A method for extracting wheat planting areas from remote sensing images based on SE-UNet deep learning network

Abstract

The present invention relates to a method for extracting wheat planting areas from remote sensing images based on a SE-UNet deep learning network. Compared with the prior art, the method solves the problem of limitations in efficiency and accuracy in extracting wheat areas from remote sensing images. The present invention comprises the following steps: obtaining basic data; constructing a wheat planting area extraction network; training a wheat planting area extraction network; obtaining remote sensing images to be extracted; and extracting wheat planting areas from remote sensing images. The present invention can improve the accuracy and efficiency of wheat area extraction from remote sensing images, and can provide a basis for statistics of large-scale winter wheat planting areas.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting wheat planting areas from remote sensing images based on a SE-UNet deep learning network. Background Art

[0002] Remote sensing technology, due to its large-area synchronous observation, low cost, strong data comprehensiveness, and ability to monitor dynamic space and time, has largely avoided the influence of human factors and has been widely used in various aspects of agricultural remote sensing monitoring, especially in research on crop planting area extraction, crop growth monitoring, crop yield estimation, and crop disease and pest monitoring.

[0003] At present, the images used for crop planting area monitoring include high-, medium- and low-resolution images. For example, Huang Jianxi et al. used time-series MODIS data as the data source to extract the spatial distribution of winter wheat. However, since the data source is medium- and low-resolution remote sensing images, it cannot meet the application requirements well. Li Xiaohui et al. used multi-temporal Landsat-8OLI images, combined with spectral angle mapping and decision tree classification to extract the classification of major crops in the eastern part of Xinrong District, Datong City, with an overall accuracy of 85.34%; He Yuan Keiko et al. used the random forest algorithm, combined with the spectral characteristics, texture characteristics and principal component characteristics of typical winter wheat planting areas, to achieve feature selection and rapid extraction of winter wheat plots on Landsat-8OLI images, with an overall accuracy of 84.85%; Zhao Ye et al. used the HSV color space transformation method based on Landsat-8OLI images, and extracted the features of winter wheat plots on H-NDVI. The difference was calculated and a reasonable threshold was set to distinguish winter wheat from non-winter wheat. Finally, the winter wheat planting area in the central and southern part of Henan Province was extracted. Wang Dongli et al. used the normalized difference vegetation index (NDVI) as the winter wheat information identification indicator, and extracted the winter wheat planting area in Xinji City, Hebei Province based on GF-1 satellite data and unsupervised classification model. Wang Rong et al. used Sentinel-2A images as the data source, selected three feature variables including principal component analysis (PCA), red edge normalized difference vegetation index (RENDVI), and texture features, and used the random forest algorithm combined with the digital elevation model (DEM) to extract the winter wheat planting area.

[0004] However, traditional supervised and unsupervised classification methods can only extract low-level features of images such as color, shape, and texture, but cannot extract higher-level semantic features. Their classification results are easily affected by phenomena such as the same object with different spectra or different objects with the same spectra.

[0005] In recent years, deep learning has made a large number of key breakthroughs in image processing, natural language processing, data analysis and other fields, becoming one of the most watched scientific research today.

[0006] Long et al. creatively proposed a fully convolutional neural network (FCN) based on CNN, which uses convolutional layers to replace the fully connected layers in CNN for pixel-by-pixel classification to achieve image segmentation; Badrinarayanan et al. proposed a symmetrical encoder-decoder network SegNet, which follows the idea of ​​FCN for image semantic segmentation, uses an encoder to extract features from the input image, and then uses a decoder to restore the resolution of the feature map, and finally uses the Softmax function to classify the feature map at the pixel level; Ronneberger et al. subsequently proposed a new FCN network, namely UNet, which has a symmetrical U-shaped structure. On the basis of SegNet, a cascade between the encoder and the decoder is introduced, and the high and low semantic features of the network are integrated to achieve better segmentation effects. At present, many scholars are also trying to apply deep learning to the field of remote sensing and explore its applicability in remote sensing classification. Fu et al. used the FCN network to classify and identify roads, vegetation, buildings, and water bodies in high-resolution remote sensing images, with a correct segmentation accuracy of more than 85%; Huang et al. used Sentinel-2 remote sensing images to improve SegNet and extracted the peanut planting area in Xincai County, Henan Province, with a test accuracy of 83.3%; Chen et al. improved SegNet for extracting buildings in remote sensing images and tested it on the Inria dataset, with good overall performance; Zhu et al. used GF-2 and BJ-2 remote sensing images to block-fuse multi-temporal remote sensing images and used the Deeplabv3+ network to extract sugarcane fields with an accuracy of 94.32%; Du et al. used the UNet network to classify rice in Arkansas, USA based on multi-temporal Landsat data, and the model performed well; Zheng et al. used GF-2 remote sensing images and the UNet network to segment five types of land features, with an overall test accuracy of 82.27%. However, UNet cannot solve the problem of edge information loss caused by multiple downsampling and upsampling.

