Crop drought detection method based on remote sensing image

Abstract

The application discloses a crop drought detection method based on remote sensing images, constructs a drought detection neural network model to detect a large range of crop drought areas, effectively prevents crop drought, and provides reliable data support for realizing precise irrigation; a feature map extraction module uses ResNet50 as a backbone network to extract a feature map; a multi-scale attention pooling module is designed in an Enconder module, multi-scale context information is obtained in a multi-scale hollow convolution mode, and then a CBAM attention mechanism is fused to focus the attention of the network on a target drought area, thereby improving the extraction effect of the features. Finally, the Deconder part effectively fuses low-level feature maps and multi-scale high-level drought feature maps of the Enconder module, and finally realizes effective detection of the land drought area in a deconvolution mode, thereby optimizing the recognition effect of the crop drought boundary.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of crop drought detection, and particularly relates to a drought detection method for semantic segmentation of a remote sensing image of crops. BACKGROUND

[0002] At present, drought detection mainly adopts a method of adding various sensor devices to an Internet of Things system to realize monitoring of a crop growth environment. This method is difficult to use in a large range of farmland, greatly increases the cost of devices, and has certain limitations because the Internet of Things devices cannot stably operate for a long time in an outdoor environment. An invention patent with a publication number CN108765906A discloses an agricultural drought monitoring and early warning method. Signal towers and telegraph poles in an existing farmland are used to set multiple detection sub-points and divide regions. Each region includes multiple detection sub-points, and a total station for collecting detection data of the detection sub-points is arranged in each region. Each detection sub-point includes an air temperature and humidity detection unit, a soil humidity detection unit, a soil pH value detection unit, a wind speed detection unit, and a picture shooting unit. The conditions of air, soil, and crops in multiple sub-point detection regions are set, and are compared and sorted with precipitation, groundwater exploitation conditions, and climate and environmental conditions in the regions to find out the causes of drought and give an early warning to prevent drought.

[0003] Considering the cost, the traditional crop drought detection method is difficult to use in a large range of farmland, and the devices cannot stably operate for a long time in an outdoor environment. In view of the above situation, more and more attention is paid to the use of remote sensing images as a data set, the expansion of the range of drought detection, and the accuracy of drought boundary segmentation. However, for the drought situation of farmland crops, the distribution of drought crops in the farmland is often irregular, and the segmented image often contains part of the background image, which cannot accurately segment the drought boundary of crops. The network structure of the traditional method is single and fixed, and the segmentation effect of a large range and relatively complex drought farmland image is poor, and it is difficult to accurately and effectively detect crop drought. Therefore, it is urgent to propose an improved network architecture to solve the above problems and realize reliable detection of a large range of crop drought. SUMMARY

[0004] The application proposes a crop drought detection method based on remote sensing images to effectively improve the recognition effect of a crop drought boundary in view of the defects of the traditional method of detecting crop drought by adding various sensor devices to an Internet of Things system.

[0005] The application is implemented by using the following technical scheme: a crop drought detection method based on remote sensing images, including the following steps:

[0006] Step A, constructing and training a drought detection neural network model, the drought detection neural network model comprising a feature map extraction module, an Encoder module and a Decoder module, and determining a final drought detection neural network model;

[0007] Step B, detecting a crop drought area based on the trained drought detection neural network model, specifically comprising the following steps:

[0008] Step B1, extracting a low-level feature map of a crop drought area remote sensing image based on the feature map extraction module;

[0009] Step B2, using multi-scale atrous convolution to obtain multi-scale context information of the remote sensing image based on the Encoder module to extract a multi-scale high-level feature map, and then fusing a CBAM attention mechanism to focus the attention of the network on the target drought area to obtain a multi-scale high-level drought feature map, thereby improving the extraction effect of the drought feature;

[0010] Step B3, the Decoder module effectively fuses the low-level feature map and the multi-scale high-level drought feature map extracted by the Encoder module, restores the detected drought area in a deconvolution manner, and realizes detection of the drought area.

[0011] Further, the feature map extraction module uses a Resnet50 network, which includes an input module, a residual module and an output module, the input module is composed of a convolution layer and a maximum pooling layer, and uses a Relu activation function and a batch normalization layer to improve the network fitting capability; the residual module includes two Conv-Block and Identity-Block basic modules connected in series, Conv-Block changes the dimension of the network, and Identity-Block deepens the network; the output module averages the input features through an average pooling layer, and finally processes the input remote sensing image into a low-level feature map through a fully connected layer.

