A crop remote sensing identification model and a crop remote sensing identification method

Abstract

The application discloses an improved DeepLabV3+ crop remote sensing recognition model and a crop remote sensing recognition method, and is applied to the technical field of remote sensing image recognition and comprises an input layer, an encoding layer, a decoding layer and an output layer; the input layer inputs normalized remote sensing images and vegetation indexes; the encoding layer extracts a multi-scale deep semantic feature map, a first shallow semantic feature map and a second shallow feature map of an input image; the decoding layer combines the deep semantic feature and the first shallow semantic feature, performs upper layer sampling, simultaneously fuses the second shallow feature map, and obtains a fused feature map; and the fused feature map is subjected to size adjustment, and is output through the output layer. The application has higher recognition precision in high-resolution remote sensing wheat and rape recognition, and the recognition result can better meet the demand of precision agriculture.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition technology, and more specifically to an improved DeepLab V3+ crop remote sensing recognition model and method. Background Technology

[0002] Timely and accurate acquisition of crop type distribution information has a significant impact on agricultural production. High-resolution Landsat remote sensing data can provide a relatively detailed depiction of crop spatial distribution information, thereby providing data support for adjusting planting structures, ensuring arable land security, and formulating reasonable food policies.

[0003] Research on remote sensing identification of crops faces the following problems: First, traditional crop identification algorithms (random forest, maximum likelihood method, artificial neural network, support vector machine, etc.) are based on the spectral, texture, and spatiotemporal characteristics of different crops. They extract and filter features to find features or feature sets that are conducive to classification and identification. However, the features extracted by this method often cannot fully and comprehensively present and express the remote sensing information of crops. Second, with the development of deep learning, semantic segmentation models such as U-Net, SegNet, and DeepLab series have been applied in the field of crop identification. Compared with traditional identification methods, semantic segmentation methods can learn the deep semantic features of different crops in remote sensing images, improving the accuracy of crop identification. However, fully supervised semantic segmentation models rely on a large number of pixel-level labeled samples, while currently there are very few publicly available crop labeled sample sets, and the labeling process is difficult and time-consuming.

[0004] Therefore, how to provide a crop remote sensing identification model and method based on the improved DeepLab V3+, so that the model has higher identification accuracy in high-resolution remote sensing image identification and the identification results better meet the needs of precision agriculture, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an improved DeepLab V3+ crop remote sensing identification model and method to solve the technical problems in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] On one hand, this invention discloses an improved DeepLab V3+ crop remote sensing identification model, comprising:

[0008] The system comprises an input layer, an encoding layer, a decoding layer, and an output layer. The input layer takes a normalized remote sensing image and a vegetation index as input. The encoding layer extracts multi-scale deep semantic feature maps, a first shallow semantic feature map, and a second shallow feature map from the input image. The decoding layer merges the deep semantic features and the first shallow semantic features, performs upper-level sampling, and fuses the second shallow feature map to obtain a fused feature map. The fused feature map is resized and output through the output layer.

[0009] Preferably, in the above-mentioned improved DeepLab V3+ crop remote sensing identification model, the coding layer includes: a Mobilenet V2 network and an ASPP module; the Mobilenet V2 network outputs a deep semantic feature map, a first shallow semantic feature map, and a second shallow feature map; the deep semantic feature map serves as the input to the ASPP module, and outputs a multi-scale deep semantic feature map that is transmitted to the decoding layer; the first shallow semantic feature map and the second shallow feature map incorporate an attention mechanism as the basis for feature fusion.

[0010] Preferably, in the above-mentioned improved DeepLab V3+ crop remote sensing recognition model, the ASPP module includes a 1*1 convolutional layer, three 3*3 convolutional layers with dilation rates of 6, 12, and 18, an average pooling layer, and a feature layer processed by a dual attention mechanism; the output of the 6 parallel processing layers is passed through a 1*1 convolutional layer and connected to the upsampling layer to output a multi-scale deep semantic feature map.