[0007] Therefore, how to combine deep learning technology with remote sensing technology to realize image extraction of wheat-growing areas has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] The purpose of the present invention is to solve the defects of the prior art in the extraction of wheat regions from remote sensing images in terms of efficiency and accuracy, and to provide a method for extracting wheat planting areas from remote sensing images based on a SE-UNet deep learning network to solve the above problems.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows:

[0010] A method for extracting wheat planting areas from remote sensing images based on a SE-UNet deep learning network comprises the following steps:

[0011] Acquisition of basic data: Landsat-8OLI remote sensing image data was obtained and preprocessed. The wheat area vector was outlined using ArcGIS software, and the wheat area in the Landsat-8OLI image was marked. The original image data size was 256×256×3, and the label data size was 256×256×1. The preprocessed remote sensing image data was obtained.

[0012] Construction of wheat planting area extraction network: Construction of wheat planting area extraction network based on SE-UNet network;

[0013] Training of wheat planting area extraction network: Use pre-processed remote sensing image data to train the wheat planting area extraction network;

[0014] Acquisition of remote sensing images to be extracted: obtaining remote sensing images to be extracted and preprocessing them;

[0015] Extraction of wheat planting areas from remote sensing images: The preprocessed remote sensing images to be extracted are input into the trained wheat planting area extraction network to extract the wheat planting areas from the remote sensing images.

[0016] The construction of the wheat planting area extraction network includes the following steps:

[0017] Build a UNet network model. The UNet network adopts a symmetrical encoder-decoder structure and uses the UNet network as the basic framework to extract features.

[0018] The encoder is set to extract feature information through 3×3 convolution, Relu activation function and 2×2 pooling. After pooling, the number of feature map channels remains unchanged and the size is reduced to half of the original size. A total of 4 downsamplings are performed with a downsampling rate of 16 times. The BN layer before the Relu activation function is used to suppress overfitting problems.

[0019] The decoder is set to perform 4 upsamplings through transposed convolution to restore the feature map to the original resolution. At the same time, the high-level semantic features are integrated with the shallow features by using skip links at each level. Finally, Softmax is used as the classification layer to segment the image pixels into two categories, one is the background and the other is the extracted wheat area.

[0020] The SE module is added after each double-layer convolution module. The SE module is used to analyze the relationship between channels in the feature map and provide corresponding weights for each channel. It automatically obtains the importance of each feature channel through model learning.

[0021] Set up the SE module.

[0022] The training of the wheat planting area extraction network includes the following steps:

[0023] The preprocessed remote sensing image data was input into the wheat planting area extraction network. During training, the Adamoptimizer function was selected as the parameter optimizer, the cross entropy function was used as the loss function, the initial learning rate was 0.0001, the number of training iterations was 100, and the step size was 8;

[0024] The feature extraction is performed through the convolution and pooling of the UNet encoder in the wheat planting area extraction network. The decoder uses the feature map generated by upsampling to fuse with the feature map on the left. In the process, the importance of each feature channel is automatically obtained through the SE module after each double-layer convolution module to strengthen the channels with strong feature expression capabilities.

[0025] After two more convolution operations, a feature map is generated;

[0026] A 1×1 convolution is used to adjust the channels of the final feature layer to the number of categories.

[0027] The setting of the SE module comprises the following steps:

[0028] The SE module is set to include compression and excitation operations. The compression operation performs global average pooling on the feature map of size W×H×C and compresses it into a 1×1×C vector. The excitation operation uses a fully connected neural network to perform a nonlinear transformation on the result after the compression operation. Finally, the result obtained by the excitation operation is used as a weight and multiplied by the input feature.