[0012] Further, the Encoder module includes a multi-scale attention pooling module, which uses different sampling rates and atrous convolution to process the low-level feature map, expands the receptive field and obtains a multi-scale high-level feature map through multiple scales; and in the process of processing the low-level feature map at different sampling rates, a CBAM convolution attention mechanism is introduced to further enhance the drought feature from two dimensions of channel and space, and finally the multi-scale high-level feature maps processed at multiple scales are fused to obtain a final fused multi-scale high-level drought feature map.

[0013] Further, the specific process of the Encoder module for extracting a multi-scale high-level drought feature map is as follows:

[0014] (1) The multi-scale attention pooling module adjusts the dilation coefficient of the dilated convolution layer to fill the number of holes into the convolution kernel to expand the convolution kernel. Assuming that the size of the convolution kernel is k, and the number of holes filled is dilations-1, then the size n of the center convolution kernel after filling dilations-1 holes is:

[0015]

[0016] (2) Assuming that the input convolution size is i, and the step is S, then the size of the convolution feature map is:

[0017]

[0018] Further, the size of the dilated convolution feature map, i.e., the multi-scale extracted crop drought feature map, is obtained:

[0019]

[0020] The convolution kernel of each scale introduces the CBAM module to enhance the extraction of drought features in the processing of low-level features. CBAM is composed of CAM spatial attention mechanism and BAM channel attention mechanism, which enhances drought features from the spatial and channel dimensions respectively;

[0021] (3) The channel attention mechanism CAM uses average pooling and maximum pooling operations on the input , respectively, and adds the corresponding elements through a shared fully connected layer to generate a one-dimensional channel attention . The channel attention is multiplied by the input element by element to obtain the drought feature map after channel dimension processing , which is the channel dimension drought feature attention , which is expressed as:

[0022]

[0023] where, is the Sigmoid function, and are the hidden layer weight and output layer weight respectively;

[0024] (4) The spatial attention mechanism SAM performs average pooling and maximum pooling on the output channel attention drought feature map , , and then performs dimension splicing to obtain the spatial dimension drought feature map through the Sigmoid function, which is expressed as:

[0025]

[0026] wherein, is a Sigmoid function, represents a convolution kernel with a kernel size of 7;

[0027] The multi-scale attention pooling module performs a maximum pooling operation after each layer scale completes the above processing, so that the sizes of the output feature maps are the same and all multi-scale high-level drought feature maps are fused in the channel dimension, and after a 1*1 convolution operation, the final output result of the multi-scale attention pooling module is obtained.

[0028] Further, the Decoder module performs a bilinear interpolation four times upsampling operation on the multi-scale high-level drought feature maps extracted by the Encoder module, and gradually recovers to the original image size; the low-level feature map is subjected to a convolution operation to compress the low-level feature channel to save the low-level feature information, and then a CBAM module is added to enhance the features of the crop drought area of the low-level feature map; finally, the multi-scale high-level drought feature map and the low-level feature map are fused and subjected to a four times magnification bilinear interpolation upsampling to decode the final prediction map, and the detection of the crop drought area is completed.

[0029] Further, in step A, when the network model is trained to determine the final drought detection neural network model, after each training batch is trained, a trained network prediction model is generated, the detected drought area image and the drought label image are subjected to IOU calculation, and the optimal solution is obtained by comparison, and the model with the highest accuracy is saved as the final model.

[0030] Compared with the prior art, the advantages and positive effects of the present application are that:

[0031] The present scheme is based on remote sensing images, and proposes a Transformer neural network method for crop drought detection, takes ResNet50 as the backbone network of the feature extraction module, designs a multi-scale attention pooling module, adopts a multi-scale dilated convolution method to reduce the loss of local information caused by the downsampling process, and introduces an attention mechanism to enhance the extraction of crop drought features; during the feature map recovery process, the high-level and low-level feature maps are effectively fused, so that the deconvolution effect is better, and the recognition effect of the crop drought boundary is further improved, which effectively prevents crop drought and provides reliable data support for realizing precision irrigation. The present scheme realizes drought detection of a large range of crops through remote sensing technology, and the pixel precision reaches 91.1%, which has wide practical application and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a schematic diagram of the Transformer network structure for crop drought detection proposed in the embodiments of the present application;

[0033] Figure 2 The feature extraction process schematic diagram is described in the embodiment of the application.

