A method for extracting desert shrubs from remote sensing images

Abstract

This invention provides a method for extracting desert shrubs from remote sensing images. Based on the EfficientNet-B0 baseline model, an improved FETNet model is obtained. The encoder retains the predefined original architecture of the baseline model, while an EdgeViT module is introduced at the skip connections between the encoder and decoder to enhance the model's global perception capability. An FEM module is introduced in the decoder to ensure the preservation of spatial detail information, resulting in more accurate and detailed desert shrub segmentation. Preprocessed remote sensing images are input into the FETNet model, and the model is trained using a stochastic gradient descent optimizer, cosine annealing strategy, and cross-entropy loss function to obtain the optimal weights. Comparing the FETNet model with conventional models under the same test conditions, this invention achieves optimal results in all evaluation metrics for desert shrub segmentation tasks in remote sensing images.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition, and more specifically to a method for extracting desert shrubs from remote sensing images. Background Technology

[0002] Desert shrubs play an important role in maintaining the ecological balance of deserts and preventing land desertification. Target identification using remote sensing images is an important means of studying the distribution of desert shrubs. However, desert shrubs are mostly dense small targets with complex morphology. Existing desert shrub information extraction methods still have many limitations when extracting desert shrub information from high-resolution remote sensing images, especially when dealing with desert shrubs with complex backgrounds and small targets.

[0003] Current methods for extracting desert shrub information from remote sensing imagery include thresholding methods, pixel-based classification methods, and object-based image analysis methods. Thresholding methods typically use spectral indices such as the Normalized Difference Vegetation Index (NDVI) to distinguish different vegetation types; however, when extracting desert shrub information, complex backgrounds such as bare ground or sand dunes can interfere with and destroy vegetation features in the remote sensing imagery. Pixel-based classification methods analyze the spectral features of each pixel in the remote sensing imagery to classify it and extract desert shrub information, but this method is mainly suitable for low-resolution remote sensing imagery. Object-based image analysis methods utilize more spatial, shape, and contextual information to improve classification accuracy; however, the data processing for extracting desert shrub information from high-resolution remote sensing imagery is very complex. While all of these methods have some effectiveness, they are mainly based on small-area mapping using UAV platforms; due to the difficulty and high cost of acquiring UAV data, they are difficult to apply on a large scale, and they usually rely on manual feature creation, resulting in poor generalization ability and a lack of self-learning capabilities. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a highly accurate method for extracting desert shrubs from remote sensing images with complex backgrounds.

[0005] Technical solution: A method for extracting desert shrubs from remote sensing images, comprising the following steps:

[0006] (1) Collect remote sensing images of the desert area, preprocess the remote sensing images to obtain a desert shrub dataset, and divide it into training set, validation set and test set;

[0007] (2) Construct the FETNet model. The FETNet model uses the EfficientNet-B0 model as the baseline model. The FETNet model includes an encoder and a decoder. The encoder retains the original architecture predefined in the baseline model. An EdgeViT module is introduced at the jump connection between the encoder and decoder, and an FEM module is introduced between the decoders to obtain the FETNet model.

[0008] (3) Train the FETNet model, set the initial learning rate, select the cross-entropy loss function and use the stochastic gradient descent method to find the parameter configuration that minimizes the loss function, input the training set into the FETNet model to start training, input the validation set into the FETNet model after each round of training, evaluate the performance of the FETNet model, and record the evaluation index and the model weights of each round of training.

[0009] (4) After reaching the set number of training rounds, select the model weight of the corresponding FETNet model training round with the best evaluation index as the optimal weight, apply the optimal weight to the test set, and generate the desert shrub segmentation and extraction results.

[0010] Specifically, the preprocessing includes: performing atmospheric correction and orthorectification on the remote sensing image, then fusing the panchromatic and multispectral bands of the remote sensing image, cropping the remote sensing image to a set size, marking the desert shrubs and background areas on the cropped remote sensing image, and generating a label map of a set size using the labeled data.

[0011] Specifically, the FETNet model includes 7 encoders and 4 decoders. Four EdgeViT modules are introduced at the jump connections between the encoders and decoders. The EdgeViT modules are connected to the encoder output and the corresponding decoder input. FEM modules are introduced between adjacent decoders.