[0029] Suppose the output feature of the network is U, and the size of U is (H, W, C), where H, W, and C are height, width, and number of channels respectively;

[0030] The c-th dimension feature map u c Perform compression operation, denoted as F sq , output z c , all feature maps are compressed and output as z, the shape of z is 1×1×C, and the calculation formula is as follows:

[0031]

[0032] Among them, F sq Indicates the compression operation on the feature, u c is the c-th dimension feature of the input feature map, W and H represent the width and height of the feature, i and j are the coordinates of the feature map in the spatial dimension;

[0033] Perform an excitation operation on z, denoted as F ex , the result is s, and the calculation formula is as follows:

[0034] s=F ex (z,w)=σ(w2,δ(w1,z)),

[0035] Where w, w1 and w2 are the weights in the fully connected layer, σ and δ represent the Sigmoid and Relu activation functions respectively;

[0036] The channel weights are multiplied and recorded as F scale , that is, weighting s to the features of each channel, the output is The calculation formula is as follows:

[0037]

[0038] Among them, F scale (u c ,s c ) represents the scalar s c With feature u c The corresponding channel product between .

[0039] Beneficial Effects

[0040] The invention discloses a remote sensing image wheat planting area extraction method based on SE-UNet deep learning network, which can improve the accuracy and efficiency of remote sensing image wheat area extraction compared with the prior art, and can provide a basis for large-scale winter wheat planting area statistics.

[0041] The present invention uses the UNet encoder part to extract feature information by downsampling through convolution and pooling, and the decoder also performs 4 upsampling through transposed convolution to restore the feature map to the original image resolution. At the same time, based on UNet, an SE module is added after each double-layer convolution, which automatically obtains the importance of each feature channel through model learning, reduces the weak feature expression ability, and strengthens the channels with strong feature expression ability.

[0042] The wheat planting area extraction algorithm proposed in the present invention is superior to the comparative segmentation model and has a significant improvement on the problem of missing edge information. It performs well under the tests of remote sensing data sets with different resolutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a method sequence diagram of the present invention;

[0044] Figure 2 The prediction results of different networks on the Landsat-8OLI test set for GF-6 training images;

[0045] Figure 3The prediction results of different networks for Sentinel-2 training images on the Landsat-8OLI test set;

[0046] Figure 4 The prediction results of different networks for Landsat-8OLI training images on the Landsat-8OLI test set. DETAILED DESCRIPTION

[0047] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0048] like Figure 1 As shown, the method for extracting wheat planting areas from remote sensing images based on the SE-UNet deep learning network of the present invention comprises the following steps:

[0049] The first step is to obtain basic data. Landsat-8OLI remote sensing image data is obtained and preprocessed in the traditional way as needed. The wheat area vector is outlined using ArcGIS software, and the wheat area in the Landsat-8OLI image is marked. The original image data size is 256×256×3, and the label data size is 256×256×1, and the preprocessed remote sensing image data is obtained.

[0050] ArcGIS software was used to open the remote sensing image, and the wheat area vector was manually delineated on the Landsat-8OLI image in accordance with GF-6, and the label was generated by outputting it in PNG format. The raster image was a single-channel 8-bit PNG image, in which the wheat pixel value was 1 and the non-wheat (background) pixel value was 0, and then it was cropped into sample label data of 256×256 size. At the same time, the original remote sensing image was saved and output in JPG format. The image was a 3-channel 8-bit JPG image, and then it was cropped into original image data of 256×256 size. The original image data corresponded to the label data file name one by one, and random data expansion operations such as flipping, rotation, and scale transformation were performed.

[0051] The second step is to construct a wheat planting area extraction network: construct a wheat planting area extraction network based on the SE-UNet network.

[0052] For a long time, the estimation of wheat planting area has been through layer-by-layer reporting or sampling data statistics, which not only wastes a lot of manpower and material resources, but also has slow speed, the results are not objective enough, and there is a lack of effective dynamic monitoring data. Remote sensing technology has been widely used in agricultural monitoring due to its advantages of large-scale simultaneous observation, low cost, and spatiotemporal dynamic monitoring. How to efficiently and intelligently obtain effective information from remote sensing data images has been a hot topic of research. In order to reduce costs and improve classification accuracy, the present invention is improved on the UNet network and introduces the SE module. The UNet network adopts a symmetrical encoder-decoder structure. The encoder part downsamples and extracts feature information through convolution and pooling, and performs 4 downsamplings in total; its decoder also performs 4 upsamplings through transposed convolution to restore the features to the original image resolution. Multiple downsampling and upsampling will cause edge information loss. Compared with other images, remote sensing images often have different target scales and uneven distribution. In order to better utilize the effective information in the feature map, the SE module is introduced to suppress the use of useless information, increase the weight of important information, improve the prediction performance of the module, and make the wheat extraction result more accurate.