[0034] Figure 3 The Encoder module architecture schematic diagram of the Transformer network structure is described in the embodiment of the application.

[0035] Figure 4 The Decoder module architecture schematic diagram of the Transformer network structure is described in the embodiment of the application.

[0036] Figure 5 The overall architecture schematic diagram of the Transformer network structure is described in the embodiment of the application.

[0037] Figure 6 The crop drought detection result schematic diagram is described in the embodiment of the application. DETAILED DESCRIPTION

[0038] In order to enable the above-mentioned purpose, features and advantages of the present application to be more clearly understood, the present application will be further described below in conjunction with the accompanying drawings and embodiments. In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, and therefore, the present application is not limited to the specific embodiments disclosed below.

[0039] The embodiment proposes a crop drought detection method based on remote sensing images, adopts satellite remote sensing technology to take long-term photos of farmland in the monitoring area, realizes large-scale drought detection of remote sensing images based on a neural network of Transformer by analyzing remote sensing images, and the crop drought detection method based on remote sensing images comprises the following steps.

[0040] Step A, a drought detection neural network model is constructed and trained, and the drought detection neural network model comprises a feature map extraction module, an Encoder module and a Decoder module;

[0041] Step B, the drought detection neural network model is trained, and the drought detection neural network model comprises a feature map extraction module, an Encoder module and a Decoder module;

[0042] Step B1, low-level feature maps of remote sensing images of the crop drought area are extracted based on the feature map extraction module;

[0043] Step B2, multi-scale context information of the remote sensing images is obtained based on the Encoder module by using multi-scale hollow convolution, multi-scale high-level feature maps are extracted, the CBAM attention mechanism is fused to focus the attention of the network on the target drought area, multi-scale high-level drought feature maps are obtained, and the extraction effect of the drought features is improved.

[0044] Step B3, the Decoder module effectively fuses the low-level feature map and the multi-scale high-level drought feature map extracted by the Encoder module, restores the detected drought area in a deconvolution manner, and realizes detection of the drought area.

[0045] In order to more clearly understand the scheme of the present application, the model construction and specific application of the present scheme will be described in detail as follows:

[0046] In step A, the main network of the drought detection neural network model adopts a Transformer network structure, and the best training model is selected by using an intersection over union calculation method after segmenting a remote sensing image of a farmland crop drought area, as shown in Figure 1 The specific steps include the following steps:

[0047] Step A1, constructing a data set: obtaining a crop drought remote sensing image data set, and randomly dividing part of the data set into a training set, a test set and a validation set at a ratio of 6:2:2;

[0048] Step A2, constructing a drought detection neural network model, the drought detection neural network model comprising a feature extraction module, an Encoder module and a Decoder module;

[0049] As shown in Figures 2-5 The structure of the feature extraction module is shown in Figure 2 Resnet50 network, including six parts, stage0 is an input module, stage1 to stage4 are residual modules, and stage5 is an output module. The input module is composed of a convolution layer and a maximum pooling layer, and uses a Relu activation function and a batch normalization layer to improve the network fitting capability; the residual module has two basic modules Conv-Block (changing the dimension of the network) and Identity-Block (deepening the network), and the two basic modules are connected in series to increase the depth of the network; the output module passes the features through an average pooling layer, and finally processes the input remote sensing image into a low-level feature map through a fully connected layer;

[0050] The structure of the Encoder module is shown in Figure 3 The Encoder module designs a multi-scale attention pooling module, which uses a multi-scale dilated convolution method to obtain multi-scale context information, and then fuses a CBAM attention mechanism to focus the network on the target drought area, outputs a multi-scale high-level feature map, and improves the feature extraction effect;

[0051] The structure of the Decoder module is shown in Figure 4As shown, the low-level feature map and the multi-scale high-level drought feature map are fused, and the image is restored in the form of deconvolution.

