Detailed parameter settings can be found in Table 2.

3.5. Evaluation Metrics

To evaluate the model’s performance in wheat lodging detection, multiple metrics were employed, including recall, precision, intersection over union (IoU), and the F1 score.
R e c a l l = T P T P + F N
P r e c i s i o n = T P T P + F P
I O U = A r e a   o f   O v e r l a p A r e a   o f   U n i o n
F 1 = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l
These metrics assessed the model’s performance at the pixel level, where true positive ( T P ) represented a case in which both the actual class and the detection class were positive (indicating a correct detection), false positive ( T P ) denoted a positive value for an incorrect detection, and false negative ( F N ) indicated a negative value for an incorrect detection. Among these metrics, particular emphasis was placed on the model’s ability to accurately identify wheat lodging situations; thus, the F1 score was selected as the evaluation metric [34].
P L = A A P A
Furthermore, in evaluating the model’s performance in wheat lodging detection, this experiment devised a formula for quantifying the extraction accuracy. The area detected by the model ( P A ) and the accurately extracted area ( A A ) were computed based on the label map and detection results. Additionally, the extraction error was incorporated as an evaluation metric, calculated by comparing the difference between the extracted area and the actual area. This method enabled the measurement of the classifier’s accuracy and reliability, facilitating a better understanding of its performance in real-world applications.

4. Results

4.1. Training Results of Various Segmentation Models

Table 3 presents the results of the experimental evaluation conducted on six distinct semantic segmentation models. The results demonstrate that the different algorithmic models exhibited commendable classification performance in terms of both health and other classes. In particular, the Deeplabv3+ model attained F1 scores of 92.48% and 93.37% for the health class and other classes, respectively. Notably, in terms of identifying lodging classes, the RTFormer model showcased a significant recall rate of 92.89%, whereas the Bisnetv2 model achieved an accuracy rate of 90.29%. All models obtained a high F1 score, with the Deeplabv3+ model demonstrating the best precision, obtaining an F1 score of 90.22%.
Figure 6 illustrates the detection outcomes achieved by the six distinct semantic segmentation models when applied to an original-resolution RGB image extracted from the test area. Upon examining the classified images, it is clear that the Deeplabv3+ model exhibits particularly good performance. Specifically, it effectively identified a small region of wheat lodging while successfully capturing the edge characteristics associated with wheat lodging.

4.2. Training Outcomes for Multispectral and RGB Data

Table 4 presents a comprehensive comparison of the modeling performance of various remote sensing image types. Notably, when considering images with equivalent resolutions, the model trained on the multispectral Dataset D exhibited a notable recall rate of 82.93% in accurately identifying the wheat lodging class. However, its overall accuracy was found to be only 79.43%, with an F1 score of 81.14% for the lodging class, indicating suboptimal performance compared to the model trained using the RGB Dataset C.
Figure 7 illustrates the comparative detection outcomes achieved by the model employing different waveband datasets. Based on the classification results, it was observed that, at the same spatial resolution, the model trained on the multispectral dataset exhibited a higher precision of wheat lodging in comparison to the RGB dataset. It effectively captured more instances of actual wheat lodging. However, there were instances of misclassification where healthy wheat was erroneously identified as wheat lodging. Furthermore, the multispectral-dataset-trained model demonstrated the proficient discrimination of other vegetation types, correctly assigning them to their respective classes. Each of these two types of remote sensing images has its distinct advantages.