[0011] Preferably, in the above-mentioned improved DeepLab V3+ crop remote sensing recognition model, the first shallow feature map in the decoding layer is fused with the multi-scale deep semantic feature map through an attention module, upsampled through a 3*3 convolutional layer, and merged with the feature map of the second shallow feature map after passing through the attention module. After passing through a 3*3 convolutional layer and an upsampling layer, a fused feature map is obtained.

[0012] On the other hand, this invention discloses a remote sensing identification method for crops. The specific steps of the improved DeepLabV3+ crop remote sensing identification model are as follows:

[0013] Obtain normalized remote sensing images and vegetation indices as data feature sets;

[0014] Using the data feature set and employing the weighted cross-entropy loss function, the improved DeepLabV3+ crop remote sensing identification model is trained to obtain the optimal crop remote sensing identification model.

[0015] Input the data to be tested into the optimal crop remote sensing identification model, and output a fused feature map;

[0016] The classification result is obtained by comparing the original labels with the fused feature map.

[0017] Preferably, in the above-mentioned remote sensing identification method for crops, obtaining the vegetation index includes: extracting four vegetation indices based on the original four bands, namely, normalized vegetation index, green light normalized difference vegetation index, optimized soil-regulated vegetation index, and rapeseed index.

[0018] Preferably, in the above-mentioned remote sensing identification method for crops, the weighted cross-entropy loss function is specifically formulated as follows:

[0019]

[0020] In the formula, N is the total number of pixels, C is the total number of categories, c represents the category of the training pixel, n represents the nth training pixel, and y n,c p is the true value of the nth pixel. n,c This is the predicted value for the nth pixel. N represents the weight of category c. c This represents the total number of pixels in category c.

[0021] As can be seen from the above technical solutions, compared with the prior art, this invention discloses an improved DeepLab V3+ crop remote sensing identification model and method, targeting the identification tasks of wheat and rapeseed. Based on prior knowledge, the DeepLab V3+ model is improved, and the improvement measures based on the traditional DeepLab V3+ model are described in detail. These include adding vegetation index features to the output layer, replacing the Xception backbone network with MobileNet, adding CBAM modules to the upsampling layer and ASPP module, changing the original 16x downsampling to 8x downsampling, and adding an upsampling layer, which can effectively improve the model's identification accuracy. At the same time, reducing the downsampling layer and adding the upsampling layer can better improve the model's crop identification effect. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the model framework of the present invention;

[0024] Figures 2(a)-(c) are schematic diagrams of standard convolution, channel-wise convolution, and point-wise convolution of the present invention;

[0025] Figure 3 This is a schematic diagram of the CBAM structure of the present invention;

[0026] Figure 4 This is a schematic diagram of the spatial attention mechanism structure of the present invention;

[0027] Figure 5 This is a schematic diagram of the channel attention mechanism structure of the present invention;

[0028] Figure 6 This is a frequency distribution diagram of different vegetation types in different indices according to the present invention;

[0029] Figures 7(a)-(b) are comparison charts of Loss curves and MIoU curves for different input layers of the present invention;

[0030] Figures 8(a)-(b) are comparison graphs of Loss curves and MIoU curves for different comparative experiments of the present invention;

[0031] Figure 9 This is a schematic diagram of the MIoU of different comparative experimental models of the present invention on the test set;

[0032] Figures 10(a)-(b) are comparison charts of Loss curves and MIoU curves of different semantic segmentation models of the present invention;

[0033] Figures 11(a)-(c) are schematic diagrams of the original image, labeled image, and prediction results of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] This invention discloses an improved DeepLab V3+ crop remote sensing identification model, such as... Figure 1 As shown, it includes:

[0036] The system comprises an input layer, an encoding layer, a decoding layer, and an output layer. The input layer takes a normalized remote sensing image and a vegetation index as input. The encoding layer extracts multi-scale deep semantic feature maps, a first shallow semantic feature map, and a second shallow feature map from the input image. The decoding layer merges the deep semantic features and the first shallow semantic features, performs upper-level sampling, and fuses the second shallow feature map to obtain a fused feature map. The fused feature map is resized and output through the output layer.