[0052] Step A3, network model training:

[0053] Step A31, input all images in the training set into the Transformer network architecture for model training, and initialize the network parameters;

[0054] Step A32: input the images in the training set into Resnet50 to extract low-level feature maps;

[0055] Step A33, Encoder module extracts multi-scale features:

[0056] The feature map extracted in step A32 is input into the multi-scale attention pooling module, which uses 3x3 convolution kernel, sampling rate of 6x6, 12x12, 24x24, etc. Different expansion rate of the hole convolution kernel is used to process the low-level feature map. This method uses different sampling rates and hole convolution to extract drought feature map, expands the receptive field and obtains high-level drought feature map through multiple scales. In the process of processing low-level feature map at different sampling rates, CBAM convolution attention mechanism is introduced to further enhance drought features from two dimensions of channel and space. Finally, the high-level feature maps processed by multiple scales are fused to obtain the final fused multi-scale high-level drought feature map, which includes:

[0057] (1) The multi-scale attention pooling module adjusts the expansion coefficient (dilations) of the expansion convolution layer to put the hole number into the convolution kernel to expand the convolution kernel. The size of the convolution kernel is k, and the number of holes put in is (dilations-1). After (dilations-1) holes are put in, the size n of the center convolution kernel is:

[0058]

[0059] (2) Assuming that the input convolution size is i and the step is S, the size of the convolution feature map is:

[0060]

[0061] Then the size of the hole convolution feature map, i.e. the multi-scale extracted crop drought feature map, can be obtained:

[0062]

[0063] The convolution kernel of each scale introduces the CBAM module to enhance the extraction of drought features in processing low-level features. CBAM is composed of CAM spatial attention mechanism and BAM channel attention mechanism, which enhances drought features from spatial and channel dimensions respectively.

[0064] (3) The channel attention mechanism CAM uses average pooling and maximum pooling operations respectively to add the corresponding elements to generate a one-dimensional channel attention through a shared fully connected layer The channel attention and the input are multiplied element by element to obtain the drought feature map after channel dimension processing adjusted by the channel attention The channel dimension drought feature attention is expressed as:

[0065]

[0066] where, is a Sigmoid function, and are hidden layer weights and output layer weights respectively;

[0067] (4) The spatial attention mechanism SAM performs average pooling and maximum pooling on the output channel attention drought feature map , then performs dimension splicing and obtains the spatial dimension drought feature map through the Sigmoid function is expressed as:

[0068]

[0069] where, is a Sigmoid function, represents a convolution kernel with a kernel size of 7;

[0070] The multi-scale attention pooling module will perform a maximum pooling operation after the above processing at each scale, so that the size of the output feature map is the same and all high-level drought feature maps are fused in the channel dimension. After 1x1 convolution operation, it is used as the final output result of the multi-scale attention pooling module.

[0071] Step A34, feature map related fusion;

[0072] ​​The multi-scale high-level drought feature map extracted by the Encoder module is subjected to a four-fold upsampling operation by bilinear interpolation to gradually restore the image to the original image. The Decoder first uses a 1x1 convolution to compress the low-level feature channels to 48 to save high-level feature information, connects two tensors in columns through the torch.cat function to restore the target boundary, and finally decodes the final drought area image through interpolate with a four-fold bilinear interpolation.

[0073] Step A4, determining the final drought detection neural network model:

[0074] After the end of each training batch, a trained network prediction model is generated, and the detected drought area image and the drought label image are subjected to IOU (intersection over union) calculation. The optimal solution is obtained by comparison, and the model with the highest accuracy is saved as the final model.

[0075] The calculation formula for MIOU (intersection over union) calculation of the detected drought area image and the label image is:

[0076]

[0077] wherein, represents the total number of pixels that are i-class pixels and the prediction result is also i-class pixels; represents the total number of pixels that are i-class pixels and the prediction result is also j-class pixels.

[0078] Based on the trained drought detection neural network model, when detecting the drought area of crops in specific applications, the following steps are included:

[0079] Step B1, during feature extraction: the input remote sensing image is subjected to down-sampling by the Resnet50 network to generate a low-level feature map.

[0080] Step B2, input the low-level feature map to the Encoder module, and process it through a 1x1 convolution kernel and a dilated convolution with different expansion rates of 6x6, 12x12, and 24x24, respectively, in the multi-scale attention pooling module. In the process of processing the low-level feature map at different sampling rates, CBAM convolution attention is added for drought feature enhancement, making the boundary profile of the crop drought area more obvious. After the pooling operation, the multi-scale features are fused to obtain a multi-scale high-level drought feature map.