4.3. Utilization of Multi-Branch Binary Classification

Based on the findings presented in Table 5, the performance of the Deeplabv3+ algorithm varied across different frameworks and balance conditions. Adopting the multi-branch binary classification framework yields a trade-off between the recall rate and the accuracy of the wheat lodging identification. Specifically, in unbalanced datasets, the multi-branch binary classification framework shows more noticeable improvements compared to the single-branch multi-classification framework, resulting in a noteworthy 2.19% enhancement of the F1 score for the lodging class.
Furthermore, the multi-branch binary framework exhibited significant optimization potential in the context of area detection. In balanced datasets, this framework enhanced the accuracy of area extraction by 7.22% when contrasted with the single-branch multi-classification model. Notably, in unbalanced datasets, the accuracy improvement became even more substantial, reaching an impressive 11.53%. It is notable that the multi-branch binary classification framework model effectively mitigates the error associated with area extraction, irrespective of the dataset’s balance status.
Upon examining the detection outcomes illustrated in Figure 8, it becomes clear that the single-branch multi-classification framework model exhibited significant inaccuracy in identifying areas of other vegetation. Specifically, it erroneously detected certain regions of other vegetation as wheat lodging. In contrast, the multi-branch binary classification framework model demonstrated superior accuracy when detecting other vegetation areas, successfully circumventing such misclassifications. Consequently, the multi-branch binary classification framework model enhanced the overall performance of the detection model.

5. Discussion

5.1. Impact of Different Segmentation Models on Wheat Lodging Recognition Accuracy

It is imperative to tailor semantic segmentation algorithms to specific scenarios. Evaluating multiple semantic segmentation models (Deeplabv3+, U-net, FastSCN, RTFormer, Bisenetv2, and HRnet) using the dataset revealed that the Deeplabv3+ model achieved the highest F1 score. This model demonstrated remarkable proficiency in accurately detecting small areas of wheat lodging and capturing the edge features of wheat lodging. The Deeplabv3+ model’s superiority stems from its utilization of advanced techniques such as null convolution and multi-scale feature fusion, enabling the effective semantic segmentation of small targets. Moreover, techniques such as global pooling and adaptive dilation convolution that are employed by the Deeplabv3+ model enhance edge detection accuracy. With its outstanding performance and generalization capabilities in detecting wheat lodging, the Deeplabv3+ model exhibits remarkable promise for practical applications.

5.2. Effect of Different Remote Sensing Data on Wheat Lodging Recognition Accuracy

The findings indicated that both types of remote sensing image data could be utilized for wheat lodging detection, with the model trained on the RGB dataset outperforming the model trained on the multispectral dataset. This observation aligns with the findings of a study conducted by Zhao et al., where the U-net model accurately detected rice lodging using vegetation indices extracted from both RGB and multispectral data, with the model trained on the RGB dataset yielding superior results [35]. Although the multispectral dataset offers richer spectral information, it also introduces additional noise and interference, thereby complicating image processing and feature extraction. Consequently, the multispectral dataset may exhibit a higher false positive rate compared to the model trained on the RGB dataset, leading to decreased accuracy. However, the multispectral dataset can provide supplementary information that enables the accurate detection of certain unidentified instances of wheat lodging, thereby enhancing the recall rate. In practical applications, a higher recall rate signifies closer proximity between the detected wheat lodging area and the actual one, arming farmers with more precise and effective information-based support to safeguard crop growth and yields; multispectral remote sensing data boast their own merits in practical applications. Both types of remote sensing images considered in the study offer advantages in terms of cost-effectiveness, a large coverage area, and operational efficiency, effectively addressing the practical requirements of wheat lodging detection.

5.3. Effect of Different Frameworks on Wheat Lodging Recognition Accuracy

The experimental results demonstrated that the adoption of a multi-branch binary classification framework enhanced the model’s performance and area extraction accuracy. Notably, the multi-branch binary classification framework exhibited superior outcomes when confronted with non-equilibrium datasets, which has significant implications for wheat lodging detection. Given the scarcity of wheat lodging samples relative to healthy wheat samples and the presence of weeds and other plants with spectral and textural similarities to wheat lodging, optimizing a single-branch multi-classification framework model to a multi-branch binary classification framework model becomes imperative for simplifying classification complexity. A related study conducted by Wen et al. revealed that leveraging class-specific subnetworks for classification, each dedicated to a distinct class, enabled more accurate segmentation and classification while reducing competition among different classes, thereby enhancing the model performance [36].

