A remote sensing extraction method for cultivated land plots driven by contour and texture features

Abstract

The invention discloses a remote sensing extraction method for cultivated land plots driven by the collaborative driving of contour and texture features. The invention adopts an encoder-decoder structure to construct a CFPENet model. Firstly, a deep residual shrinkage network module is introduced into the encoder to eliminate redundant information that hinders the task of extracting farmland blocks. Then, a feature search module, a contour-texture attention module, and a contour-texture guidance module are constructed to fully extract the contour and texture features of the farmland plots. Finally, a decoder of hierarchical information fusion is constructed to strengthen the interaction ability of features at various scales, reduce the loss of useful information, and improve the extraction ability of the farmland plots. The method of the invention effectively solves the problem that the extraction results are stuck and seriously missed due to the different areas and shapes of cultivated land plots. The method of the invention can not only better extract the boundaries of cultivated land plots, but also enhance the integrity of the plot extraction results, significantly improve the extraction accuracy, and have good robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing imagery farmland plot extraction, and specifically relates to a farmland plot remote sensing extraction method driven by the collaborative efforts of contour and texture features, which is suitable for high-precision extraction of farmland plots. Background Art

[0002] As the global population continues to grow, in order to meet the huge demand for food, on the one hand, the cultivated land area has expanded, and on the other hand, agricultural production has gradually become mechanized and intensified. Among them, the growth of cultivated land area often comes at the expense of forests, posing a threat to biodiversity and mitigation of global climate change. In order to cope with this threat, reasonable policies are very important, and cultivated land plot information is an important data support. In addition, cultivated land plot data is also needed to understand the level of agricultural intensification and mechanization, crop diversity and rotation, and provide effective decision-making for agricultural production and management.

[0003] Remote sensing technology is an important means of obtaining farmland information and is widely used around the world. However, traditional manual remote sensing interpretation is time-consuming, labor-intensive and costly. Methods such as region segmentation, edge detection and machine learning have accelerated the automation of farmland remote sensing interpretation, but the extraction results suffer from serious adhesion and omission, and are prone to produce fragmented patches, making it impossible to obtain accurate farmland information.

[0004] Deep learning technology represented by convolutional neural networks has been widely used in the field of remote sensing because it can automatically obtain typical features in images for image classification. However, conventional convolutional neural network segmentation methods can only extract large areas of cultivated land and cannot be subdivided into cultivated land plots. Therefore, instance segmentation models are introduced into the task of extracting cultivated land plots, but they are less effective for cultivated land plots with small areas and long and narrow shapes. At this time, by adding boundary constraints of cultivated land plots to the convolutional neural network, the model can be promoted to learn the boundaries of cultivated land plots, thereby improving the accuracy of cultivated land plot extraction results.

[0005] Although high-resolution remote sensing images have rich detailed features of objects, and can achieve certain results even in areas with complex planting patterns and fragmented agricultural landscapes, the "same object, different spectrum" and "same spectrum, different object" phenomena are serious, which will affect the boundary clarity of cultivated land plots to a certain extent, bringing challenges to the accurate extraction of cultivated land plots. Contour features are the shape and structure of the boundary of objects in the image, reflecting the external shape characteristics of the object, while texture features describe the repeated patterns of the image surface in a local area, reflecting the local changes and relationships within the object. Therefore, the contour and texture features of cultivated land plots are combined to more comprehensively describe the appearance and shape of cultivated land plots to improve the accuracy of cultivated land plot extraction results. Summary of the invention

[0006] The purpose of the present invention is to provide a remote sensing extraction method for cultivated land plots driven by the coordinated use of contour and texture features in order to address the above-mentioned problems existing in the prior art.

[0007] The above-mentioned purpose of the present invention is achieved by the following technical means:

[0008] A remote sensing extraction method for cultivated land plots driven by contour and texture features comprises the following steps:

[0009] Step 1: Obtain images of the existing study area and preprocess the images to obtain multispectral images;

[0010] Step 2, according to the multispectral image, a corresponding black-and-white binary image and a contour-texture feature map are obtained, and the multispectral image, the black-and-white binary image, and the contour-texture feature map are respectively cropped to a set size, and a plurality of multispectral sub-images and black-and-white binary images and contour-texture feature maps of the corresponding areas of each multispectral sub-image are obtained and used as input samples, a first true label, and a second true label, respectively, and each input sample and the first true label and the second true label of the corresponding area are taken as a sample, and all samples constitute a sample set, and then the sample set is expanded to obtain an expanded sample set;

[0011] Step 3: Divide the expanded sample set into a training data set and a validation data set in proportion, then reselect the study area as the test area, obtain the image of the test area, and then repeat step 2 to obtain multiple samples corresponding to the test area to form a test set;