[0081] Step B3, the Decoder module performs a bilinear interpolation four times upsampling operation on the multi-scale high-level drought feature map extracted by the Encoder module, gradually restores to the original image size, and uses a 1*1 convolution to compress the low-level feature channel to 48 to save the low-level feature information for the original low-level feature map (the crop drought feature map obtained after rough division), and then adds a CBAM module to enhance the low-level feature and enhance the features of the crop drought area of the low-level feature map. Finally, the high and low level feature maps are fused and subjected to 4 times magnification bilinear interpolation upsampling to decode the final prediction map, completing the detection of the crop drought area.

[0082] The present embodiment uses the International Conference on Computer Vision and Pattern Recognition (CVPR) 2020 agricultural vision challenge dataset to test the method of the present application, and the results are as follows:

[0083] 21061 aerial farmland images are used in the dataset, each image consists of four color channels of 512*512, each image has a boundary map and a mask, the boundary map represents the area of the farmland, and the mask represents the effective pixels in the image. In this embodiment, part of the dataset (3000) is selected for division, of which 60% (1800) is used to train the image segmentation network, and 600 images are used to verify the trained network model.

[0084] The residual neural network ResNet50 in the original network architecture is fused with Senet attention mechanism to form a Transformer(Se-Resnet) network, and 600 images are used to train the Transformer(Se-Resnet);

[0085] The Encoder part in the original network structure is fused with the convolution block attention mechanism (CBAM), and 600 images are used to train the Transformer(increase CBAM);

[0086] The Encoder part in the original network structure is fused with multi-scale dilated convolution, and 600 images are used to train the Transformer;

[0087] The multi-scale attention pooling module of the Encoder part in the original network structure is fused with the convolution block attention mechanism (CBAM), and 600 images are used to train the method, which is the method of the present application;

[0088] The network architecture is replaced with Unet, and 600 image samples are used to train the Unet;

[0089] The network architecture is replaced with FCN, and 600 image samples are used to train the FCN;

[0090] The network architecture was replaced with SegNet, and SegNet was trained using 600 image samples;

[0091] The network architecture was replaced with PSPNet, and PSPNet was trained using 600 image samples;

[0092] The network architecture was replaced with HRNet, and HRNet was trained using 600 image samples;

[0093] With all other parameter settings kept constant, all networks were trained for 100 epochs to obtain pixel accuracy. The training results are shown in Table 1.

[0094] Table 1. Comparison of training methods for different networks

[0095]

[0096] As shown in Table 1, on the original image test set, the accuracy of the method of this invention is higher than the other four methods, with a recognition accuracy as high as 91.05%. The recognition accuracy of the method of this invention is 2 percentage points higher than the original Deeplabv3+ network, proving that fusing the Convolutional Block Attention (CBAM) mechanism can enable the feature extraction network to extract features more effectively and improve the model's recognition performance. Compared with Transformer (with added CBAM), this method fuses the Convolutional Block Attention (CBAM) mechanism into the multi-scale attention pooling module, improving the recognition accuracy by 1 percentage point, proving that fusing the CBAM mechanism into the hollow spatial pyramid can enhance the feature extraction performance. The recognition accuracy of the method of this invention is higher than the other four models, proving that this method can enhance the feature extraction performance and enable the feature extraction network to extract features more efficiently.

[0097] like Figure 6 The image shown is a remote sensing satellite image of a drought-stricken farmland. After segmentation, the drought-stricken area of ​​the crop is obtained. It can be seen that the drought area is completely preserved. When it is input into the recognition network, the performance loss caused by image noise can be reduced, thus showing good recognition performance.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting crop drought based on remote sensing images, characterized in that: The method comprises the following steps: Step A, constructing and training a drought detection neural network model, the drought detection neural network model comprising a feature map extraction module, an Encoder module and a Decoder module, and determining a final drought detection neural network model; Step B, detecting a crop drought area based on the trained drought detection neural network model, specifically comprising the following steps: Step B1, extracting a low-level feature map of a crop drought area remote sensing image based on the feature map extraction module; Step B2, extracting a multi-scale high-level feature map by using a multi-scale atrous convolution based on the Encoder module to obtain multi-scale context information of the remote sensing image, and then fusing a CBAM attention mechanism to focus the attention of the network on a target drought area to obtain a multi-scale high-level drought feature map; Step B3, effectively fusing the low-level feature map and the multi-scale high-level drought feature map extracted by the Encoder module through the Decoder module, restoring the detected drought area in a deconvolution manner, and realizing detection of the drought area.