5.4. Identification of Wheat Lodging Areas

The accurate determination of wheat lodging in various area ranges was achieved using the raster transect function of ArcGIS10.7. The experimental results (Table 6) indicated that, while the optimal model exhibited relatively high accuracy when detecting the total area of wheat lodging, significant errors arose when identifying small lodging areas, specifically those within the ranges of [0.01, 1], [1, 5], and [5, 10].
By conducting a comparative analysis between the original image and the detection image (Figure 9), it became evident that certain instances of weeds, trees, and houses were erroneously classified as wheat lodging during the detection process. This misclassification could be attributed to the similarities in the spectral information between these objects and wheat lodging, an issue compounded by the dataset’s extensive range of land-cover types, which introduces complexity to the classification task. In a study conducted by Liu et al., a supervised classification approach demonstrated that favorable performance was achieved in wheat lodging detection by incorporating spectral features, vegetation index features, and texture features [37]. Building upon this research, further model optimization could involve combining spectral features, vegetation index features, and texture features to collectively construct classification features, thereby enhancing the algorithm’s classification accuracy and mitigating misclassifications of other objects, such as wheat lodging. By employing such an approach, the accurate detection of wheat lodging areas could be achieved more reliably.

5.5. Effect of Different External Environments on the Accuracy of Wheat Lodging Recognition

The Deeplabv3+ model, with its end-to-end feature, has the advantage of focusing solely on the task’s input and output without the need for intricate feature extraction from the input data. This facilitates swift iterations in processing the task, distinguishing it from traditional machine learning algorithms. In contrast to previous studies, this research included varying lighting conditions, diverse wheat varieties, different growth periods within the study area, and a range of land-cover types. Additionally, data augmentation techniques were applied to the training dataset, resulting in a more heterogeneous and diverse dataset. Consequently, the methodology proposed in this study demonstrates strong adaptability to the actual environment, exhibiting good performance even in the presence of complex external factors.
When using the semantic segmentation approach, the lodging class segmentation model showcased exceptional performance, achieving an impressive F1 score of 90.30%. This achievement demonstrates its potential to accurately detect both healthy and wheat lodging areas in expansive farmland encompassing diverse land-cover types, using only consumer-grade RGB data acquired through unmanned aerial vehicles (UAVs) in conjunction with deep learning neural network models.

6. Conclusions

The findings of this study highlighted the superior performance of the Deeplabv3+ model over five alternative semantic segmentation models in terms of recognition accuracy, establishing its suitability for practical wheat lodging detection applications. Training the model using both multispectral and RGB data yielded excellent results, with RGB data proving particularly effective for wheat lodging detection in large-scale wheat fields. The adoption of the multi-branch binary classification framework significantly enhanced the area detection accuracy, particularly in non-equilibrium classes.
In summary, the utilization of consumer-grade UAV-captured ultra-high-resolution RGB images combined with deep neural networks presents a viable approach for accurately detecting wheat lodging under heterogeneous field conditions. This study involved a comprehensive examination of different algorithms, remote sensing data types, and model frameworks within the deep learning neural network model. The extensive experimental results affirm the stability and effectiveness of the proposed deep neural network model in large-scale data scenarios, characterized by varying location conditions, field types, and lighting characteristics. This methodology provides a valuable solution for accurately identifying wheat lodging across extensive areas, with a focus on high efficiency and cost-effectiveness.

Author Contributions

Conceptualization, R.Z. and K.Z.; methodology, R.Z., J.D. and Z.Y.; software, R.Z. and K.Z.; validation, K.Z., R.Z. and Z.Y.; formal analysis, R.Z. and X.L.; investigation, Z.Y., X.L. and J.D.; resources, R.Z., K.Z. and J.D.; data curation, K.Z. and R.Z.; writing—original draft preparation, K.Z. and R.Z.; writing—review and editing, J.D., C.Z. and A.A.; visualization, R.Z., K.Z., R.W. and X.L.; supervision, Z.M.; project administration, K.Z.; funding acquisition, K.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Chongqing Meteorological Department Operational Technical Research Project (YWJSGG-202319).