[0012] Specifically, the formula for the cross-entropy loss function is:

[0013] ,

[0014] In the formula: For cross-entropy loss, For the actual label value, These are the model's predicted values;

[0015] Specifically, stochastic gradient descent updates the parameters for each sample by calculating the gradient for each sample, and then randomly selects one sample or a batch of samples in each update. The formula is as follows:

[0016] ,

[0017] In the formula: w represents the model parameters, and t represents the update round. To update the gradient of the model parameters after a sample input, η is the learning rate, and i is the index of the sample selected in the training set at the current iteration.

[0018] Specifically, step (3) also includes: updating the learning rate of the FETNet model using the cosine annealing algorithm, as shown in the following formula:

[0019] ,

[0020] In the formula: Let be the learning rate for the t-th training round. This is the lower bound of the learning rate. This is the upper limit of the learning rate, i.e., the initial learning rate. This represents the current number of training rounds. This represents the maximum number of training rounds.

[0021] Specifically, the evaluation metrics include one or more of the following: pixel accuracy, average pixel accuracy, F1 score, frequency-weighted cross overlap, and average cross-union ratio.

[0022] Specifically, the formula for the average intersection-union ratio is:

[0023] ,

[0024] In the formula: This represents the average crossover ratio. It is a true positive. It was a false positive. It was a false negative. It is a true negative. This represents the number of categories.

[0025] Preferably, in step (4), the model weight corresponding to the training round of the FETNet model with the largest average crossover ratio is taken as the optimal weight.

[0026] Preferably, the method for extracting desert shrubs from remote sensing images further includes the following steps:

[0027] (5) Using the desert shrub dataset, the FETNet model and several different deep learning models were tested under the same conditions. The evaluation metrics of the FETNet model and several different deep learning models were compared to verify their effectiveness. The deep learning models included HRRNet, DFANet, DeepLabv3+, HRNet, FCN16s, ESPNetv2, ExtremeC3Net, FCN8s and SegNet.

[0028] Beneficial Effects: Compared with existing technologies, the significant advantages of this invention are: The FETNet model is obtained by improving upon a deep learning baseline model. The model automatically scales and expands the network length, width, and resolution without altering the predefined original architecture of the baseline network, thus achieving better model performance. The encoder uses EfficientNet-B0 as the backbone network to extract multi-scale features of desert shrubs and introduces EdgeViT (Transformer module) to synthesize contextual information, enhancing the model's global perception capability. The decoder introduces an FEM (Feature Enhancement) module to ensure the preservation of spatial detail information, resulting in more accurate and detailed desert shrub segmentation. Preprocessed remote sensing images are input into the FETNet model, and the model is trained using a stochastic gradient descent optimizer, cosine annealing strategy, and cross-entropy loss function to obtain the optimal weights for the FETNet model. Comparing the FETNet model proposed in this invention with commonly used deep learning models, under the same testing conditions, the FETNet model achieves optimal results in all evaluation metrics. Attached Figure Description

[0029] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0030] Figure 2 This is a structural diagram of the FETNet model in Embodiment 1 of the present invention.

[0031] Figure 3 This is the visualization result of desert shrub extraction in Example 1 of the present invention. Figure 1 .

[0032] Figure 4 This is a magnified view of a portion of the visualization results of desert shrub extraction in Embodiment 1 of the present invention.

[0033] Figure 5 This is the visualization result of desert shrub extraction in Example 1 of the present invention. Figure 2 . Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings.

[0035] Example 1

[0036] Please see Figure 1 As shown in the figure, this embodiment provides a method for extracting desert shrubs from remote sensing images, including the following steps:

[0037] (1) Collect remote sensing images of the desert area, preprocess the remote sensing images to obtain a desert shrub dataset, and divide it into training set, validation set and test set.

[0038] In this embodiment, high-resolution remote sensing imagery from the GF-2 satellite is used to acquire remote sensing imagery data located in a desert area, ensuring that the quality of the remote sensing imagery is suitable for the input of the deep learning model.