[0037] Furthermore, to make the model more suitable for remote sensing identification of wheat and rapeseed, this embodiment makes the following improvements based on DeepLabV3+: First, a vegetation index is added to the input layer to increase the distinguishability between land features; to reduce computational load, lower memory usage, and improve computational speed, the backbone network Xception is replaced with the lighter MobileNetV2 network; second, to reduce the loss of edge information of wheat and rapeseed due to multiple downsampling, which would make the edges blurry, the downsampling factor of 16x is reduced to 8x, while an upsampling layer is added; to increase the model's sensitivity to wheat and rapeseed areas, a dual attention mechanism (Convolutional Block Attention Module, CBAM) is added to the ASPP module and the upsampling layer; finally, to address the sample imbalance problem among wheat, rapeseed, and other land features, a weighted cross-entropy loss function is introduced.

[0038] To further optimize the above technical solution, the coding layer includes a Mobilenet V2 network and an ASPP module; the Mobilenet V2 network outputs a deep semantic feature map, a first shallow semantic feature map, and a second shallow feature map; the deep semantic feature map serves as the input to the ASPP module, and outputs a multi-scale deep semantic feature map that is transmitted to the decoding layer; the first shallow semantic feature map and the second shallow feature map are enhanced with an attention mechanism as the basis for feature fusion.

[0039] Specifically, the encoding layer: The MobileNetV2 network is a deep separable network, designed to extract deep semantic features from the input image through multiple convolutional and pooling layers, and to reduce computation through channel-wise and pointwise convolution operations. The output of MobileNetV2 is divided into three parts: two are used to preserve shallow features of ground objects during upsampling, with output sizes of 16*64*64 and 24*32*32 respectively; the third is used as input to the ASPP module to further extract multi-scale deep semantic features, with an output size of 320*16*16.

[0040] It's important to understand that MobileNet V2 is an updated version of the original MobileNet architecture, designed to improve efficiency and reduce computational cost. MobileNet V2 is built around the concept of depthwise separable convolutions, which allows the network to learn more complex features with fewer parameters and lower computational cost compared to traditional convolutional networks. Furthermore, MobileNet V2 features an inverted residual structure, which helps increase the network's representational power while maintaining its efficiency.

[0041] Depthwise separable convolution is a method for feature extraction in convolutional neural networks. It consists of two steps: channel-wise convolution and pointwise convolution. In channel-wise convolution, the kernel for each channel processes the features of that channel independently. This method preserves the uniqueness of each channel while reducing the number of parameters and computational complexity. In pointwise convolution, the features of each channel are convolved in separate kernels to capture the spatial information of the entire feature map. This method captures the spatial relationships of the entire feature map and generates a unified feature map. Depthwise separable convolution reduces the number of parameters and computational complexity while preserving the uniqueness and spatial relationships of features.

[0042] With C1 input feature map channels, C2 output feature map channels, and a kernel size of k1*k2, the number of parameters required by the standard convolution operation model is C1*C2*k1*k2. However, the number of parameters required by the depthwise separable convolution operation model is only C1*k1*k2 + 1*1*C1*C2. This is the sum of the number of parameters required for channel-wise convolution (C1*k1*k2) and the number of parameters required for pointwise convolution (1*1*C1*C2). When C1, C2, k1, and k2 are 3, 4, 3, and 3 respectively, the number of parameters required by the standard convolution operation is 108, while the number of parameters required by the depthwise separable convolution operation is 39. It can be clearly seen that using depthwise separable convolution can greatly reduce the amount of computation in the model operation process. Figures 2(a)-(c) show schematic diagrams of standard convolution, channel-wise convolution, and pointwise convolution.