Data Availability Statement

Not applicable.

Acknowledgments

The authors acknowledge all those who helped us to conduct the field experiments and assisted with logistics. We thank Chao-Yan Huang, Lei Shi, and the other Xiangyang City Plant Protection Station staff for providing information and support during our data collection. Thanks to Junyan Wang and the other staff of Henan Feixiang Zhihang Electronics Technology Co., Ltd., for providing us with technical support for drone aerial photography and image stitching. Thanks are also due to the Chongqing Meteorological Department Operational Technical Research Project (YWJSGG-202319) for the financial support.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. National Bureau of Statistics National Bureau of Statistics Announcement on 2022 Early Rice Production Data. Available online: http://www.stats.gov.cn/sj/zxfb/202302/t20230203_1901559.html (accessed on 25 June 2023).
  2. Joseph, G.M.D.; Mohammadi, M.; Sterling, M.; Baker, C.J.; Gillmeier, S.G.; Soper, D.; Jesson, M.; Blackburn, G.A.; Whyatt, J.D.; Gullick, D.; et al. Determination of Crop Dynamic and Aerodynamic Parameters for Lodging Prediction. J. Wind. Eng. Ind. Aerodyn. 2020, 202, 104169. [Google Scholar] [CrossRef]
  3. Berry, P.M.; Spink, J. Predicting Yield Losses Caused by Lodging in Wheat. Field Crops Res. 2012, 137, 19–26. [Google Scholar] [CrossRef]
  4. Py, C.; de Langre, E.; Moulia, B.; Hémon, P. Measurement of Wind-Induced Motion of Crop Canopies from Digital Video Images. Agric. For. Meteorol. 2005, 130, 223–236. [Google Scholar] [CrossRef]
  5. Py, C.; Langre, E.D.; Moulia, B. A Frequency Lock-in Mechanism in the Interaction between Wind and Crop Canopies. J. Fluid Mech. 2006, 568, 425–449. [Google Scholar] [CrossRef]
  6. De Langre, E.; Penalver, O.; Hémon, P.; Frachisse, J.-M.; Bogeat-Triboulot, M.-B.; Niez, B.; Badel, E.; Moulia, B. Nondestructive and Fast Vibration Phenotyping of Plants. Plant Phenomics 2019, 2019, 6379693. [Google Scholar] [CrossRef] [PubMed]
  7. Li, L.; Zhang, Q.; Huang, D. A Review of Imaging Techniques for Plant Phenotyping. Sensors 2014, 14, 20078–20111. [Google Scholar] [CrossRef]
  8. Weiss, M.; Jacob, F.; Duveiller, G. Remote Sensing for Agricultural Applications: A Meta-Review. Remote Sens. Environ. 2020, 236, 111402. [Google Scholar] [CrossRef]
  9. Yang, G.; Liu, J.; Zhao, C.; Li, Z.; Huang, Y.; Yu, H.; Xu, B.; Yang, X.; Zhu, D.; Zhang, X.; et al. Unmanned Aerial Vehicle Remote Sensing for Field-Based Crop Phenotyping: Current Status and Perspectives. Front. Plant Sci. 2017, 8, 1111. [Google Scholar] [CrossRef]
  10. Liu, L.; Wang, J.; Song, X.; Li, C.; Huang, W. The Canopy Spectral Features and Remote Sensing of Wheat Lodging. J. Remote Sens.-Beijing 2005, 9, 323. [Google Scholar] [CrossRef]
  11. Yang, H.; Chen, E.; Li, Z.; Zhao, C.; Yang, G.; Pignatti, S.; Casa, R.; Zhao, L. Wheat Lodging Monitoring Using Polarimetric Index from RADARSAT-2 Data. Int. J. Appl. Earth Obs. Geoinf. 2015, 34, 157–166. [Google Scholar] [CrossRef]
  12. Du, M.; Noguchi, N. Multi-Temporal Monitoring of Wheat Growth through Correlation Analysis of Satellite Images, Unmanned Aerial Vehicle Images with Ground Variable. IFAC-PapersOnLine 2016, 49, 5–9. [Google Scholar] [CrossRef]
  13. Bendig, J.; Yu, K.; Aasen, H.; Bolten, A.; Bennertz, S.; Broscheit, J.; Gnyp, M.L.; Bareth, G. Combining UAV-Based Plant Height from Crop Surface Models, Visible, and near Infrared Vegetation Indices for Biomass Monitoring in Barley. Int. J. Appl. Earth Obs. Geoinf. 2015, 39, 79–87. [Google Scholar] [CrossRef]
