An image extraction method for winter wheat planting areas combining GF-6 and Sentinel-2

Abstract

The present invention relates to a method for extracting images of winter wheat planting areas by combining GF-6 and Sentinel-2, which solves the defect that the remote sensing image extraction technology for winter wheat planting areas is difficult to meet the actual use requirements compared with the prior art. The present invention includes the following steps: creating a remote sensing image data set; constructing a network for extracting images of winter wheat planting areas; training the network for extracting images of winter wheat planting areas; obtaining the remote sensing image to be extracted; and extracting the image result of the winter wheat planting area. The present invention can more accurately extract the winter wheat planting area from the remote sensing image. By designing and improving the model structure according to the crop spatial distribution information extraction task and the characteristics of high-resolution remote sensing data, good results are achieved in terms of efficiency, speed, applicability and accuracy.

Description

Technical Field

[0001] The invention relates to the field of remote sensing image processing, and in particular to a method for extracting winter wheat planting area images by combining GF-6 and Sentinel-2. Background Art

[0002] Due to the influence of my country's crop planting structure and farming methods, such as the complexity of crop types and the existence of "same species, different spectra" and "different species, same spectra" phenomena caused by regional differences, many difficulties have been brought to crop identification and planting area extraction. The traditional survey method based on sampling and statistical departments reporting agricultural conditions step by step has the disadvantages of being time-consuming and labor-intensive and often lacking spatial information. Satellite remote sensing technology has the characteristics of wide coverage, multi-band, multi-phase and multi-level resolution. While obtaining dynamic information on crop planting areas, it can greatly improve work efficiency and make up for the shortcomings of traditional agricultural monitoring. It is a good data source for obtaining winter wheat spatial distribution data.

[0003] With the development of remote sensing technology, remote sensing images have gradually become the main data source for extracting crop spatial distribution information. Pixel-by-pixel classification technology is the main means of obtaining crop spatial distribution data from remote sensing images. In order to improve the accuracy of pixel-by-pixel classification results, many scholars have conducted a lot of research work.

[0004] Zhang Jiahua et al. (2013) used MODIS–EVI time series data and crop phenology information and the EVI threshold model to extract the corn planting area in Northeast China. Ge Guangxiu et al. (2014) used HJ-1A remote sensing images as the data source and used the normalized difference vegetation index NDVI density segmentation method to extract the winter wheat area in Shuyang County, Jiangsu Province. Wang Limin et al. (2015) used GF-1 satellite as the main data source and used the hierarchical decision tree classification method to extract winter wheat. You Jiong et al. (2016) used GF-1 remote sensing images as the data source and used SVM to extract winter wheat information. Zhang Sha et al. (2018) used MODIS-EVI data combined with the spectral mutation method to extract the winter wheat area in the Huanghuaihai Plain.

[0005] The above studies have successfully extracted the planting area and spatial distribution data of crops such as winter wheat, but they mainly use traditional supervised classification methods and unsupervised classification methods, which can only extract low-level features such as color, shape, and texture of the image, but cannot extract higher-level semantic features. When applied to high spatial resolution images for feature extraction, the obtained features have poor distinguishing ability, which often leads to unsatisfactory final classification results.

[0006] Deep convolutional neural networks have achieved great success in many fields and have demonstrated excellent performance in many applications. This trend has also attracted many researchers to apply deep convolutional neural networks to the field of remote sensing image semantic segmentation.

[0007] Zhang Meng et al. (2018) used the method of image block division to train a convolutional neural network based on multi-temporal Landsat 8 data, so as to extract the spatial distribution information of rice. Jiao Jihan et al. (2018) proposed an improved AlexNet model for rapeseed planting area extraction. Liu Xiangnan et al. (2018) used a convolutional neural network to extract the rice planting area and fine-tuned the pre-trained convolutional neural network model using a transfer learning strategy. The experimental results showed that the segmentation accuracy of this method was better than that of the support vector machine. Li Qianjing et al. (2021) proposed a convolutional neural network RE-CNN remote sensing image crop extraction method suitable for the red-edge band of GF-6 WFV and achieved good results. Kussul et al. (2017) used multi-temporal image data of Landsat 8 and Sentinel-1A and used a convolutional neural network model for crop extraction. Compared with the multi-layer perceptron and random forest methods, this method could better distinguish crop types. Zhong et al. (2019) used Landsat enhanced vegetation index time series data and used a convolutional neural network to classify summer crops. The results showed that this method was better than the extraction results of random forest and support vector machine.