[0053] The specific steps are as follows:

[0054] (1) Build a UNet network model. The UNet network adopts a symmetrical encoder-decoder structure and uses the UNet network as the basic framework to extract features.

[0055] (2) The encoder is set to downsample and extract feature information through 3×3 convolution, Relu activation function and 2×2 pooling. After pooling, the number of feature map channels remains unchanged and the size is reduced to half of the original size. A total of 4 downsamplings are performed with a downsampling rate of 16 times. The BN layer before the Relu activation function is used to suppress the overfitting problem.

[0056] (3) The decoder is set to perform four upsamplings through transposed convolution to restore the feature map to the original image resolution. At the same time, a skip link method is used at each horizontal stage to merge high-level semantic features with shallow features. Finally, Softmax is used as the classification layer to segment the image pixels into two categories, one is the background and the other is the extracted wheat area.

[0057] (4) A Squeeze-and-Excitation module (SE module) is added after each double-layer convolution module. The SE module is used to analyze the relationship between channels in the feature map and provide corresponding weights for each channel. It automatically obtains the importance of each feature channel through model learning, reduces the channels with weaker feature expression capabilities, and strengthens the channels with stronger feature expression capabilities.

[0058] (5) Set up the SE module.

[0059] A1) The SE module is set to include compression and excitation operations. The compression operation performs global average pooling on the feature map of size W×H×C and compresses it into a 1×1×C vector. The excitation operation uses a fully connected neural network to perform a nonlinear transformation on the result after the compression operation. Finally, the result obtained by the excitation operation is used as a weight and multiplied by the input feature.

[0060] A2) Let the output feature of the network be U, and the size of U is (H, W, C), where H, W, and C are the height, width, and number of channels respectively;

[0061] The c-th dimension feature map u c Perform compression operation, denoted as F sq , output z c , all feature maps are compressed and output as z, the shape of z is 1×1×C, and the calculation formula is as follows:

[0062]

[0063] Among them, F sq Indicates the compression operation on the feature, u c is the c-th dimension feature of the input feature map, W and H represent the width and height of the feature, i and j are the coordinates of the feature map in the spatial dimension;

[0064] A3) Perform an excitation operation on z, denoted as F ex , the result is s, and the calculation formula is as follows:

[0065] s=F ex (z,w)=σ(w2,δ(w1,z)),

[0066] Where w, w1 and w2 are the weights in the fully connected layer, σ and δ represent the Sigmoid and Relu activation functions respectively;

[0067] A4) multiplies the channel weights, denoted as F scale , that is, weighting s to the features of each channel, the output is The calculation formula is as follows:

[0068]

[0069] Among them, F scale (u c ,s c ) represents the scalar s c With feature u c The corresponding channel product between .

[0070] The third step is to train the wheat planting area extraction network: use the preprocessed remote sensing image data to train the wheat planting area extraction network.

[0071] (1) The preprocessed remote sensing image data was input into the wheat planting area extraction network. During training, the Adamoptimizer function was selected as the parameter optimizer, the cross entropy function was used as the loss function, the initial learning rate was 0.0001, the number of training iterations was 100, and the step size was 8.

[0072] (2) Feature extraction is performed through the convolution and pooling of the UNet encoder in the wheat planting area extraction network. The decoder uses the feature map generated by upsampling to fuse with the feature map on the left. In the process, the importance of each feature channel is automatically obtained through the SE module after each double-layer convolution module to strengthen the channels with strong feature expression capabilities.

[0073] (3) After two more convolution operations, a feature map is generated.

[0074] (4) Use a 1×1 convolution to perform channel adjustment and adjust the number of channels in the final feature layer to the number of categories.

[0075] The fourth step is to obtain the remote sensing image to be extracted: obtain the remote sensing image to be extracted and perform preprocessing.