[0043] To further optimize the above technical solution, the ASPP module includes a 1*1 convolutional layer, three 3*3 convolutional layers with dilation rates of 6, 12, and 18, an average pooling layer, and a feature layer processed by a dual attention mechanism. The outputs of the six parallel processing layers are merged into a single convolutional layer, which is connected to the upsampling layer to output a multi-scale deep semantic feature map.

[0044] Specifically, the ASPP module includes six parallel convolutional layers: a 1x1 convolution, three 3x3 convolutions with dilation rates of 6, 12, and 18, an average pooling layer, and a feature layer processed by a dual attention mechanism. The output size of each of the six parallel processing modules is 256x16x16. The outputs of these six parallel processing modules are then merged into a single convolutional layer, resulting in an output size of 1536x16x16. Next, a 1x1 convolution is used to reduce the number of output feature maps, shrinking the 1536 to 256, resulting in an output size of 256x16x16.

[0045] To further optimize the above technical solution, the first shallow feature map in the decoding layer is fused with the multi-scale deep semantic feature map through an attention module, upsampled through a 3*3 convolutional layer, and merged with the feature map of the second shallow feature map after passing through the attention module. After passing through a 3*3 convolutional layer and an upsampling layer, a fused feature map is obtained.

[0046] Decoding layer: The output of ASPP module 256*16*16 is restored to 256*32*32 through linear upsampling, and merged with the 24*32*32 output of MobileNet V2 through attention mechanism and shallow feature map obtained through 1*1 convolution layer. The merged result is then passed through 3*3 convolution to obtain a fused feature map with an output size of 256*32*32.

[0047] The obtained 256*32*32 feature map is linearly upsampled to obtain a 256*64*64 feature map, which is then merged with the 16*64*64 shallow feature map output by MobileNet V2 through an attention mechanism and a 1*1 convolutional layer. The result is then passed through a 3*3 convolutional layer and an upsampling layer to obtain a feature map with an output size of 256*64*64.

[0048] Furthermore, it is important to understand that attention plays a crucial role in human perception. By utilizing localized observation and focusing attention on salient parts, we can better capture visual structures.

[0049] In this embodiment of the invention, the attention module, namely the CBAM module, is composed of a channel attention module and a spatial attention module. By adding the CBAM module, the more important parts of the channel and spatial aspects for the recognition result can be highlighted, thereby improving the accuracy of the recognition result. Figure 3 The diagram shown is a CBAM structure diagram.

[0050] The spatial attention module aims to highlight the spatial location distribution that is important for recognition in the convolutional layer. Specifically, it compresses the channels of the input feature map F using both global average pooling and global max pooling, obtaining... and The two feature maps are merged and reduced to 1 channel by a 7x7 convolution operation. Then, the sigmoid function is used to obtain the spatial attention weight map. The obtained weight map is multiplied by the original feature map F to obtain the spatial attention weighted map F. Sout .

[0051] like Figure 4 The diagram shows the structure of the spatial attention mechanism, and its formula is as follows:

[0052]

[0053] In the formula, σ represents the Sigmoid function, f 7*7 This represents a 7x7 convolution, and ⊕ indicates channel merging. This indicates element-wise multiplication.

[0054] The purpose of the channel attention module is to highlight channels in the convolutional layer that are of significant value for recognition. Specifically, it performs global average pooling and global max pooling to compress the spatial dimension of the feature map F, resulting in... and Next and The inputs are fed into a multilayer perceptron (MLP) and the results are summed. The sum is then fed into a sigmoid function to generate channel attention weights. These weights are multiplied by the original feature map F to obtain the channel attention weighted map F. Cout .