[0012] Step 4: Build the CFPENet model;

[0013] Step 5: Construct a composite loss function L;

[0014] Step 6: Input the training data set into the CFPENet model for training, use the weight parameter of the last iteration as the final weight parameter of the CFPENet model, save the final weight parameter, obtain the trained CFPENet model, and then input each input sample in the test set into the trained CFPENet model for prediction, and obtain the prediction result of each input sample in the test set;

[0015] Step 7, obtain the image to be processed, and execute step 1 for the image to be processed to obtain the multispectral image to be processed, then crop the multispectral image to be processed into multiple multispectral sub-images to be processed of a set size, and then input the multiple spectral sub-images to be processed into the CFPENet model in sequence for prediction, and obtain the prediction results corresponding to each multispectral sub-image to be processed.

[0016] As mentioned above, the preprocessing of the image in step 1 to obtain a multispectral image specifically includes the following processes: first, the multispectral band of the image is subjected to radiometric calibration, atmospheric correction, and geometric correction; secondly, the panchromatic band of the image is subjected to radiometric calibration and orthorectification; finally, the processed panchromatic band and multispectral band are fused using an image fusion method to obtain a multispectral image of a set resolution.

[0017] As mentioned above, step 2 specifically includes the following steps:

[0018] Step 2.1, using the vector surface construction method to draw the vector data of the cultivated land based on the multispectral image, without drawing the non-cultivated land area, to obtain the vector data of the cultivated land;

[0019] Step 2.2: Use the vector-to-raster tool to convert the drawn cultivated land vector data into a black-and-white binary image, with cultivated land in white and non-cultivated land in black.

[0020] Step 2.3, using edge recognition operator to calculate the contour feature map based on the black and white binary image;

[0021] Step 2.4, using the gray level co-occurrence matrix to calculate the variance based on the multispectral image to obtain the texture feature map;

[0022] Step 2.5, adding the contour feature map and the texture feature map to obtain a contour-texture feature map;

[0023] Step 2.6, crop the multispectral image into multiple multispectral sub-images of set sizes, each of which is an input sample; crop the black-and-white binary image into multiple black-and-white binary sub-images of set sizes, each of which is a first true label; crop the contour-texture feature image into multiple contour-texture feature sub-images of set sizes, each of which is a second true label; each multispectral sub-image corresponds to a black-and-white binary sub-image and a contour-texture feature sub-image of the same area, and take each input sample and the corresponding first true label and second true label as a sample, and all samples constitute a sample set;

[0024] Step 2.7: Expand the sample set in step 2.6 to obtain an expanded sample set.

[0025] As mentioned above, the CFPENet model in step 4 includes I deep residual shrinkage modules, I contour-texture guidance modules, J feature search modules, J level information fusion modules, and a contour-texture attention module. The input sample passes through a CBR module with a convolution kernel size of 3×3 and the number of convolutions C to obtain feature F1;

[0026] The input of the i-th deep residual shrinkage module is feature F i, the output of the i-th deep residual contraction module is feature F i+1 ;

[0027] The input of the j-th feature search module is the feature F output by the j-th deep residual shrinkage module j+1 , the output of the j-th feature search module is feature FC j ;

[0028] Features FC1~FC output by each feature search module J They are input into the contour-texture attention module respectively, and the output of the contour-texture attention module is the feature FCT;

[0029] The input of the i-th contour-texture guidance module is the feature FCT output by the contour-texture attention module and the feature F output by the i-th deep residual shrinkage module. i+1 , the output of the i-th contour-texture guided module is the feature FCTG i ;

[0030] When j is 1, 2, ..., J-1, the input of the j-th level information fusion module is the feature FCTG output by the j-th contour-texture guidance module j And the feature FRH output by the j+1th level information fusion module j+1 ; When j is J, the input of the j-th level information fusion module is the output FCTG of the j-th contour-texture guidance module j and the output feature FCTG of the j+1th contour-texture guidance module j+1 ; The output of the first level information fusion module is also the output of the CFPENet model;

[0031] Among them, i and j are both serial numbers, serial number i is 1, 2, ..., I, and serial number j is 1, 2, ..., J.