[0039] Atmospheric and orthorectification corrections were performed on the acquired remote sensing images to eliminate atmospheric interference and image distortion, ensuring that the images accurately reflect the features of ground features. Panchromatic and multispectral bands were fused to generate high-resolution color images, and suitable experimental areas were cropped. Pixel-level annotations were applied to the cropped images, manually marking desert shrub areas and background areas to ensure accurate differentiation between them. After annotation, corresponding label images were generated using the labeled data and cropped to the size required for deep learning model input, forming a desert shrub dataset. This dataset was then divided into training, validation, and test sets according to a predefined ratio.

[0040] (2) Constructing the FETNet model: The FETNet model uses the EfficientNeT-B0 model as the baseline model. The FETNet model includes an encoder and a decoder. The encoder retains the predefined original architecture in the baseline model to extract multi-scale features of desert shrubs. The model automatically scales and expands the network length, width, and resolution without changing the predefined original architecture in the baseline network, thereby obtaining better model performance. An EdgeViT module is introduced at the skip connections between the encoder and decoder of the FETNet model to synthesize contextual information and enhance the model's global perception capability. The decoder part is used to progressively restore the resolution of the image, combining shallow and deep features. At the same time, an FEM module is introduced between the decoders to ensure the preservation of spatial detail information, making the final segmentation result more accurate and detailed, thus obtaining the FETNet model proposed in this invention.

[0041] Please refer to Figure 2As shown, in this embodiment, the FETNet model uses the EfficientNet-B0 model as the baseline model and designs 7 encoder blocks and 4 decoder blocks. Four EdgeViT modules are placed at the skip connections between the encoder and decoder in the FETNet model; the EdgeViT modules are connected to the encoder output and then to the corresponding decoder input, ensuring that feature information can be effectively transferred during encoding and decoding. In addition, the FETNet model adds 3 layers of FEM (Feature Enhancement Modules) between adjacent decoder modules to further enhance information interaction and feature transfer between different decoder modules. Through this structural design, the FETNet model can effectively integrate the efficient feature extraction capabilities of the backbone network and further enhance the information flow through the EdgeViT and FEM modules, giving the overall architecture better feature processing capabilities. The overall structure of the FETNet model consists of the encoder, EdgeViT modules, decoders, and FEM modules, forming a complete deep learning network architecture.

[0042] The basic module of the EfficientNet-B0 model is the MBConv (Mobile Inverted Bottleneck Convolution) module, which includes the following parts: Depthwise Separable Convolution, SE (SESqueeze-and-Excitation) module, and Swish activation function. First, 1x1 convolutions are used to increase the dimensionality. After the depthwise separable convolution operation, an attention mechanism (SENet) is added. Finally, 1x1 convolutions are used to reduce the dimensionality again, and the result is merged with the output.

[0043] The EdgeViT module overcomes the bottleneck of local-global-local information exchange through three key operations. First, local aggregation integrates local information from neighboring labels using deep convolution. Second, sparse attention provides a small number of representative labels globally, promoting long-distance information sharing through a self-attention mechanism. Finally, local propagation uses transposed convolution to propagate the learned global context information from representative labels to neighboring labels. Therefore, the EdgeViT module can effectively extract small-scale desert shrubs while identifying large-scale shrub regions. The self-focusing technique effectively learns global information and long-distance dependencies, helping to avoid interference from terrain undulations, diverse imaging conditions, and the similarity of desert shrubs to background features, thus achieving efficient desert shrub segmentation.

[0044] The Feature Enhancement Module (FEM) enhances the feature representation of small targets through multi-branch convolution and rotational convolution techniques. The FEM module has four branches. Each branch first performs a 1×1 convolution on the input feature image to initially adjust the number of channels in preparation for subsequent processing. The first branch uses a residual structure to preserve key feature information of small targets; the other three branches perform cascaded standard convolution operations with kernel sizes of 1×3, 3×1, and 3×3, respectively. Additional texture convolution layers are added to the two middle branches, allowing the extracted feature maps to retain more contextual information, thereby further improving the accuracy of small target detection. In this way, the FEM module significantly improves the performance of desert shrub extraction networks, especially when dealing with small targets and complex backgrounds. The FEM module not only extracts rich semantic information but also expands the receptive field, making the model perform better when dealing with densely distributed small-scale desert shrubs.