[0008] The convolutional neural network has made great progress in the task of extracting crop spatial distribution information, proving the effectiveness of the convolutional neural network in the task of extracting crop spatial distribution information. However, there are also some problems. Most of the current research is based on the standard convolutional neural network structure for crop extraction, and less attention is paid to designing and improving the model structure according to the characteristics of the crop spatial distribution information extraction task and high-resolution remote sensing data. Moreover, the current extraction methods have limitations in terms of efficiency, speed, applicability, and accuracy.

[0009] Therefore, how to implement a remote sensing image extraction method for winter wheat planting areas based on convolutional neural network technology has become an urgent technical problem to be solved. Summary of the Invention

[0010] The purpose of the present invention is to solve the defect that the existing remote sensing image extraction technology for winter wheat planting areas is difficult to meet the actual use needs, and to provide a method for extracting winter wheat planting area images by combining GF-6 and Sentinel-2 to solve the above problems.

[0011] In order to achieve the above purpose, the technical solution of the present invention is as follows:

[0012] A method for extracting winter wheat planting area images by combining GF-6 and Sentinel-2, comprising the following steps:

[0013] 11) Creation of remote sensing image dataset: Obtain the GF-6 remote sensing images with 8m resolution and Sentinel-2 remote sensing images with 10m resolution and perform preprocessing to form a remote sensing image dataset;

[0014] 12) Construction of winter wheat planting area image extraction network: Using the UNet network as the basic network, embed the CBAM attention module in the basic convolutional units of its encoding layer and decoding layer to construct an improved UNet network structure as the winter wheat planting area image extraction network;

[0015] 13) Training of winter wheat planting area image extraction network: Input the preprocessed remote sensing image dataset into the improved UNet network structure for training to obtain the trained winter wheat planting area image extraction network;

[0016] 14) Acquisition of remote sensing image to be extracted: Obtain the remote sensing image to be extracted and perform preprocessing;

[0017] 15) Extraction of winter wheat planting area image results: Input the preprocessed remote sensing image to be extracted into the trained winter wheat planting area image extraction network to obtain the winter wheat planting area image extraction results.

[0018] The construction of the winter wheat planting area image extraction network includes the following steps:

[0019] 21) Extract features based on the UNet network as the basic framework. Set the Unet model to include an encoder, a decoder, and a skip connection part;

[0020] 22) Set that the image feature information extracted by the encoder through the convolutional layer consists of a 3×3 convolutional layer, a ReLU function, and a 2×2 max pooling layer; perform four downsamplings; after each pooling operation, the feature image becomes smaller and the number of channels doubles;

[0021] The decoder performs upsampling through a 2×2 transposed convolution. The decoder part completes four upsamplings; after each upsampling, the size of the feature image increases and the number of channels is reduced by half, and then the low-level details and high-level semantics of the feature map are combined through skip connection operations;

[0022] 23) Set to input the features extracted by convolution and max pooling in the UNet structure into the embedded attention mechanism CBAM module;

[0023] 24) Set the CBAM attention module:

[0024] Use the CBAM attention module to generate the feature map F ∈ R C×H×W Calculate the channel attention M C ∈ R C ×1×1 ;

[0025] Multiply F and M C (F) element-wise multiplication to obtain F', and then calculate the spatial attention M for F' S ∈R 1×H×W , multiply F' and M S (F) element-wise multiplication to obtain the output result F” of the CBAM module;

[0026] Among them, M C (F) The formula representation includes:

[0027]

[0028] Among them, C, H, and W are the number of channels, height, and width of the feature map respectively, δ is the Sigmoid function, W 0 ∈R C / r×C , W 1 ∈R C ×C / r , r is the parameter reduction rate, W 0 Then use Relu as the activation function; and respectively represent average and max pooling of F in the spatial dimension, M C (F) is the channel attention map;

[0029] Spatial attention M S (F) The calculation formula includes:

[0030]

[0031] Among them, f 7×7 represents a 7×7 convolutional layer, and respectively represent max and average pooling of F in the channel dimension.