[0032] As mentioned above, the i-th deep residual shrinkage module performs the input feature F i The specific processing steps include:

[0033] Step 4.1.1: Set feature F i The feature F is obtained by sequentially passing through two CBR modules with a convolution kernel size of 3×3 and a convolution number of C. i，1 ;

[0034] Step 4.1.2: Feature F i，1 After taking the absolute value and adaptive average pooling operations, we get the feature F i，2 ;

[0035] Step 4.1.3: Set feature F i，2After being calculated through full connection, batch normalization-Relu function-full connection, and Sigmoid function, the feature F is obtained. i，3 ;

[0036] Step 4.1.4: Set feature F i，2 With feature F i，3 Multiply to get the threshold ;

[0037] Step 4.1.5: Set the threshold With feature F i，1 Perform soft thresholding to obtain feature F i，4 ;

[0038] The soft thresholding operation is calculated based on the following formula: ,

[0039] Where y is the characteristic F i，4 , x is the feature F i，1 ;

[0040] Step 4.1.6: Set feature F i，4 With feature F i After adding, we get the feature F output by the i-th deep residual shrinkage module i+1 .

[0041] As mentioned above, the j-th feature search module performs the input feature F j+1 The specific processing steps include:

[0042] Step 4.2.1: Set feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 3×1, a convolution kernel size of 1×3, and a convolution kernel size of 3×3 with a dilation rate of 3. j+1，1 ;

[0043] Step 4.2.2: Feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 5×1, a convolution kernel size of 1×5, and a convolution kernel size of 3×3 with a dilation rate of 5. j+1，2 ;

[0044] Step 4.2.3: Set feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 7×1, a convolution kernel size of 1×7, and a convolution kernel size of 3×3 with a dilation rate of 7. j+1，3 ;

[0045] Step 4.2.4: Set feature F j+1 Through the CBR module with a convolution kernel size of 3×3, the feature F is obtained j+1，4 ;

[0046] Step 4.2.5: Use the Concat operation to convert feature F j+1，1 , Feature F j+1，2 , and feature F j+1，3 The convolution kernel is stacked in the channel dimension and passes through the CBR module with a convolution kernel size of 1×1, and then combined with F j+1，4 Add them together to get the feature FC output by the j-th feature search module j .

[0047] As mentioned above, the contour-texture attention module specifically includes the following processing steps for the input:

[0048] Step 4.3.1: First, for features FC2~FC J Up-sample so that the feature FC1~FC J The size remains the same, and the updated FC2~FC J ,

[0049] Step 4.3.2, concatenate feature FC1, updated feature FC2, and updated feature FC3 in the channel dimension to obtain feature FT;

[0050] Step 4.3.3: Pass the feature FT through two CBR modules with a convolution kernel size of 3×3, a 1×1 convolution, and a sigmoid function to obtain the feature FCT output by the contour-texture attention module.

[0051] As mentioned above, the i-th contour-texture guidance module is used to input the feature F i+1 The feature FCT specifically includes the following processing steps:

[0052] Step 4.4.4: The contour-texture guidance module performs 2 i-1 Upsample by times to get the feature FCT i+1 ;

[0053] Step 4.4.5: Convert feature FCT i+1 With feature F i+1 Multiply them to get the feature FCT′ i+1 ;

[0054] Step 4.4.6: Feature FCT i+1 、F i+1 and FCT′ i+1 Add them together to get the feature FH i ;

[0055] Step 4.4.7: Set feature FH i Through two CBR modules with a convolution kernel size of 3×3, the feature FHC is obtained i ;

[0056] Step 4.4.8: Feature FHC i After the average pooling operation, multiply it with itself to get the feature FHC′ i , and finally the feature FHC′ i With feature FH i Add them together to get the feature FCTG output by the i-th contour-texture guidance module i .

[0057] As mentioned above, the j-th level information fusion module specifically includes the following processing steps for the input features:

[0058] Step 4.5.1. Characteristic FCTG j+1 or characteristic FRH j+1 The feature size is restored by deconvolution operation, and the feature FCTG j Keep the size of the same, and stack them in the channel dimension to get the feature FG j ;

[0059] Step 4.5.2: Feature FG j Group by channel to get feature FG j，1 , FG j，2 , FG j，3 and FG j，4 ;

[0060] Step 4.5.3: FG j，1 , FG j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 1 j，1 ;

[0061] Step 4.5.4: Feature FGH j，1 , FG j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 2 j，2 ;

[0062] Step 4.5.5: Feature FGH j，1 , FGH j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 3 j，3 ;

[0063] Step 4.5.6: Feature FGHj，1 , FGH j，2 , FGH j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a dilation rate of 4 and a convolution kernel size of 3×3 j，4 ;

[0064] Step 4.5.7: Feature FGH j，1 , FGH j，2 , FGH j，3 and FGH j，4 Stacking is performed on the channel dimension to obtain the feature FGH j ;

[0065] Step 4.5.8: Feature FGH j After a 1×1 convolution operation, it is combined with the feature FG j Add them together and then perform a 3×3 convolution operation to obtain the fused feature FRH j .