2. The remote-sensing image-based crop drought detection method according to claim 1, wherein, The feature map extraction module uses a Resnet50 network, comprising an input module, a residual module and an output module, the input module being composed of a convolution layer and a maximum pooling layer, and using a Relu activation function and a batch normalization layer to improve the network fitting capability; the residual module comprises two Conv-Block and Identity-Block basic modules connected in series, the Conv-Block changes the dimension of the network, and the Identity-Block deepens the network; the output module averages the input features through an average pooling layer, and finally processes the input remote sensing image into a low-level feature map through a fully connected layer. 3.The remote-sensing image based crop drought detection method according to claim 1, characterized in that, The Encoder module comprises a multi-scale attention pooling module, which processes the low-level feature map by using different sampling rates and atrous convolutions, expands the receptive field and obtains a multi-scale high-level feature map through multiple scales; and introduces a CBAM convolution attention mechanism to further enhance the drought features from two dimensions of channel and space, finally fuses the high-level feature maps processed in multiple scales to obtain a final fused multi-scale high-level drought feature map.

4. The remote sensing image-based crop drought detection method according to claim 3, characterized in that: The specific process of the Encoder module for extracting a multi-scale high-level drought feature map is as follows: (1) The multi-scale attention pooling module inserts dilations-1 into the convolution kernel to expand the convolution kernel by adjusting the expansion coefficient of the dilated convolution layer, assuming that the size of the convolution kernel is k and dilations-1 is inserted, then the size n of the center convolution kernel after inserting dilations-1 is: (2) Assuming that the input convolution size is i and the step is S, the size of the convolution feature map is: and then the size of the atrous convolution feature map, i.e. the size of the multi-scale extracted crop drought feature map, is obtained: The CBAM module is introduced in the convolution kernel of each scale to enhance the extraction of drought features in the low-level features, and the CBAM is composed of a CAM spatial attention mechanism and a BAM channel attention mechanism, which enhances the drought features from two dimensions of space and channel respectively; (3) The channel attention mechanism CAM is used for input , , respectively, using average pooling and max pooling operation, through the shared full connection layer to add corresponding elements to generate one-dimensional channel attention , the channel attention and the input are multiplied element by element, and the channel attention adjusted drought feature map after channel dimension processing is obtained , the channel dimension drought feature attention is expressed as: wherein, is a Sigmoid function, and are hidden layer weights and output layer weights, respectively. (4) The spatial attention mechanism SAM averages and maximizes the output channel attention drought feature map , , and obtains a spatial dimension drought feature map after dimension splicing through a Sigmoid function is expressed as: wherein, is a Sigmoid function, denotes a convolution kernel with a kernel size of 7; After the above processing is completed at each layer scale, the multi-scale attention pooling module performs a max pooling operation, so that the size of the output feature map is the same and all multi-scale high-level drought feature maps are fused in the channel dimension. After a 1*1 convolution operation, the final output result of the multi-scale attention pooling module is obtained. 5.The remote-sensing image based crop drought detection method according to claim 1, characterized in that: The Decoder module performs a bilinear interpolation four times upsampling operation on the multi-scale high-level drought feature map extracted by the Encoder module, gradually restores to the original image size, and performs a convolution operation on the low-level feature map to compress the low-level feature channel to save low-level feature information. Then, the CBAM module is added to enhance the features of the crop drought area in the low-level feature map. Finally, the multi-scale high-level drought feature map and the low-level feature map are fused and subjected to a four times magnification bilinear interpolation upsampling to decode the final prediction map, thereby completing the detection of the crop drought area. 6.The remote-sensing image based crop drought detection method according to claim 1, characterized in that: In step A, when the network model is trained to determine the final drought detection neural network model, after each training batch is completed, a trained network prediction model is generated. The detected drought area image and the drought label image are subjected to IOU calculation, the optimal solution is obtained by comparison, and the model with the highest accuracy is saved as the final model.

Images