  14. Li, G.; Zhang, L.; Song, C.; Peng, M.; Han, W. Extraction Method of Wheat Lodging Information Based on Multi-Temporal UAV Remote Sensing Data. Nongye Jixie Xuebao/Trans. Chin. Soc. Agric. Mach. 2019, 50, 211–220. [Google Scholar] [CrossRef]
  15. Zhang, Z.; Flores, P.; Igathinathane, C.; Naik, D.L.; Kiran, R.; Ransom, J.K. Wheat Lodging Detection from UAS Imagery Using Machine Learning Algorithms. Remote Sens. 2020, 12, 1838. [Google Scholar] [CrossRef]
  16. Yu, J.; Cheng, T.; Cai, N.; Zhou, X.-G.; Diao, Z.; Wang, T.; Du, S.; Liang, D.; Zhang, D. Wheat Lodging Segmentation Based on Lstm_PSPNet Deep Learning Network. Drones 2023, 7, 143. [Google Scholar] [CrossRef]
  17. Zhang, D.; Ding, Y.; Chen, P.; Zhang, X.; Pan, Z.; Liang, D. Automatic Extraction of Wheat Lodging Area Based on Transfer Learning Method and Deeplabv3+ Network. Comput. Electron. Agric. 2020, 179, 105845. [Google Scholar] [CrossRef]
  18. Guo, G.; Li, C.; Li, X.; Zhao, W.; Zhang, H.; Zou, J.; Zhu, Z.; Gao, C. Investigation on the Lodging of Wheat in Xiangzhou District of Xiangyang City. Hubei Agric. Sci. 2018, 57, 41. [Google Scholar] [CrossRef]
  19. Buda, M.; Maki, A.; Mazurowski, M.A. A Systematic Study of the Class Imbalance Problem in Convolutional Neural Networks. Neural Netw. 2018, 106, 249–259. [Google Scholar] [CrossRef]
  20. Cubuk, E.D.; Zoph, B.; Mane, D.; Vasudevan, V.; Le, Q.V. AutoAugment: Learning Augmentation Strategies From Data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 15–20 June 2019; pp. 113–123. [Google Scholar]
  21. Thoma, M. A Survey of Semantic Segmentation. arXiv 2016, arXiv:1602.06541. [Google Scholar] [CrossRef]
  22. Chen, L.-C.; Zhu, Y.; Papandreou, G.; Schroff, F.; Adam, H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018; pp. 801–818. [Google Scholar]
  23. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015, Munich, Germany, 5–9 October 2015; Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F., Eds.; Springer International Publishing: Cham, Switzerland, 2015; pp. 234–241. [Google Scholar]
  24. Yu, C.; Gao, C.; Wang, J.; Yu, G.; Shen, C.; Sang, N. BiSeNet V2: Bilateral Network with Guided Aggregation for Real-Time Semantic Segmentation. Int. J. Comput. Vis. 2021, 129, 3051–3068. [Google Scholar] [CrossRef]
  25. Wang, J.; Sun, K.; Cheng, T.; Jiang, B.; Deng, C.; Zhao, Y.; Liu, D.; Mu, Y.; Tan, M.; Wang, X.; et al. Deep High-Resolution Representation Learning for Visual Recognition. IEEE Trans. Pattern Anal. Mach. Intell. 2021, 43, 3349–3364. [Google Scholar] [CrossRef] [PubMed]
  26. Wu, H.; Zhang, J.; Huang, K.; Liang, K.; Yu, Y. FastFCN: Rethinking Dilated Convolution in the Backbone for Semantic Segmentation. arXiv 2019, arXiv:1903.11816. [Google Scholar]
  27. Wang, J.; Gou, C.; Wu, Q.; Feng, H.; Han, J.; Ding, E.; Wang, J. RTFormer: Efficient Design for Real-Time Semantic Segmentation with Transformer. Adv. Neural Inf. Process. Syst. 2022, 35, 7423–7436. [Google Scholar] [CrossRef]
  28. Deng, L.; Mao, Z.; Li, X.; Hu, Z.; Duan, F.; Yan, Y. UAV-Based Multispectral Remote Sensing for Precision Agriculture: A Comparison between Different Cameras. ISPRS J. Photogramm. Remote Sens. 2018, 146, 124–136. [Google Scholar] [CrossRef]
  29. Candiago, S.; Remondino, F.; De Giglio, M.; Dubbini, M.; Gattelli, M. Evaluating Multispectral Images and Vegetation Indices for Precision Farming Applications from UAV Images. Remote Sens. 2015, 7, 4026–4047. [Google Scholar] [CrossRef]