[0066] The composite loss function L in step 5 as described above is based on the following formula: ,

[0067] Where L is the composite loss function of the CFPENet model, is the loss function of contour-texture features, is the loss function of the plot range;

[0068] Loss function of contour-texture features The definition is as follows: ,

[0069] In the formula, is the loss function of contour-texture features, is the Dice loss function, P is the prediction result of the CFPENet model, G is the second true label, To perform the intersection operation, To take the union operation;

[0070] Plot-wide loss function The definition is as follows: ,

[0071] In the formula, is the loss function of the plot range, is the weighted binary cross entropy loss function, is the weighted intersection-over-union loss function.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] (1) The method of the present invention effectively solves the serious problem of adhesion and omission of extraction results caused by the different areas and shapes of cultivated land plots. The method of the present invention can not only better extract the boundaries of cultivated land plots, but also enhance the integrity of the plot extraction results, significantly improve the extraction accuracy, and have good robustness.

[0074] (2) The present invention adopts an encoder-decoder structure and introduces a deep residual shrinkage network for feature encoding to eliminate redundant information of the farmland plot extraction task, constructs a contour-texture feature attention module and a contour-texture feature guidance module, fully mines the contour and texture features of the farmland plot, and also constructs a decoder for hierarchical information fusion to enhance the interactive ability of features at various scales and reduce the information loss caused by feature restoration, thereby improving the extraction ability of farmland plots. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of the structure of the CFPENet model of the present invention;

[0076] Figure 2 is a schematic diagram of the structure of the deep residual shrinkage module of the present invention (where C is the number of channels and W is the image width);

[0077] Figure 3 It is a structural schematic diagram of a feature search module of the present invention;

[0078] Figure 4 It is a structural diagram of the hierarchical information fusion module of the present invention;

[0079] Figure 5 is a schematic diagram of the structure of the contour-texture attention module of the present invention;

[0080] Figure 6 is a schematic structural diagram of a contour-texture guiding module of the present invention;

[0081] Figure 7 It is a display diagram of the results of the extraction of each model cultivated land plot in Example 2 of the present invention, wherein:

[0082] (a) is an actual image diagram of an example of the present invention;

[0083] (b) is the present invention Figure 7 Schematic diagram of the actual boundaries of the plot in (a);

[0084] (c) is the Unet model used in the present invention to Figure 7 The result of extracting cultivated land from the image in (a);

[0085] (d) is the R2UNet model used in the present invention Figure 7The result of extracting cultivated land from the image in (a);

[0086] (e) is the ResUNet-a model used in the present invention Figure 7 The result of extracting cultivated land from the image in (a);

[0087] (f) is the BsiNet model used in the present invention Figure 7 The result of extracting cultivated land from the image in (a);

[0088] (g) is the SEANet model used in the present invention Figure 7 The result of extracting cultivated land from the image in (a);

[0089] (h) is the CFPENet model constructed by the present invention. Figure 7 The result of extracting cultivated land from the image in (a);

[0090] Figure 8 CFPENet model attention heat map with contour-texture module and without contour-texture module in embodiment 2 of the present invention, where:

[0091] (a) is a first example of an actual image map including regular plots of land according to the present invention;

[0092] (b) is a second example of an actual image diagram including regular plots of land according to the present invention;

[0093] (c) is an example of an actual image diagram including irregular plots of land according to the present invention;

[0094] (d) is an actual image map of the present invention including both regular plots and irregular plots;

[0095] (e) is the CFPENet model with contour-texture module of the present invention. Figure 8 The attention heat map after the cultivated land plots are extracted from the actual image in (a);

[0096] (f) is the CFPENet model with contour-texture module of the present invention. Figure 8 The attention heat map after the cultivated land plots are extracted from the actual image in (b);

[0097] (g) is the CFPENet model with contour-texture module of the present invention. Figure 8 The attention heat map after the cultivated land plots are extracted from the actual image in (c);

[0098] (h) is the CFPENet model with contour-texture module of the present invention. Figure 8(d) is the attention heat map after the cultivated land plots are extracted from the actual image;

[0099] (i) is the CFPENet without contour-texture module of the present invention NT Model pair Figure 8 The attention heat map after the cultivated land plots are extracted from the actual image in (a);

[0100] (j) is the CFPENet without contour-texture module of the present invention NT Model pair Figure 8 The attention heat map after the cultivated land plots are extracted from the actual image in (b);

[0101] (k) is the CFPENet without contour-texture module of the present invention NT Model pair Figure 8 The attention heat map after the cultivated land plots are extracted from the actual image in (c);

[0102] (m) is the CFPENet without contour-texture module of the present invention NT Model pair Figure 8 (d) is the attention heat map after the cultivated land plots are extracted from the actual image;