[0055] like Figure 5 The diagram shows the structure of the channel attention mechanism, and its formula is as follows:

[0056]

[0057] On the other hand, this invention discloses a remote sensing identification method for crops, which applies the improved DeepLab V3+ remote sensing identification model for crops described above. The specific steps are as follows:

[0058] Obtain normalized remote sensing images and vegetation indices as a data sample set;

[0059] Using the data sample set and employing the weighted cross-entropy loss function, the improved DeepLabV3+ crop remote sensing identification model is trained to obtain the optimal crop remote sensing identification model;

[0060] Input the data to be tested into the optimal crop remote sensing identification model, and output a fused feature map;

[0061] The classification result is obtained by comparing the original labels with the fused feature map.

[0062] To further optimize the above technical solution, four vegetation indices were extracted from the original four bands: Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Optimal Soil Adjusted Vegetation Index (OSAVI), and Canola Index (CI).

[0063] Specifically, the calculation formulas for the four vegetation indices are shown in Table 1.

[0064] Table 1. Formula for calculating vegetation index

[0065]

[0066] To evaluate the effectiveness of the four indices in distinguishing different vegetation types within the study area, this embodiment extracts the values ​​corresponding to wheat, rapeseed, and woodland sample points from the feature maps of the four indices and studies the differences in their values. For example... Figure 6 The figure shows the frequency distribution of the extracted values ​​of the three vegetation sample points in the four index feature maps. It can be seen from the figure that the peaks in NDVI for wheat, rapeseed, and forest are located at approximately 0.73, 0.55, and 0.52, respectively; the peaks in GNDVI are located at approximately 0.68, 0.52, and 0.47, respectively; the peaks in OSAVI are located at approximately 0.84, 0.63, and 0.59, respectively; and the peaks in CI are located at approximately 0.6*10⁻⁶. 7 1.2*10 7 0.45*10 7 By observing the degree of overlap of the curves and the distribution of the peaks, it can be concluded that the NDVI, GNDVI, and OSAVI indices can effectively distinguish between wheat and rapeseed, and between wheat and forest land; the CI index can effectively distinguish between rapeseed and wheat, and between rapeseed and forest land.

[0067] To further optimize the above technical solution, the weighted cross-entropy loss function is specifically formulated as follows:

[0068]

[0069] In the formula, N is the total number of pixels, C is the total number of categories, c represents the category of the training pixel, n represents the nth training pixel, and y n,c p is the true value of the nth pixel. n,c This is the predicted value for the nth pixel. N represents the weight of category c. c This represents the total number of pixels in category c.

[0070] In practical applications, the impact of different vegetation indices on identification results

[0071] To evaluate the impact of different vegetation indices on the extraction of wheat and rapeseed from GF-2 images during the rapeseed flowering period, this embodiment uses an improved DeepLab V3+ network as the training model. The four vegetation indices were added to the model's input layer (the initial input layer contained only 4 bands). Table 2 shows the model's accuracy performance on the test set after adding different vegetation indices.

[0072] Experimental data reveals that adding different vegetation indices to the four-band model improves the overall accuracy of the model and the recognition accuracy of wheat and rapeseed. Specifically, the impact of different vegetation indices on the model's classification accuracy is evident. Observing the changes in the overall classification accuracy evaluation indicators MIoU and PA, it can be seen that the MIoU obtained using CI and OSAVI is significantly higher than that obtained using NDVI and GNDVI, while the PA values ​​of CI and OSAVI are similar to those of NDVI and are higher than those of GNDVI. Therefore, it can be concluded that adding CI and OSAVI is more effective than NDVI and GNDVI in improving the overall classification accuracy of the model. Furthermore, from the impact of vegetation indices on the classification accuracy of each category, it can be seen that rapeseed shows the best classification accuracy when using CI, followed by OSAVI, both of which outperform NDVI and GNDVI. As for the impact on wheat classification accuracy, the various vegetation indices have a relatively small effect. In summary, the conclusion is that adding CI and OSAVI significantly improves the model's recognition accuracy, followed by NDVI; while GNDVI has the least impact on the model's recognition accuracy.

[0073] Table 2 shows the recognition accuracy with different vegetation features.