  30. Khaliq, A.; Comba, L.; Biglia, A.; Ricauda Aimonino, D.; Chiaberge, M.; Gay, P. Comparison of Satellite and UAV-Based Multispectral Imagery for Vineyard Variability Assessment. Remote Sens. 2019, 11, 436. [Google Scholar] [CrossRef]
  31. Abdulridha, J.; Ampatzidis, Y.; Roberts, P.; Kakarla, S.C. Detecting Powdery Mildew Disease in Squash at Different Stages Using UAV-Based Hyperspectral Imaging and Artificial Intelligence. Biosyst. Eng. 2020, 197, 135–148. [Google Scholar] [CrossRef]
  32. Deng, J.; Zhou, H.; Lv, X.; Yang, L.; Shang, J.; Sun, Q.; Zheng, X.; Zhou, C.; Zhao, B.; Wu, J.; et al. Applying Convolutional Neural Networks for Detecting Wheat Stripe Rust Transmission Centers under Complex Field Conditions Using RGB-Based High Spatial Resolution Images from UAVs. Comput. Electron. Agric. 2022, 200, 107211. [Google Scholar] [CrossRef]
  33. Smith, L.N.; Topin, N. Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates. arXiv 2018, arXiv:1708.07120. [Google Scholar] [CrossRef]
  34. Powers, D.M.W. Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness and Correlation. arXiv 2020, arXiv:2010.16061. [Google Scholar] [CrossRef]
  35. Zhao, X.; Yuan, Y.; Song, M.; Ding, Y.; Lin, F.; Liang, D.; Zhang, D. Use of Unmanned Aerial Vehicle Imagery and Deep Learning UNet to Extract Rice Lodging. Sensors 2019, 19, 3859. [Google Scholar] [CrossRef] [PubMed]
  36. Wen, S.; Dong, M.; Yang, Y.; Zhou, P.; Huang, T.; Chen, Y. End-to-End Detection-Segmentation System for Face Labeling. IEEE Trans. Emerg. Top. Comput. Intell. 2021, 5, 457–467. [Google Scholar] [CrossRef]
  37. Liu, H.Y.; Yang, G.J.; Zhu, H.C. The Extraction of Wheat Lodging Area in UAV’s Image Used Spectral and Texture Features. Appl. Mech. Mater. 2014, 651–653, 2390–2393. [Google Scholar] [CrossRef]
Remotesensing 15 04572 g001
Figure 1. (a) Map of China; (b) Xiangcheng District, Hubei Province, China, Revision number: GS(2016)2923; (c) map and sample image projection of the study area in Oumiao Town: WGS84 UTM ZONE 50N. Data produced by the authors.
Figure 1. (a) Map of China; (b) Xiangcheng District, Hubei Province, China, Revision number: GS(2016)2923; (c) map and sample image projection of the study area in Oumiao Town: WGS84 UTM ZONE 50N. Data produced by the authors.
Remotesensing 15 04572 g001
Remotesensing 15 04572 g002
Figure 2. Detailed overview of labeled samples.
Figure 2. Detailed overview of labeled samples.
Remotesensing 15 04572 g002
Remotesensing 15 04572 g003
Figure 3. Adapted Deeplabv3 + CNN architecture for lodging area segmentation [22].
Figure 3. Adapted Deeplabv3 + CNN architecture for lodging area segmentation [22].
Remotesensing 15 04572 g003
Remotesensing 15 04572 g004
Figure 4. RGB imagery and multispectral imagery.
Figure 4. RGB imagery and multispectral imagery.
Remotesensing 15 04572 g004
Remotesensing 15 04572 g005
Figure 5. The framework used for single-branch multi-classification models and multi-branch binary classification models.
Figure 5. The framework used for single-branch multi-classification models and multi-branch binary classification models.
Remotesensing 15 04572 g005
Remotesensing 15 04572 g006
Figure 6. Prediction results of different algorithm models.
Figure 6. Prediction results of different algorithm models.
Remotesensing 15 04572 g006
Remotesensing 15 04572 g007
Figure 7. Comparison of the detection results of different types of remote sensing data.
Figure 7. Comparison of the detection results of different types of remote sensing data.
Remotesensing 15 04572 g007
Remotesensing 15 04572 g008
Figure 8. Comparison of detection results under different frameworks.