[0032] The training of the improved UNet network structure includes the following steps:

[0033] 31) Determine the hyperparameters during the training process, and initialize the parameters of the CBAM-UNet model. The Batch size is 4, the learning rate is 0.0001, the parameters of the Adam optimizer are Beta1 = 0.5, Beta2 = 0.999, and Epochs = 100;

[0034] 32) Respectively take the images and labels in the GF-6 image dataset and the Sentinel-2 image dataset as the training dataset and input them into the winter wheat planting area image extraction network;

[0035] 33) Use the winter wheat planting area image extraction network to perform forward propagation on the current training data;

[0036] 34) Calculate the loss and backpropagate it to the winter wheat planting area image extraction network;

[0037] 35) Use the Adam optimizer to update the parameters of the winter wheat planting area image extraction network according to the loss value, and repeat steps 32)-34) until the loss is less than a predetermined threshold.

[0038] Beneficial effects

[0039] A method for extracting winter wheat planting area images by combining GF-6 and Sentinel-2 according to the present invention can extract the winter wheat planting area from remote sensing images more accurately compared with the prior art. By designing and improving the model structure for the crop spatial distribution information extraction task and the characteristics of high-resolution remote sensing data, good results are achieved in terms of efficiency, speed, applicability, and accuracy.

[0040] Based on the improvement of UNet, the present invention embeds an attention module in the basic convolutional units of its encoding layer and decoding layer. It has good scalability and high extraction accuracy, and can adjust parameters to be applied to remote sensing satellite images of different sensors. Through verification, the high-resolution multi-source remote sensing image classification results obtained by the present invention are superior to the comparative classification algorithm in terms of evaluation indicators, and the generated classification results can better maintain the smoothness and integrity of the edges. Brief description of the drawings

[0041] Figure 1 It is the sequence diagram of the method of the present invention;

[0042] Figure 2a It is the image of the GF-6 dataset;

[0043] Figure 2b It is Figure 2a the corresponding label map;

[0044] Figure 3a It is the image of the Sentinel-2 dataset;

[0045] Figure 3b It is Figure 3a the corresponding label map;

[0046] Figure 4 It is the prediction result map obtained by using the GF-6 dataset as the training set;

[0047] Figure 5 It is the prediction result map obtained by using the Sentinel-2 dataset as the training set. Detailed implementation manners

[0048] To further understand and recognize the structural features and achieved effects of the present invention, the following is a detailed description with preferred embodiments and accompanying drawings:

[0049] As Figure 1 shown, a method for extracting winter wheat planting area images by combining GF-6 and Sentinel-2 according to the present invention includes the following steps:

[0050] First step, creation of a remote sensing image data set. Obtain the GF-6 remote sensing image with a resolution of 8m and the Sentinel-2 remote sensing image with a resolution of 10m and perform preprocessing to form a remote sensing image data set.

[0051] In practical applications, first fuse the 2m panchromatic image and the 8m multispectral data image of GF-6 to obtain 2m multispectral data as a reference basis for manual annotation. For example, Zhengding County, Shijiazhuang City can be selected as the training set, and Zengcun Town, Gaocheng City can be selected as the test set. Using GIS software, manually draw the boundaries of the winter wheat planting areas on the GF-6 and Sentinel-2 remote sensing images in the study area respectively; using the ROI (Region of Interesting) tool in ENVI software, based on the boundaries obtained above, obtain the pixel numbers within the ROI area and perform pixel-by-pixel marking. Use Photoshop software to divide the remote sensing image and the marking results into blocks in the order from left to right and from top to bottom. The size of each block is 256×256 pixels. The image block and its corresponding label file form an image-label pair.

[0052] Second step, construct a network for extracting winter wheat planting area images: use the UNet network as the basic network, and embed the CBAM attention module in the basic convolutional units of its encoding layer and decoding layer to construct an improved UNet network structure as the network for extracting winter wheat planting area images.

[0053] Most current research is based on the standard convolutional neural network structure for crop extraction, and there are few model structures designed and improved specifically for the crop planting area extraction task and the characteristics of high-resolution remote sensing data.

[0054] Based on the research on extracting crop spatial distribution information by convolutional neural networks, for the winter wheat planting area extraction task, this invention analyzes the characteristics of winter wheat in high-resolution remote sensing images: 1) Remote sensing images are different from natural images with high resolution and rich details. The detailed information of winter wheat patches in the images is much lower than that of natural images; 2) The sizes of winter wheat planting areas are different. When the planting area is large, the number of pixels occupied by winter wheat patches is relatively large, while when the planting area is small, the winter wheat patches only contain dozens of pixels or even fewer. There are situations where the sizes of winter wheat patches in remote sensing images are different. Therefore, an attention mechanism is introduced into the UNet network. By placing more attention on informative features, a large number of intermediate features extracted from the hidden layer are reduced, effectively improving the extraction accuracy of the winter wheat planting area and providing theoretical and technical methods for the extraction of other crop planting areas.