[0074]

[0075]

[0076] To improve model performance and enhance the recognition accuracy of wheat and rapeseed, based on the above experiments, CI, OSAVI, NDVI, and GNDVI were added to the input layer in order of their impact on the recognition results. Table 3 shows the model's recognition accuracy when two, three, and four vegetation indices were added, respectively.

[0077] Based on the experimental data, it can be concluded that when only CI and OSAVI are added, the overall recognition accuracy of the model, as well as the recognition accuracy of wheat and rapeseed, reaches the optimal level. Its MIoU and PA reach 85.63% and 95.30%, respectively, while the IoU and F1_Score for wheat and rapeseed are 93.76% and 96.78%, and 74.24% and 85.51%, respectively. In contrast, adding three or four vegetation indices not only does not further improve the model's recognition accuracy but may also increase memory consumption. Therefore, this study chose to add CI and OSAVI to the original four bands as the feature set for this embodiment's experiment and used it as the model's input layer. The input layers of the following experiments all use the feature set with CI and OSAVI added.

[0078] Table 3. Recognition accuracy with different amounts of vegetation indices added.

[0079]

[0080]

[0081] To further evaluate the impact of the feature set construction in this embodiment on model performance, the changes in Loss and MIoU on the validation set during each iteration of the model training process were recorded. As shown in Figures 7(a)-(b), the Loss and MIoU curves of the validation set during model training under two different input layer conditions are shown. Origin uses only the original four bands as the input layer, while origin+VI adds CI and OSAVI features to the original four bands, for a total of six bands, as the input layer.

[0082] The graph showing the variation of the loss function reveals that adding vegetation index features slows down the model's convergence speed, but it still reaches convergence before the end of 200 iterations, and the difference in the loss value after convergence is not significant in either case. The graph showing the variation of MIoU demonstrates that adding vegetation index features significantly improves the model's accuracy on the validation set. Therefore, it can be concluded that the feature set proposed in this embodiment has a significant performance improvement effect on the model.

[0083] To further verify the improved crop recognition performance of the DeepLab V3+ model in this embodiment, an ablation experiment was conducted based on the original DeepLab V3+ model. Five schemes were developed according to the improvement method proposed in this embodiment, as follows:

[0084] Option 1: DeepLab V3+ network with Xception as the backbone network;

[0085] Option 2: DeepLab V3+ network based on Mobilenet V2 as the backbone network;

[0086] Option 3: Based on the DeepLab V3+ architecture of Option 2, add CBAM to the upsampling layer and ASPP layer;

[0087] Option 4: Based on the DeepLab V3+ structure of Option 2, reduce the number of downsampling layers and increase the number of upsampling layers;

[0088] Option 5: Based on the DeepLab V3+ structure of Option 2, add CBAM to the upsampling layer and ASPP layer, while reducing the downsampling layer and increasing the upsampling layer (MyDeepLab V3+).

[0089] This embodiment compares the impact of different backbone networks on model performance in Scheme 1 and Scheme 2, and the results are shown in Table 4. The data in the table shows that replacing the backbone network from Xception to the lightweight MobileNet V2 reduces the model's parameter size from 208.7MB to 27.91MB, a reduction of approximately 7.5 times, significantly reducing memory consumption. The training time is reduced from 265ms to 100ms. Comparing the prediction time for a single image, MobileNet V2 can shorten the prediction time from 10ms to 4ms. Furthermore, the model's accuracy on the test set shows that modifying the backbone network improves the overall accuracy (MIoU) by 4.6%. Therefore, this embodiment chooses MobileNetV2 as the backbone network, which reduces memory consumption, improves training speed, and enhances the model's recognition speed and accuracy.