Figure 8. Comparison of detection results under different frameworks.
Remotesensing 15 04572 g008
Remotesensing 15 04572 g009
Figure 9. Examples of model misclassification.
Figure 9. Examples of model misclassification.
Remotesensing 15 04572 g009
Table 1. Details of datasets.
Table 1. Details of datasets.
DatasetSpatial
Resolution (cm)		Data TypesData BalanceOccurrenceOccurrences aArea-Related Shares b (%)
OtherHealthLodgingOtherHealthLodging
a1.8 cmRGB ImageryBalanced290222102398202830.97%45.41%23.62%
b1.8 cmRGB ImageryUnbalanced31012261155616358.65%39.67%1.68%
c8.0 cmRGB ImageryUnbalanced17281576126431657.39%40.83%1.78%
d8.0 cmMultispectral
Imagery		Unbalanced17281576126431657.39%40.83%1.78%
a Occurrence of the class in the number of tiles. b Area-related share of the class in the dataset.
Table 2. Architecture network parameters.
Table 2. Architecture network parameters.
DL Framework Backbone Network Optimizer Loss Function
Paddle 2.4.1Resnet50MomentumDice+CE
LR scheduler Training batch size Max Epochs Annual strategy
OneCycleLR4100cos
Learn rate(LR) Max LR End LR Phase pct
0.010.10.00010.3
Table 3. Model accuracy when using different algorithmic frameworks.
Table 3. Model accuracy when using different algorithmic frameworks.
ModelsU-NetFastSCNRTFormerBisenetv2HRNetDeeplabv3+
Recall (%)Other87.73%89.03%88.68%89.96%88.21%91.56%
Health92.77%93.29%92.40%93.45%94.46%93.51%
Lodging88.00%89.69%92.89%88.87%89.13%91.19%
Precision (%)Other92.06%92.83%93.77%92.63%94.43%93.41%
Health89.22%90.64%91.92%90.83%89.99%93.22%
Lodging89.03%89.63%87.18%90.29%89.53%89.26%
IOU (%)Other81.56%83.30%83.74%83.96%83.85%86.01%
Health83.41%85.09%85.45%85.40%85.48%87.56%
Lodging79.40%81.25%81.72%81.12%80.71%82.18%
F1 score (%)Other89.84%90.89%91.15%91.28%91.21%92.48%
Health90.96%91.95%92.16%92.13%92.17%93.37%
Lodging88.51%89.66%89.94%89.57%89.33%90.22%
Table 4. Model performance for the different dataset types.
Table 4. Model performance for the different dataset types.
Datasetcd
Recall (%)Other97.83%98.03%
Health97.29%96.88%
Lodging80.79%82.93%
Precision (%)Other98.23%98.40%
Health96.06%96.57%
Lodging90.32%79.43%
IOU (%)Other96.14%96.49%
Health93.56%93.66%
Lodging74.35%68.27%
F1 score (%)Other98.03%98.21%
Health96.67%96.73%
Lodging85.29%81.14%
Table 5. Result of models with different frameworks.
Table 5. Result of models with different frameworks.
Datasetab
FrameworkSingle-Branch Multi-Classification FrameworkMulti-Branch Binary Classification FrameworkSingle-Branch Multi-Classification FrameworkMulti-Branch Binary Classification Framework
Recall (%)Other91.56%90.38%95.61%88.14%
Health93.51%94.67%95.86%94.37%
Lodging91.19%89.19%82.17%71.58%
Precision (%)Other93.41%94.96%97.23%93.65%
Health93.22%90.38%94.01%88.12%
Lodging89.26%91.45%70.09%85.29%
IOU (%)Other86.01%86.25%93.08%83.16%
Health87.56%86.00%90.34%83.72%
Lodging82.18%82.32%60.84%63.72%
F1 score (%)Other92.48%92.61%96.42%90.81%
Health93.37%92.47%94.93%91.14%
Lodging90.22%90.30%75.65%77.84%
Label a/km20.0339 0.03390.03390.0339
PA/km20.0396 0.03530.03860.0313
AA/km20.0316 0.03070.02780.0262
Extraction error b/km20.0057 0.00130.00470.0026
PL (%)79.72%86.94%72.03%83.56%
a Label derived from visual interpretation. b Extraction error is PA minus label.
Table 6. Area of wheat lodging in different ranges.
Table 6. Area of wheat lodging in different ranges.
Size (m2)LabelDeeplabv3+
Area[0.01, 1]40.77 748.40
[1, 5)259.67 1232.05
[5, 10]452.20 903.50
[10, 20]978.95 971.25
[20, 50]2096.74 1841.32
[50, +∞)30,114.00 29,572.30
Total33,942.33 35,268.82
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