[0103] Fig. 9 1 is a diagram showing the results of the CF PENet model with and without the contour-texture module in Example 2 of the present invention, wherein:

[0104] (a) is the present invention Figure 8 Schematic diagram of the actual scope of the plot in the actual image map (a);

[0105] (b) is obtained by using the CFPENet model with contour-texture module of the present invention Figure 8 Schematic diagram of the actual scope of the plot in the actual image map (a);

[0106] (c) is the CFPENet of the present invention using no contour-texture module NT Model Acquisition Figure 8 Schematic diagram of the actual scope of the plot in the actual image map (a);

[0107] (d) is the present invention Figure 8 Schematic diagram of the actual extent of the plot in the actual image map (b);

[0108] (e) is obtained by using the CFPENet model with contour-texture module of the present invention Figure 8 Schematic diagram of the actual extent of the plot in the actual image map (b);

[0109] (f) is the CFPENet without contour-texture module of the present invention NT Model Acquisition Figure 8 Schematic diagram of the actual extent of the plot in the actual image map (b);

[0110] (g) is the present invention Figure 8 Schematic diagram of the actual extent of the plot in the actual image map (c);

[0111] (h) is obtained by using the CFPENet model with contour-texture module of the present invention Figure 8 Schematic diagram of the actual extent of the plot in the actual image map (c);

[0112] (i) is the CFPENet without contour-texture module of the present invention NT Model Acquisition Figure 8 Schematic diagram of the actual extent of the plot in the actual image map (c);

[0113] (j) is the present invention Figure 8 (d) is a schematic diagram of the actual extent of the plot in the actual image;

[0114] (k) is obtained by using the CFPENet model with contour-texture module of the present invention Figure 8 (d) is a schematic diagram of the actual extent of the plot in the actual image;

[0115] (m) is the CFPENet without contour-texture module of the present invention NT Model Acquisition Figure 8 (d) is a schematic diagram of the actual extent of the plot in the actual image. DETAILED DESCRIPTION

[0116] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. The embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0117] Example 1

[0118] A remote sensing extraction method for cultivated land plots driven by the coordinated use of contour and texture features comprises the following steps:

[0119] Step 1: Obtain images of the existing study area and preprocess the images to obtain multispectral images;

[0120] The preprocessing process is as follows: first, the multispectral band of the image is subjected to radiometric calibration, atmospheric correction, and geometric correction; secondly, the panchromatic band of the image is subjected to radiometric calibration and orthorectification; finally, the processed panchromatic band and multispectral band are fused using the image fusion method (NNDiffuse panchromatic sharpening) to obtain a multispectral image with a resolution of 1m.

[0121] In this embodiment, the image of the study area is obtained from the land observation satellite data service platform of the China Resources Satellite Application Center. The obtained Gaofen-2 image is preprocessed using ENVI software to obtain the Gaofen-2 multispectral image;

[0122] Step 2: Perform manual visual interpretation of the GF-2 multispectral image, obtain the corresponding black-and-white binary image and contour-texture feature map based on the GF-2 multispectral image, and crop the GF-2 multispectral image, black-and-white binary image, and contour-texture feature map to 256 pixels respectively. 256 size, among which, is a multiplication sign, and a plurality of GF-2 multispectral sub-images and black-and-white binary images and contour-texture feature maps of the corresponding regions of each GF-2 multispectral sub-image are obtained and used as input samples, a first true label, and a second true label, respectively. Each input sample and the first true label and the second true label of the corresponding region are taken as a sample, and all samples constitute a sample set, and then the sample set is expanded to obtain an expanded sample set;

[0123] Step 2.1, using the vector surface construction method in ArcMap software to draw the vector data of the cultivated land based on the GF-2 multispectral image, and not drawing the non-cultivated land area, to obtain the vector data of the cultivated land;

[0124] Step 2.2: Use the vector-to-raster tool in ArcMap software to convert the drawn cultivated land plot vector data into a black-and-white binary map, with cultivated land plots in white and non-cultivated land plots in black.

[0125] Step 2.3, set the low threshold parameter and the high threshold parameter in the edge recognition operator, and use the edge recognition operator (Canny function) in the OpenCV library to calculate the contour feature map based on the black and white binary image;

[0126] Step 2.4, using the gray level co-occurrence matrix to calculate the variance based on the GF-2 multispectral image to obtain the texture feature map;

[0127] Step 2.5, adding the contour feature map and the texture feature map to obtain a contour-texture feature map;

[0128] Step 2.6: To facilitate model input, the GF-2 multispectral image is cropped into multiple 256 256-size GF-2 multispectral sub-images, each GF-2 multispectral sub-image is used as an input sample; the black and white binary image is cropped into multiple 256 256-size black-and-white binary sub-images, each of which is used as a first true label; the contour-texture feature map is cropped into multiple 256 256-size contour-texture feature submaps, each of which serves as a second true label; each GF-2 multispectral sub-image corresponds to a black-and-white binary submap and contour-texture feature submap of the same area, and each input sample and the corresponding first true label and second true label are taken as a sample, and all samples constitute a sample set.