[0055] The specific steps are as follows:

[0056] (1) Based on the UNet network as the basic framework to extract features, it is set that the Unet model includes an encoder, a decoder, and a skip connection part.

[0057] (2) It is set that the image feature information extracted by the encoder through the convolutional layer consists of a 3×3 convolutional layer, a ReLU function, and a 2×2 max pooling layer; four downsamplings are performed; after each pooling operation, the feature image becomes smaller and the number of channels doubles;

[0058] The decoder performs upsampling through a 2×2 transposed convolution, and the decoder part completes four upsamplings; after each upsampling, the size of the feature image increases and the number of channels is reduced by half, and then the low-level details and high-level semantics of the feature map are combined through skip connection operations. The shallow network can more effectively preserve detailed position information and assist in segmentation by cascading the corresponding feature patterns of the encoder and decoder.

[0059] (3) It is set to utilize the features extracted by the convolutional layer, upsampling, and downsampling in the encoder or decoder in the UNet structure and input them into the embedded attention mechanism CBAM module.

[0060] (4) Set the CBAM attention module:

[0061] Use the CBAM attention module to generate the feature map F∈R of the UNet network C×H×W Calculate the channel attention M C ∈R C ×1×1 ;

[0062] Multiply F and M C (F) element-wise to obtain F’, and then calculate the spatial attention M for F' S ∈R 1×H×W , multiply F' and M S (F) element-wise to obtain the output result F” of the CBAM module;

[0063] Among them, M C (F) The formula representation includes:

[0064]

[0065] Among them, C, H, and W are the number of channels, height, and width of the feature map respectively, δ is the Sigmoid function, W 0 ∈R C / r×C , W 1 ∈R C ×C / r , r is the parameter reduction rate, and W 0 Then use Relu as the activation function; and respectively represent the use of average and max pooling on F in the spatial dimension, M C (F) is the channel attention map;

[0066] Spatial attention M S (F) calculation formula includes:

[0067]

[0068] where f 7×7 represents a 7×7 convolutional layer, and respectively represent the use of max and average pooling on F in the channel dimension.

[0069] Step 3, training of the winter wheat planting area image extraction network: Input the preprocessed remote sensing image dataset into the improved UNet network structure for training to obtain the trained winter wheat planting area image extraction network.

[0070] The training of the improved UNet network structure includes the following steps:

[0071] (1) Determine the hyperparameters during training and initialize the parameters of the CBAM-UNet model. The Batch size is 4, the learning rate is 0.0001, the parameter Beta1 of the Adam optimizer is 0.5, Beta2 is 0.999, and Epochs is 100.

[0072] (2) Input the images and labels in the GF-6 image dataset and Sentinel-2 image dataset into the winter wheat planting area image extraction network as the training dataset respectively.

[0073] (3) Use the winter wheat planting area image extraction network to perform forward propagation on the current training data.

[0074] (4) Calculate the loss and backpropagate it to the winter wheat planting area image extraction network.

[0075] (5) Use the Adam optimizer to update the parameters of the winter wheat planting area image extraction network according to the loss value, and repeat steps (2)-(4) until the loss is less than a predetermined threshold.

[0076] Step 4, acquisition of the remote sensing image to be extracted: Acquire the remote sensing image to be extracted and perform preprocessing.

[0077] Use the Photoshop software to divide the remote sensing images and labeling results of Gaocheng City, Shijiazhuang City into blocks in the order from left to right and from top to bottom. The size of each block is 256×256 pixels. The image block and its corresponding label file form an image-label pair. Randomly select an image-label pair from each of the two datasets, such as Figure 2a , 2b , 3a, 3b.

[0078] Step 5, extraction of the winter wheat planting area image result: Input the preprocessed remote sensing image to be extracted into the trained winter wheat planting area image extraction network to obtain the winter wheat planting area image extraction result.