Zhang, K.; Zhang, R.; Yang, Z.; Deng, J.; Abdullah, A.; Zhou, C.; Lv, X.; Wang, R.; Ma, Z. Efficient Wheat Lodging Detection Using UAV Remote Sensing Images and an Innovative Multi-Branch Classification Framework. Remote Sens. 2023, 15, 4572. https://doi.org/10.3390/rs15184572

AMA Style

Zhang K, Zhang R, Yang Z, Deng J, Abdullah A, Zhou C, Lv X, Wang R, Ma Z. Efficient Wheat Lodging Detection Using UAV Remote Sensing Images and an Innovative Multi-Branch Classification Framework. Remote Sensing. 2023; 15(18):4572. https://doi.org/10.3390/rs15184572

Chicago/Turabian Style

Zhang, Kai, Rundong Zhang, Ziqian Yang, Jie Deng, Ahsan Abdullah, Congying Zhou, Xuan Lv, Rui Wang, and Zhanhong Ma. 2023. "Efficient Wheat Lodging Detection Using UAV Remote Sensing Images and an Innovative Multi-Branch Classification Framework" Remote Sensing 15, no. 18: 4572. https://doi.org/10.3390/rs15184572

APA Style

Zhang, K., Zhang, R., Yang, Z., Deng, J., Abdullah, A., Zhou, C., Lv, X., Wang, R., & Ma, Z. (2023). Efficient Wheat Lodging Detection Using UAV Remote Sensing Images and an Innovative Multi-Branch Classification Framework. Remote Sensing, 15(18), 4572. https://doi.org/10.3390/rs15184572

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