[0129] Step 2.7: In order to prevent the model from overfitting and improve the robustness of the model, based on the sample set in step 2.6, data enhancement processing is performed by image rotation, image flipping, or random noise addition to expand the sample set, and finally the expanded sample set is obtained;

[0130] Step 3: Divide the expanded sample set into a training data set and a validation data set in proportion. At the same time, in order to ensure the correctness of the model test, reselect other study areas (i.e., other study areas different from the study areas corresponding to the training data set and the validation data set) as the test areas, obtain the GF-2 image of the test area, and then repeat step 2 to obtain multiple samples corresponding to the test area to form a test set.

[0131] Set the serial number i to {1, 2, ..., I} and the serial number j to {1, 2, ..., J (also I-1)};

[0132] In this embodiment, I is 4 and J is 3;

[0133] Step 4. Construct the CFPENet model. The CFPENet model includes I deep residual shrinkage modules (DBS-Block modules), I contour-texture guidance modules (CTG-Module modules), J feature search modules (FS-Module modules), J hierarchical information fusion modules (HIF-Module modules), and a contour-texture attention module (CTA-Module module). The input sample passes through a CBR (Convolution-BatchNorm-ReLU) module with a convolution kernel size of 3×3 (the number of convolutions is C) to obtain the feature F1;

[0134] The input of the i-th deep residual shrinkage module is feature F i , the output of the i-th deep residual contraction module is feature F i+1 ;

[0135] The input of the j-th feature search module is the feature F output by the j-th deep residual shrinkage module j+1 , the output of the j-th feature search module is feature FC j ;

[0136] Features FC1~FC output by each feature search module J They are input into the contour-texture attention module respectively, and the output of the contour-texture attention module is the feature FCT;

[0137] The input of the i-th contour-texture guidance module is the feature FCT output by the contour-texture attention module and the feature F output by the i-th deep residual shrinkage module. i+1 , the output of the i-th contour-texture guided module is the feature FCTG i ;

[0138] When j is 1, 2, ..., J-1 (in this embodiment, J is 3, then in this embodiment, when j is 1 and 2), the input of the j-th level information fusion module is the feature FCTG output by the j-th contour-texture guidance module j And the feature FRH output by the j+1th level information fusion module j+1 ; When j is J, the input of the j-th level information fusion module is the output FCTG of the j-th contour-texture guidance module j and the output feature FCTG of the j+1th contour-texture guidance module j+1 ; The output of the first-level information fusion module is also the output of the CFPENet model.

[0139] The present invention adopts an encoder-decoder structure to construct a CFPENet model. The structure of the CFPENet model is as follows: Figure 1 As shown in the figure, firstly, in the encoder, a deep residual shrinkage network module is introduced to eliminate the redundant information that hinders the farmland block extraction task, and then a feature search module, a contour-texture attention module, and a contour-texture guidance module are constructed to fully extract the contour and texture features of the farmland plots. Finally, a hierarchical information fusion decoder is constructed to enhance the interaction ability of features at each scale, reduce the loss of useful information, and improve the extraction ability of farmland plots. Specifically, the following steps are included:

[0140] Step 4.1, the present invention is based on the deep convolutional neural network structure of the encoder-decoder, and introduces a deep residual shrinkage module as a model encoder. The CFPENet model includes a deep residual shrinkage module (DBS-Block module). The internal structure of the deep residual shrinkage module is as follows: Figure 2As shown in Figure 1, the input sample (i.e., the sub-image of the GF-2 multispectral image) passes through a CBR module with a convolution kernel size of 3×3 (the number of convolutions is C) to obtain feature F1, and the input of the i-th deep residual contraction module is feature F i , the output of the i-th deep residual contraction module is feature F i+1 , the i-th deep residual contraction module performs the input feature F i The specific processing steps include:

[0141] Step 4.1.1: Set feature F i The feature F is obtained by sequentially passing through two CBR modules with a convolution kernel size of 3×3 (the number of convolutions is C) i，1 ;

[0142] Step 4.1.2: Feature F i，1 After taking the absolute value and adaptive average pooling (GAP), the feature F is obtained. i，2 ;

[0143] Step 4.1.3: Set feature F i，2 After being calculated by full connection, batch normalization-Relu function-full connection (BatchNorm-ReLU-FC, BRF), and Sigmoid function, the feature F is obtained. i，3 ;

[0144] Step 4.1.4: Set feature F i，2 With feature F i，3 Multiply to get the threshold ;

[0145] Step 4.1.5: Set the threshold With feature F i，1 Perform soft thresholding to obtain feature F i，4 ;

[0146] The calculation formula for the soft thresholding operation is as follows: (1)

[0147] Where y represents the calculation result of the soft thresholding operation (i.e., feature F i，4 ), x represents the input feature of the soft thresholding operation (i.e., feature F i，1 ).