[0079] The following further illustrates the effect of the present invention in combination with simulation experiments:

[0080] The computer hardware environment for the experiments of the present invention is an Intel Xeon Gold 6248R processor, 192 Gb of memory, an NVIDIA Quadro P4000 graphics card, the GPU acceleration library uses CUDA 10.0, and the deep learning framework uses Pytorch. All subsequent training and testing experiments are based on this platform.

[0081] To verify the effectiveness of the proposed CBAM-UNet method for remote sensing image classification of wheat planting areas, the GF-6 image dataset and the Sentinel-2 image dataset are used. The present invention uses five popular criteria, namely accuracy, mean intersection over union, recall, overall accuracy, and F 1 value to evaluate the performance of the proposed model.

[0082] Content and result analysis of the simulation experiment:

[0083] SegNet and DeepLabv3+ are classical semantic segmentation models for images and have achieved good results in camera image processing. In addition, the working principles of these two models are similar to the research in this article. Therefore, these two models are selected as comparison models to better reflect the advantages of the UNet-CBAM model in classification. The SegNet, DeepLabv3+, UNet, and UNet-CBAM models are trained using the GF-6 image dataset and the Sentinel-2 image dataset respectively.

[0084] After the model is trained, input the Sentinel-2 test set images into the model. The model automatically extracts features from the images and predicts each pixel to determine its type, and finally obtains the classification result.

[0085] To compare the performance of the four models, namely SegNet, DeepLabv3+, UNet, and UNet-CBAM, the experimental results of two representative small regions were selected for comparison. One small region is mainly farmland, and the other region is mixed with facility agriculture and buildings. These two regions can represent the land use structure of the experimental area.

[0086] Figure 4 and Figure 5 shows two images selected from the test images and the corresponding results using the four methods. It can be seen that the UNet-CBAM model only misclassified a small number of pixels in the corner of the winter wheat planting area. In the SegNet results, Deeplabv3+ results, UNet results, and UNet-CBAM results, the misclassified pixels are mainly distributed at the junction of the winter wheat and non-winter wheat areas, including the edges and corners. The number of misclassified pixels in the UNet-CBAM model results is less than that in DeepLabv3+.

[0087] The error of the SegNet results is the largest, and most of the misclassified pixels are located at the edges, corners, and planting areas. Among them, in the areas with a large growth amount of winter wheat, the shape extracted by the UNet-CBAM algorithm fits well with the actual area, and there are more edge errors in other algorithms (the areas in columns c, d, and e). The UNet-CBAM algorithm shows better performance than other algorithms when processing images.

[0088] Table 1 shows the confusion matrices of the segmentation results of the models trained with the GF-6 image dataset and the Sentinel-2 image dataset respectively. Each row of the confusion matrix represents the proportion of the actual category, and each column represents the proportion of the predicted category. UNet-CBAM achieved better classification results on both the GF-6 image dataset and the Sentinel-2 image dataset. The average proportions of "winter wheat" misclassified as "non-winter wheat" are 0.052 and 0.031 respectively, and the average proportions of "non-winter wheat" misclassified as "winter wheat" are 0.029 and 0.028 respectively.

[0089] In the confusion matrices of the four models, there is almost no confusion between winter wheat and urban areas, which may be due to the different characteristics of the two land use types. However, the confusion between winter wheat and farmland is serious because the growth conditions of most winter wheat areas misclassified as farmland are poor. In these areas, their characteristics are similar to those of winter farmland, which leads to a high probability of misclassification. There is also a certain degree of confusion between winter wheat and forest areas. This is because some trees are still green in winter, similar to the characteristics of winter wheat areas. However, in this case, due to the use of texture and high-level semantic information, the degree of confusion is significantly lower than that of farmland.

[0090] Table 1: Confusion Matrix Comparison Table for Winter Wheat Classification

[0091]

[0092]

[0093] In the present invention, five popular criteria, namely accuracy, mean intersection over union, recall, overall accuracy, and F1-score, are used to evaluate the performance of the proposed model. Table 2 shows the evaluation criterion values of four models on the GF-6 image dataset and the Sentinel-2 image dataset respectively.