[0090] Table 4 Comparison of model parameters and recognition accuracy of different backbone networks

[0091]

[0092] To compare the impact of adding an attention mechanism and reducing downsampling while increasing upsampling on model accuracy, this embodiment designed four sets of comparative experiments: Scheme 2, Scheme 3, Scheme 4, and Scheme 5. As shown in Figures 8(a)-(b), the MIoU and loss function curves of these four schemes converged within 200 epochs on the validation set. The method used in this embodiment (Scheme 5) has the highest average test accuracy and convergence accuracy. The model accuracy at convergence is lower than that of the method in this embodiment, but slightly higher than that of the original model (Scheme 2), for both adding only the CBAM module (Scheme 3) and reducing only the downsampling layer while increasing the upsampling layer (Scheme 4). Figure 9The figure shows the MIoU obtained by the models of these four schemes on the test set. It can be seen from the figure that Scheme 5 has the highest MIoU, followed by Schemes 4 and 3, while Scheme 2 has the lowest MIoU. Experimental data shows that adding a CBAM module or reducing downsampling layers and increasing upsampling layers to the original DeepLab V3+ model can effectively improve the model's recognition accuracy. Furthermore, adding a CBAM module while simultaneously reducing downsampling layers and increasing upsampling layers can further improve the model's performance in identifying crops.

[0093] To demonstrate that the improved DeepLab V3+ model in this embodiment has certain advantages in remote sensing identification of wheat and rapeseed, its crop identification accuracy is compared with that of SegNet, U-Net, and the original DeepLab V3+ network, which are widely used in the field of crop remote sensing identification. The identification performance of different models is evaluated from the following two aspects: (1) analyzing the accuracy and convergence of the loss value of each model on the validation set; (2) comparing the average intersection-over-union ratio (MIoU), pixel precision (PA), and intersection-over-union ratio (IoU) and F1 score of each model on the test set.

[0094] Figures 10(a)-(b) show the Loss and MIoU curves of SegNet, U-Net, DeepaLab V3+, and MyDeepLabV3+ on the validation set during training. The figures show that the Loss and MIoU curves of the four models exhibit different trends during the 200 epochs of training. The Loss curves of all four models show an overall decreasing trend and converge within 200 epochs. Therefore, it can be determined that all four models have good convergence. Meanwhile, it can be observed that the MyDeepLab V3+ model proposed in this embodiment has the fastest convergence speed, indicating that the MyDeepLab V3+ model has better training performance. The MIoU curves of the four models exhibit different trends. The trend of the MyDeepLab V3+ model is more stable than that of the SegNet and U-Net models, and the average test accuracy and convergence accuracy of the MyDeepLab V3+ model are the highest among the four models. Therefore, through the analysis of Loss and MIoU on the test set during the training process of the four models, it can be concluded that MyDeepLab V3+ is the best performing model in identifying wheat and rapeseed among the four models.

[0095] To evaluate the generalization ability of the four models, the recognition accuracy obtained by the four models on the test set was compared, and the MIoU, PA, IoU and F1_Score of the four models on the test set were statistically analyzed. Table 5 shows the recognition accuracy of SegNet, U-Net, DeepLab V3+, and the MyDeepLab V3+ model proposed in this embodiment for wheat and rapeseed. The results show that the MyDeepLab V3+ model proposed in this embodiment has the highest overall accuracy (MIoU, PA) and the highest recognition accuracy (IoU, F1_Score) for wheat and rapeseed. Specifically, compared with SegNet, U-Net, and DeepLabV3+, the MIoU of the method in this embodiment is improved by 3.36%, 1.65%, and 1.01%, respectively. Comparing the recognition accuracy of wheat among the four models, the IoU of the method in this embodiment is improved by 1% to 2% compared with the other three models, while the F1_Score improvement is less than 1%. Comparing the rapeseed recognition accuracy of the four models, the IoU of the method in this embodiment is improved by 7.03%, 4.03%, and 2.10% compared to SegNet, U-Net, and DeepLab V3+, respectively; the F1 score is improved by 5.05%, 3.03%, and 1.64% compared to SegNet, U-Net, and DeepLab V3+, respectively. Based on the above comparisons, the MyDeepLab V3+ model proposed in this embodiment has higher model generalization performance and more accurate performance in identifying the planting range of wheat and rapeseed.