[0148] Step 4.1.6: Set feature F i，4 With feature F i After adding, we get the feature F output by the i-th deep residual shrinkage module i+1 .

[0149] Step 4.2: In the present invention, a feature search module is added before the contour-texture module. The CFPENet model includes J feature search modules. The input of the jth feature search module is the feature F output by the jth deep residual shrinkage module. j+1 , the output of the j-th feature search module is feature FC j , the internal structure of the feature search module is as follows Figure 3 As shown, the j-th feature search module performs an operation on the input feature F j+1 The specific processing steps include:

[0150] Step 4.2.1: Set feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 3×1, a convolution kernel size of 1×3, and a convolution kernel size of 3×3 with a dilation rate of 3. j+1，1 ;

[0151] Step 4.2.2: Feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 5×1, a convolution kernel size of 1×5, and a convolution kernel size of 3×3 with a dilation rate of 5. j+1，2 ;

[0152] Step 4.2.3: Set feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 7×1, a convolution kernel size of 1×7, and a convolution kernel size of 3×3 with a dilation rate of 7. j+1，3 ;

[0153] Step 4.2.4: Set feature F j+1 Through a CBR module with a convolution kernel size of 3×3, the feature F is obtained j+1，4 ;

[0154] Step 4.2.5: Use the Concat operation to convert feature F j+1，1 , Feature F j+1，2 , and feature F j+1，3 The concatenation is performed in the channel dimension and passed through a CBR module with a convolution kernel size of 1×1 before being combined with F j+1，4 Add them together to get the feature FC output by the j-th feature search module j .

[0155] Step 4.3, the present invention adds a contour-texture module to the encoder part, and the contour-texture module includes a contour-texture attention module and a contour-texture guidance module. The contour-texture attention module is used to obtain the contour and texture features of the cultivated land in the image. The CFPENet model includes a contour-texture attention module. The internal structure of the contour-texture attention module is as follows: Figure 5As shown, the features FC1~FC output by each feature search module J The contour-texture attention module outputs the feature FCT. The contour-texture attention module specifically includes the following processing steps for the input:

[0156] Step 4.3.1: Contour-texture attention module for features FC2~FC J Up-sample so that the feature FC1~FC J The size remains the same, and the updated FC2~FC J ,

[0157] In this embodiment, three feature search modules are included, and the outputs of the three feature search modules are feature FC1, feature FC2, and feature FC3, respectively. Therefore, features FC1 to FC3 are respectively input into the contour-texture attention module, and feature FC2 and feature FC3 are respectively upsampled by 2 times and 4 times to keep the size consistent with feature FC1, thereby obtaining updated features FC2 and FC3;

[0158] Step 4.3.2, concatenate feature FC1, updated feature FC2, and updated feature FC3 in the channel dimension to obtain feature FT;

[0159] Step 4.3.3, the feature FT is calculated by two CBR modules with a convolution kernel size of 3×3, a 1×1 convolution, and a sigmoid function to obtain the feature FCT output by the contour-texture attention module;

[0160] Step 4.4: The contour-texture guidance module of the present invention is to introduce the extracted contour-texture features of the cultivated land into the model for reinforcement learning. The CFPENet model includes I contour-texture guidance modules, and the input of the i-th contour-texture guidance module is the feature F output by the i-th deep residual shrinkage module. i+1 The output of the contour-texture attention module is the feature FCT, and the output of the i-th contour-texture guidance module is the feature FCTG i , the internal structure of the contour-texture guidance module is as follows Figure 6 As shown, the i-th contour-texture guidance module is used to input feature F i+1 The feature FCT specifically includes the following processing steps:

[0161] Step 4.4.4: The contour-texture guidance module performs 2 i-1 Upsample by times to get the feature FCT i+1 ;

[0162] Step 4.4.5: Convert feature FCT i+1 With feature Fi+1 Multiply them to get the feature FCT′ i+1 ;

[0163] Step 4.4.6: Feature FCT i+1 、F i+1 and FCT′ i+1 Add them together to get the feature FH i ;

[0164] Step 4.4.7: Set feature FH i Through two CBR modules with a convolution kernel size of 3×3, the feature FHC is obtained i ;

[0165] Step 4.4.8: Feature FHC i After the average pooling operation, multiply it with itself to get the feature FHC′ i , and finally the feature FHC′ i With feature FH i Add them together to get the feature FCTG output by the i-th contour-texture guidance module i .