[0045] (3) Train the FETNet model, set the initial learning rate, select the cross-entropy loss function and use the stochastic gradient descent method to find the parameter configuration that minimizes the loss function, input the training set of the desert shrub dataset into the FETNet model to start training, input the validation set of the desert shrub dataset into the FETNet model after each round of training, evaluate the performance of the FETNet model, and record the evaluation index and the model weights of each round of training.

[0046] In this embodiment, the PyTorch framework is used, and the model is trained using stochastic gradient descent (SGD), cosine annealing, and cross-entropy loss function.

[0047] In this embodiment, the initial learning rate of the FETNet model is set to 0.001, which gradually decreases during training. Furthermore, the batch size is set to 2, and the maximum number of iterations is 300.

[0048] Stochastic gradient descent is a gradient-based optimization algorithm used to find parameter configurations that minimize the loss function. It updates parameters by calculating the gradient for each sample, and randomly selects one or a batch of samples in each update. The formula is as follows:

[0049]

[0050] In the formula: w represents the model parameters, and t represents the update round. To update the gradient of the model parameters after a sample input, η is the learning rate, and i is the index of the sample selected in the training set at the current iteration.

[0051] The cross-entropy loss function compares the model's predicted values ​​with the true values ​​through a logarithmic operation, reflecting the model's performance in binary classification problems. The formula is:

[0052]

[0053] In the formula: For cross-entropy loss, For the actual label value, These are the predicted values ​​from the model.

[0054] Cosine annealing is a method for adjusting the learning rate. During training, the learning rate is gradually decreased following the curve of a cosine function. This reduces the learning rate in the later stages of training, helping the model converge to the global optimum more effectively. In cosine annealing, the learning rate changes as training progresses according to the following formula:

[0055]

[0056] In the formula: Let be the learning rate for the t-th training round. This is the lower bound of the learning rate. This is the upper limit of the learning rate, i.e., the initial learning rate. This represents the current number of training rounds. This represents the maximum number of training rounds.

[0057] In each training round, the model's performance on the validation set is measured using five evaluation metrics: pixel precision (PA), average pixel precision (MPA), F1 score, frequency-weighted cross overlap (FWIOU), and average crossover ratio (MIOU), as shown in the following formulas:

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] ,

[0063] ,

[0064] ,

[0065] In the formula: It is a true positive. It was a false positive. It was a false negative. It is a true negative. Number of categories; For the model's accuracy, This represents the model recall rate.

[0066] (4) After reaching the set number of training rounds, select the model weight of the corresponding FETNet model training round with the best evaluation index as the optimal weight, apply the optimal weight to the test set of the desert shrub dataset, and generate the desert shrub segmentation and extraction results.

[0067] In this embodiment, the Mean Intersection over Union (MIOU) ratio is used as the primary evaluation metric to assess the overall performance of the model in the semantic segmentation task. During each training epoch, the MIOU is continuously monitored, and the model weights for each epoch are recorded. During the set number of training rounds, when the model achieves its maximum MIOU value on the validation set, the corresponding model weights are saved as the optimal model weights. Finally, the optimal model weights are selected for prediction on the test set to ensure the model performs best in the desert shrub extraction task.

[0068] (5) Using the desert shrub dataset constructed in step (1), the FETNet model and several different deep learning models are tested under the same conditions. The evaluation metrics of the FETNet model and several different deep learning models are compared to verify its effectiveness. In this embodiment, the deep learning models used for comparison testing include the HRRNet model, DFANet model, DeepLabv3+ model, HRNet model, FCN16s model, ESPNetv2 model, ExtremeC3Net model, FCN8s model, and SegNet model.

[0069] In this embodiment, five indicators were used for evaluation: PA, MPA, F1 score, MIOU, and FWIOU. The results are shown in Table 1 below.

[0070] Table 1

[0071]

[0072] The FETNet model exhibits the highest segmentation accuracy in the desert shrub segmentation task, outperforming other models on all metrics.