[0094] Table 2: Evaluation Table of Experimental Results

[0095]

[0096] Through the statistical analysis of the experimental results, we can intuitively observe that the five evaluation indicators of the improved model (UNet-CBAM) are all better than the original model. The recall and mean intersection over union have increased, the F1-score and accuracy have increased to a certain extent, and the overall accuracy has increased slightly. Compared with SegNet, DeepLabv3+, and UNet, UNet-CBAM performs best in terms of accuracy, MioU, overall accuracy, and F1-score. 1 In addition, the accuracy and mean intersection over union of UNet-CBAM are significantly higher than those of UNet, being 4.21% and 2.94% respectively on the GF-6 dataset, and 4.07% and 2.9% respectively on the Sentinel-2 dataset. This means that the introduction of the attention mechanism is effective in the improvement of the UNet network. 1

[0097] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection required by the present invention is defined by the appended claims and their equivalents.​

Claims

1. An image extraction method for winter wheat planting areas combining GF-6 and Sentinel-2, characterized in that, it includes the following steps: 11) Creation of a remote sensing image dataset: Obtain the GF-6 remote sensing image with an 8m resolution and the Sentinel-2 remote sensing image with a 10m resolution and perform preprocessing to form a remote sensing image dataset; 12) Construction of an image extraction network for winter wheat planting areas: Use the UNet network as the basic network, and embed the CBAM attention module in the basic convolutional units of its encoding layer and decoding layer to construct an improved UNet network structure as the image extraction network for winter wheat planting areas; The construction of the image extraction network for winter wheat planting areas includes the following steps: 121) Extract features based on the UNet network as the basic framework, and set that the Unet model includes an encoder, a decoder, and a skip connection part; 122) Set that the image feature information extracted by the encoder through the convolutional layer consists of a 3×3 convolutional layer, a ReLU function, and a 2×2 max pooling layer; perform four downsamplings; after each pooling operation, the feature image becomes smaller and the number of channels doubles; The decoder performs upsampling through a 2×2 transposed convolution, and the decoder part completes four upsamplings; after each upsampling, the size of the feature image increases and the number of channels is reduced by half, and then the low-level details and high-level semantics of the feature map are combined through skip connection operations; 123) Set to use the features extracted by convolution and max pooling in the UNet structure and input them into the embedded attention mechanism CBAM module; 124) Set the CBAM attention module: Using the CBAM attention module for the feature map F ∈ R generated by the UNet network C×H×W Calculate the channel attention M C ∈ R C×1×1 ; Multiply F and M element-wise C (F) Multiply element-wise to obtain F', and then calculate the spatial attention M for F' S ∈R 1×H×W , multiply F' and M S (F) Multiply element-wise to obtain the output result F” of the CBAM module; Among them, M C (F) The formula representation includes: Among them, C, H, and W are the number of channels, height, and width of the feature map respectively, δ is the Sigmoid function, W 0 ∈R C / r×C , W 1 ∈R C×C / r , r is the parameter reduction rate, and W 0 uses Relu as the activation function later; and respectively represent using average and max pooling on F in the spatial dimension, and M C (F) is the channel attention map; Spatial attention M S (F) The calculation formula includes: Among them, f 7×7 represents a 7×7 convolutional layer, and respectively represent using max and average pooling on F in the channel dimension; 13) Training of the image extraction network for winter wheat planting areas: Input the preprocessed remote sensing image dataset into the improved UNet network structure for training to obtain the trained image extraction network for winter wheat planting areas; 14) Acquisition of the remote sensing image to be extracted: Obtain the remote sensing image to be extracted and perform preprocessing; 15) Extraction of the image result of the winter wheat planting area: Input the preprocessed remote sensing image to be extracted into the trained image extraction network for the winter wheat planting area to obtain the image extraction result of the winter wheat planting area.

2. An image extraction method for winter wheat planting areas combining GF-6 and Sentinel-2 according to claim 1, characterized in that, the training of the improved UNet network structure includes the following steps: 21) Determine the hyperparameters during the training process and initialize the parameters of the CBAM-UNet model. The Batch size is 4, the learning rate is 0.0001, the parameters of the Adam optimizer are Beta1 = 0.5, Beta2 = 0.999, and Epochs = 100; 22) Respectively input the images and labels in the GF-6 image dataset and the Sentinel-2 image dataset as the training dataset into the image extraction network for winter wheat planting areas; 23) Use the image extraction network for winter wheat planting areas to perform forward propagation on the current training data; 24) Calculate the loss and backpropagate it to the image extraction network for winter wheat planting areas; 25) Use the Adam optimizer to update the parameters of the winter wheat planting area image extraction network according to the loss value, and repeat steps 22)-24) until the loss is less than a predetermined threshold.

Images