[0096] Table 5 Comparison of recognition accuracy of different semantic segmentation models

[0097]

[0098] In summary, applying the MyDeepLab V3+ model proposed in this embodiment to remote sensing identification of wheat and rapeseed achieves higher identification accuracy and extraction results. Therefore, this model was used to identify three regions in the validation set: mountainous areas, planting areas around urban areas, and areas where wheat and rapeseed are concentrated. Figures 11(a)-(c) show a comparison between the original labels and the identification results. The figures show that the model proposed in this embodiment can accurately identify the planting areas of wheat and rapeseed in different scenarios, indicating that the model has good generalization ability even in complex planting structures. Compared to the wheat identification results, the model proposed in this embodiment achieves high accuracy, basically consistent with the actual planting area, but there is still room for improvement in the detailed identification of some small field ridges. Compared to the rapeseed identification results, the model in this embodiment can basically accurately identify the rapeseed planting area without too many errors or omissions, but the preservation of rapeseed plot boundary information differs from the actual planting. Overall, the model proposed in this embodiment has good identification performance in practical applications.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for remote sensing identification of crops, characterized in that, The specific steps are as follows: Obtain normalized remote sensing images and vegetation indexes as a data feature set; An improved DeepLab V3+ crop remote sensing recognition model is trained using the data feature set and a weighted cross-entropy loss function, and an optimal crop remote sensing recognition model is obtained; The to-be-tested data is input into the optimal crop remote sensing recognition model, and a fusion feature map is output; The classification result is obtained by comparing the original label with the fusion feature map; The improved DeepLab V3+ crop remote sensing recognition model comprises: An input layer, an encoding layer, a decoding layer and an output layer; the input layer inputs normalized remote sensing images and vegetation indexes; the encoding layer extracts a multi-scale deep semantic feature map, a first shallow semantic feature map and a second shallow feature map of the input image; the decoding layer combines the deep semantic feature and the first shallow semantic feature, performs upper layer sampling, and fuses the second shallow feature map to obtain a fusion feature map; the fusion feature map is size-adjusted and output through the output layer; The encoding layer comprises a Mobilenet V2 network and an ASPP module; the Mobilenet V2 network outputs a deep semantic feature map, a first shallow semantic feature map and a second shallow feature map; the deep semantic feature map is input into the ASPP module, and a multi-scale deep semantic feature map is output and transmitted to the decoding layer; the first shallow semantic feature map and the second shallow feature map increase the attention mechanism and serve as the basis for feature fusion; The ASPP module includes a 1 1 convolutional layer, three 3 convolutional layers with a hollow rate of 6, 12, and 18 respectively, and one average pooling layer and one feature layer after a double attention mechanism operation; the outputs of the 6 layers of parallel processing pass through a 1 convolutional layer, are connected with the up-sampling layer, and output multi-scale deep semantic feature maps; The first shallow feature map in the decoding layer is fused with the multi-scale deep semantic feature map through an attention module, is up-sampled through a 3 3 convolutional layer, and is combined with the feature map after the second shallow feature map through an attention module, is processed through a 3 3 convolutional layer and an up-sampling layer, and a fused feature map is obtained. The weighted cross-entropy loss function has the following specific formula: ； In the formula N N is the total number of pixels, C is the total number of categories, c represents the category of the training pixel point, n represent the first n training pixel point, y n,c is the true value of the first n pixel point, p n,c is the predicted value of the first n pixel point; represents the weight of category c, N c represents the total number of pixels of category c.

2. The method according to claim 1, wherein, The vegetation index comprises four vegetation indexes extracted on the basis of the original four wave bands, namely, a normalized vegetation index, a green normalized difference vegetation index, an optimized soil-adjusted vegetation index and a rape index.

Images