[0073] In high-resolution remote sensing imagery, desert shrubs vary in size and shape and are often densely packed small targets, leading to differences in the training and prediction results of different deep learning models. To visually demonstrate the effectiveness of the FETNet model in desert shrub segmentation on high-resolution remote sensing imagery, this example selects experimental results from several representative deep learning models and compares the extracted images of several representative types of desert shrubs. Please refer to [link / reference]. Figures 3 to 5 As shown ( Figures 3 to 5 In the image, each row represents the original remote sensing image / segmentation result of the same area, and each column represents the segmentation result of the same original remote sensing image / same model. Figure 3 The images show the original remote sensing images of three regions and the desert shrub segmentation results of different models in the corresponding regions. Sub-image a is the original remote sensing image input to the model, sub-image b is the accurate desert shrub segmentation result after manual annotation, sub-image c is the desert shrub segmentation result of SegNet, and sub-image d is the desert shrub segmentation result of FETNet model. Figure 4 for Figure 3 The images are magnified views of the three regions in the image. Sub-image a is the original remote sensing image input to the model, sub-image b is the accurate result of desert shrub segmentation after manual annotation, sub-image c is the desert shrub segmentation result of SegNet, and sub-image d is the desert shrub segmentation result of FETNet model. Figure 5 To and Figure 3 The desert shrub segmentation results of three models with the same region are selected in the image. Sub-image a is the desert shrub segmentation result of DeepLabv3+, sub-image b is the desert shrub segmentation result of ESPNetv2, and sub-image c is the desert shrub segmentation result of FCN16s. Figure 3 and Figure 4 The areas segmented by typical desert shrubs are marked with red boxes. From Figure 3 As can be seen from subgraph d, the FETNet model performs exceptionally well in desert shrub segmentation, particularly outperforming other models in small-scale shrub detection and detail capture. In contrast, from... Figure 5 As can be seen in subgraph a, DeepLabv3+'s classification results miss a large number of small-scale shrub targets, resulting in a less than ideal classification performance. From Figure 5 Subgraphs b and c show that ESPNetv2 and FCN16s perform reasonably well in capturing the overall distribution, but they fall short in handling edge details and complex backgrounds. Figure 3 subgraph c and Figure 4As shown in sub-figure c, SegNet exhibits some similarity to the FETNet model in segmentation results, but it also has significant shortcomings: In the first row marked with red boxes, SegNet misidentifies building shadows as shrubs; in the second row marked with red boxes, it incorrectly identifies roads in the desert as shrubs; and in the third row marked with red boxes, SegNet misses detections in dense, small-scale shrub regions. In contrast, the FETNet model can more accurately capture the edges of shrubs and maintain high segmentation accuracy in complex backgrounds, making the segmentation results closer to the actual shrub distribution. In summary, FETNet demonstrates higher accuracy and stability in desert shrub segmentation tasks, especially in the extraction of small-target shrubs, effectively addressing the deficiencies of existing technologies.

[0074] To verify the effectiveness of introducing the EdgeViT and FEM modules into the baseline model of the FETNet model, an ablation experiment was conducted. Using the desert shrub dataset constructed in step (1), different modules were added step by step under the same test conditions to evaluate their impact on model performance. The results are shown in Table 2 below.

[0075] Table 2

[0076]

[0077] First, feature extraction was performed using only the baseline model EfficientNet-B0, and segmentation results were generated by direct upsampling. Then, EdgeViT module, FEM module, and combination of EdgeViT module and FEM module were added in sequence to observe the improvement effect of EdgeViT module and FEM module on segmentation performance. After introducing the EdgeViT module, the pixel accuracy (PA) improved slightly, and the mean intersection-over-union ratio (MIOU) also increased, indicating that the EdgeViT module can enhance the ability to capture global information. The FEM module is mainly used for feature enhancement, especially to enhance the feature extraction ability of small targets. After adding the FEM module, the pixel accuracy and mean intersection-over-union ratio improved significantly, indicating that the module performs particularly well when dealing with small-scale shrubs, further improving the segmentation accuracy of the model. When both the EdgeViT and FEM modules are added, the pixel accuracy of the model reaches 98.72%, and the mean intersection-over-union ratio increases to 92.49%, which is significantly better than the results of using the baseline model alone. This result proves that the performance of the FETNet model in the desert shrub segmentation task is maximized after adding the EdgeViT and FEM modules.