[0076] In practical applications, the acquired Landsat-8OLI images, Sentinel-2 images and GF-6 images can be preprocessed using ENVI (Environment for Visualizing Images) software. The preprocessing steps include: radiation calibration, atmospheric correction, image mosaicking and cropping, and band synthesis, etc., and the Landsat-8OLI images are subjected to image fusion processing. The amount of image information acquired by a single sensor is limited, and it is often difficult to meet application needs. Through image fusion, more useful information can be obtained from different remote sensing images to supplement the shortcomings of a single sensor. Panchromatic images generally have a higher spatial resolution, and multispectral images have richer spectral information. In order to improve the spatial resolution of multispectral images, panchromatic images can be fused with multispectral images, thereby both improving the spatial resolution of multispectral images and retaining their multispectral characteristics. The present invention uses the Gram-Schmidt Pan Sharpening method to fuse the panchromatic spectrum and the multispectral bands. The spatial resolution of the Landsat-8OLI panchromatic image is 15m, and it has only one spectral band. The multispectral image has rich spectral information, but its spatial resolution is 30m. By fusing the two, we can finally get a Landsat-8OLI multispectral image with a resolution of 15m.

[0077] The fifth step is to extract the wheat planting area from the remote sensing image: input the pre-processed remote sensing image to be extracted into the trained wheat planting area extraction network to extract the wheat planting area from the remote sensing image.

[0078] The following is a comparative verification using Sentinel-2 and GF-6 data with different resolutions, and the effect of the present invention is further illustrated in combination with simulation experiments:

[0079] 1. Simulation experiment conditions:

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

[0081] 2. Simulation experiment content and result analysis:

[0082] In order to verify the method of the present invention, SegNet, Deeplabv3+, and U-Net were selected as comparison models. The three comparison models and SE-UNet all used the same training data and test data, and used mean pixel accuracy (MPA) and mean intersection over union (MIoU) as indicators of the evaluation method. The trained network model was used to predict the data in the test set. The experiments all used Landsat-8 as the test set, and three training sets with different resolutions for model training, and the results were divided into three groups to analyze the results. Both the original image and the labeled image are Landsat-8 images. The three groups are the prediction results of different networks on the Landsat-8 test set using GF-6 as the training image, as shown in the figure. Figure 2 As shown in Figure 2, the prediction results of different networks on the Landsat-8 test set using Sentinel-2 as the training image are as follows: Figure 3 As shown in Figure 2, the prediction results of different networks on the Landsat-8 test set using Landsat-8 as the training image are as follows: Figure 4As shown in the figure. Since the resolutions of Landsat-8OLI, Sentinel-2 and GF-6 are 15m, 10m and 8m respectively, the wheat area with higher resolution in the result will be larger than the area with lower resolution. From the experimental prediction results, the SegNet segmentation results obviously have some under-prediction situations, and there are many places with dotted prediction parts, resulting in unsatisfactory segmentation results. The segmentation results of Deeplabv3+ are obviously smoother than other methods. Many wheat areas are connected together in a large area, which cannot well show the angular shape of the wheat field. The edge information of the wheat area is seriously missing. Deeplabv3+ is better when used to identify larger objects such as buildings. When used to identify winter wheat, due to the small changes in the details of the pixel blocks in the winter wheat planting area, there is less information available, resulting in the role of Deeplabv3+ not being fully utilized, and it cannot be well applied to the extraction of small fields and blocks. The overall effect of U-Net is good, but there is still a problem of missing edge information. After adding the SE module, the overall result has been significantly improved.

[0083] Overall, the method proposed in the present invention is superior to other methods in the test results based on three different data training models. MPA and MIoU are the highest. The MPA and MIoU predicted by the Landsat8 image model reach 88.97% and 81.14% respectively, the MPA and MIoU predicted by the Sentinel-2 image model reach 88.38% and 76.15% respectively, and the MPA and MIoU predicted by the GF-6 image model reach 86.37% and 75.03% respectively. Due to the different resolutions, the predicted wheat areas of Sentinel-2 and GF-6 will be larger, and there will be some deviations in calculating MPA and MIoU using Landsat8 OLI tags. However, from the prediction map, there is basically no large deviation in the overall wheat area position, which verifies the authenticity of the prediction results. The evaluation indicators corresponding to different prediction models based on three data sets are shown in Table 1.

[0084] Table 1 Comparison of evaluation indicators corresponding to different models

[0085]

[0086] 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 to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for extracting wheat planting areas from remote sensing images based on SE-UNet deep learning network, characterized in that: The following steps are involved: 11) Acquisition of basic data: Landsat-8OLI remote sensing image data was obtained and preprocessed. The wheat area vector was delineated using ArcGIS software and the wheat area in the Landsat-8OLI image was marked. The original image data size was 256×256×3 and the label data size was 256×256×1. The preprocessed remote sensing image data was obtained. 12) Construction of wheat planting area extraction network: Construction of wheat planting area extraction network based on SE-UNet network; The construction of the wheat planting area extraction network includes the following steps: 121) Build a UNet network model. The UNet network adopts a symmetrical encoder-decoder structure and uses the UNet network as the basic framework to extract features; 122) Set the encoder to downsample and extract feature information through 3×3 convolution, Relu activation function and 2×2 pooling. After pooling, the number of feature map channels remains unchanged and the size is reduced to half of the original size. A total of 4 downsamplings are performed with a downsampling rate of 16 times. The BN layer before the Relu activation function is used to suppress overfitting problems. 123) Set the decoder to perform 4 upsamplings through transposed convolution to restore the feature map to the original resolution, and use skip links to merge high-level semantic features with shallow features at each level. Finally, Softmax is used as the classification layer to segment the image pixels into two categories, one is the background and the other is the extracted wheat area; 124) After each double-layer convolution module, an SE module is added. The SE module is used to analyze the relationship between channels in the feature map and provide corresponding weights for each channel. The importance of each feature channel is automatically obtained through model learning. 125) Setting SE module; 13) Training of wheat planting area extraction network: Use pre-processed remote sensing image data to train the wheat planting area extraction network; 14) Obtaining the remote sensing image to be extracted: obtaining the remote sensing image to be extracted and performing preprocessing; 15) Extraction of wheat planting areas from remote sensing images: The pre-processed remote sensing image to be extracted is input into the trained wheat planting area extraction network to extract the wheat planting area from the remote sensing image.

2. According to claim 1, a method for extracting wheat planting areas from remote sensing images based on SE-UNet deep learning network is characterized in that: The training of the wheat planting area extraction network includes the following steps: 21) Input the preprocessed remote sensing image data into the wheat planting area extraction network. During training, select the Adamoptimizer function as the parameter optimizer, use the cross entropy function as the loss function, the initial learning rate is 0.0001, the number of training iterations is 100, and the step size is 8; 22) Feature extraction is performed through convolution and pooling of the UNet encoder in the wheat planting area extraction network. The decoder uses the feature map generated by upsampling to fuse with the feature map on the left. In the process, the importance of each feature channel is automatically obtained through the SE module after each double-layer convolution module to strengthen the channel with strong feature expression ability; 23) After two more convolution operations, a feature map is generated; 24) Use a 1×1 convolution to perform channel adjustment and adjust the number of channels of the final feature layer to the number of categories.

3. According to claim 1, a method for extracting wheat planting areas from remote sensing images based on SE-UNet deep learning network is characterized in that: The setting of the SE module comprises the following steps: 31) Set the SE module to include compression and excitation operations. The compression operation performs global average pooling on the feature map of size W×H×C and compresses it into a 1×1×C vector. The excitation operation uses a fully connected neural network to perform a nonlinear transformation on the result after the compression operation. Finally, the result obtained by the excitation operation is used as a weight and multiplied by the input feature. 32) Let the output feature of the network be U, and the size of U is (H, W, C), where H, W, and C are the height, width, and number of channels respectively; The c-th dimension feature map u c Perform compression operation, denoted as F sq , output z c , all feature maps are compressed and output as z, the shape of z is 1×1×C, and the calculation formula is as follows: Among them, F sq Indicates the compression operation on the feature, u c is the c-th dimension feature of the input feature map, W and H represent the width and height of the feature, i and j are the coordinates of the feature map in the spatial dimension; 33) Perform an excitation operation on z, denoted as F ex , the result is s, and the calculation formula is as follows: s=F ex (z,w)=σ(w2,δ(w1,z)), Where w, w1 and w2 are the weights in the fully connected layer, σ and δ represent the Sigmoid and Relu activation functions respectively; 34) Multiply the channel weights and record them as F scale , that is, weighting s to the features of each channel, the output is The calculation formula is as follows: Among them, F scale (u c ,s c ) represents the scalar s c With feature u c The corresponding channel product between .

Images