Claims

1. A method for extracting desert shrubs from remote sensing images, characterized in that, Includes the following steps: (1) Collect remote sensing images of the desert area, preprocess the remote sensing images to obtain a desert shrub dataset, and divide it into a training set, a validation set and a test set. The desert shrub dataset contains small target shrubs and large-scale shrub areas. (2) Construct the FETNet model, which uses the EfficientNet-B0 model as the baseline model and includes an encoder and a decoder. The encoder retains the original architecture predefined by the baseline model. An EdgeViT module is introduced at the skip connection between the encoder and the decoder. In the EdgeViT module, local aggregation integrates the local information of neighboring labels through deep convolution. Sparse attention provides a small number of representative labels globally, promoting long-distance information sharing through a self-attention mechanism; local propagation uses transposed convolution to propagate the learned global context information from representative labels to neighboring labels, enabling the EdgeViT module to extract small target shrubs while recognizing large-scale shrub regions; an FEM module is introduced between the decoders, which has four branches. Each branch first performs a 1×1 convolution operation on the input feature image to initially adjust the number of channels in preparation for subsequent processing; the first branch adopts a residual structure to retain key feature information of small targets; The other three branches perform cascaded standard convolution operations, and texture convolution layers are added to the two middle branches, so that the extracted feature maps can retain more contextual information, improve the accuracy of small target shrub detection, and obtain the FETNet model. (3) Train the FETNet model, set the initial learning rate, select the cross-entropy loss function and use stochastic gradient descent to find the parameter configuration that minimizes the loss function, input the training set into the FETNet model to start training, input the validation set into the FETNet model after each round of training, evaluate the performance of the FETNet model, and record the evaluation metrics and the model weights of each round of training; the evaluation metrics include one or more of pixel accuracy, average pixel accuracy, F1 score, frequency-weighted cross overlap and average cross-union ratio; (4) After reaching the set number of training rounds, select the model weight of the corresponding FETNet model training round with the best evaluation index as the optimal weight, apply the optimal weight to the test set, and generate the desert shrub segmentation and extraction results.

2. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: The preprocessing includes: performing atmospheric correction and orthorectification on the remote sensing image, then fusing the panchromatic and multispectral bands of the remote sensing image, cropping the remote sensing image to a set size, marking the desert shrubs and background areas on the cropped remote sensing image, and generating a label map of a set size using the labeled data.

3. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: The FETNet model includes 7 encoders and 4 decoders. Four EdgeViT modules are introduced at the jump connection between the encoder and decoder. The EdgeViT modules are connected to the encoder output and the corresponding decoder input. FEM modules are introduced between adjacent decoders.

4. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: The formula for the cross-entropy loss function is: ， In the formula: For cross-entropy loss, For the actual label value, These are the model's predicted values; The formula for the stochastic gradient descent method is as follows: ， In the formula: w represents the model parameters, and t represents the update round. To update the gradient of the model parameters after a sample input, η is the learning rate, and i is the index of the sample selected in the training set at the current iteration.

5. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: Step (3) further includes updating the learning rate of the FETNet model using the cosine annealing algorithm, as shown in the following formula: ， In the formula: Let be the learning rate for the t-th training round. This is the lower bound of the learning rate. This is the upper limit of the learning rate, i.e., the initial learning rate. This represents the current number of training rounds. This represents the maximum number of training rounds.

6. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: The formula for the average intersection-union ratio is: ， In the formula: This represents the average crossover ratio. It is a true positive. It was a false positive. It was a false negative. It is a true negative. This represents the number of categories.

7. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: In step (4), the model weights corresponding to the training rounds of the FETNet model with the largest average crossover-union ratio are taken as the optimal weights.

8. The method for extracting desert shrubs from remote sensing images according to claim 1, characterized in that: The method for extracting desert shrubs from remote sensing images also includes the following steps: (5) Using the desert shrub dataset, the FETNet model and several different deep learning models were tested under the same conditions. The evaluation metrics of the FETNet model and several different deep learning models were compared to verify the effectiveness.

9. The method for extracting desert shrubs from remote sensing images according to claim 8, characterized in that: The deep learning models include HRRNet, DFANet, DeepLabv3+, HRNet, FCN16s, ESPNetv2, ExtremeC3Net, FCN8s, and SegNet.

Images