[0166] Step 4.5, the present invention adds a decoder module of hierarchical information fusion to the CFPENet model. The CFPENet model includes J hierarchical information fusion modules. When j is 1, 2, ..., J-1, the input of the j-th hierarchical information fusion module is the feature FCTG output by the j-th contour-texture guidance module. j And the feature FRH output by the j+1th level information fusion module j+1 ; When j is J, the input of the j-th level information fusion module is the output FCTG of the j-th contour-texture guidance module j and the output feature FCTG of the j+1th contour-texture guidance module j+1 ; The output of the first level information fusion module is also the output of the CFPENet model; the internal structure of the level information fusion module is as follows Figure 4 As shown, the j-th level information fusion module specifically includes the following processing steps for the input features:

[0167] Step 4.5.1. Characteristic FCTG j+1 or characteristic FRH j+1 (Right now Figure 4 F low ) The feature size is restored by deconvolution operation, and the feature FCTG j (Right now Figure 4 F high ) keep the size consistent and stack them in the channel dimension to get the feature FG j ;

[0168] Step 4.5.2: Feature FG j Group by channel to get feature FG j，1 , FG j，2 , FG j，3 and FG j，4 ;

[0169] Step 4.5.3: FG j，1 , FG j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained by using a CBR module with a convolution kernel size of 3×3 and a dilation rate (d) of 1. j，1 ;

[0170] Step 4.5.4: Feature FGH j，1 , FG j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 2 j，2 ;

[0171] Step 4.5.5: Feature FGH j，1 , FGH j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 3 j，3 ;

[0172] Step 4.5.6: Feature FGH j，1 , FGH j，2 , FGH j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 4 j，4 ;

[0173] Step 4.5.7: Feature FGH j，1 , FGH j，2 , FGH j，3 and FGH j，4 Stacking is performed on the channel dimension to obtain the feature FGH j ;

[0174] Step 4.5.8: Feature FGH j After a 1×1 convolution operation (Conv(1×1)), it is combined with the feature FG j Add them together and then perform a 3×3 convolution operation to obtain the fused feature FRH j .

[0175] It should be noted that the subscripts i and j in the present invention only represent serial numbers, such as F i and F j , when i=j, F i and F j For the same quantity, if i=j=1, it represents F1.

[0176] Step 5: Construct a composite loss function L. The composite loss function L is based on the following formula: (2)

[0177] Where L represents the composite loss function of the CFPENet model, Represents the loss function of contour-texture features, Loss function representing the extent of a plot.

[0178] Loss function of contour-texture features The definition is as follows: (3)

[0179] In the formula, Represents the loss function of contour-texture features, represents the Dice loss function, P represents the prediction result of the CFPENet model, G represents the second true label, represents the intersection operation. represents the union operation, is the multiplication sign;

[0180] Plot-wide loss function The definition is as follows: (4)

[0181] In the formula, represents the loss function of the plot range, represents the weighted binary cross entropy loss function, represents the weighted intersection-over-union loss function.

[0182] Step 6: Input the training data set into the CFPENet model for training, use the weight parameter of the last iteration as the final weight parameter of the CFPENet model, save the final weight parameter, obtain the trained CFPENet model, and then input each input sample in the test set into the trained CFPENet model for prediction, obtain the prediction result of each input sample in the test set, extract the cultivated land plots and evaluate the results according to the prediction results of each input sample in the test set, which specifically includes the following steps:

[0183] Step 6.1, the initial learning rate of the CFPENet model is set to 0.0001; the learning rate decay method is the exponential transformation (Poly) strategy, and the decay coefficient is 0.9; the model training batch is set to 4; the number of model iterations is set to 100;

[0184] Step 6.2: Input the training data set into the CFPENet model for training, use the weight parameter of the last iteration as the final weight parameter of the CFPENet model, save the final weight parameter, and obtain the trained CFPENet model;

[0185] Step 6.3: Input each input sample in the test set into the trained CFPENet model for prediction, and obtain the prediction results of each input sample in the test set;

[0186] Step 6.4, based on the prediction results of each input sample in the test set, the prediction results of each input sample in the test set are calculated with the first true label corresponding to each input sample. According to the results of the difference calculation, the true positive (True Positive, TP), false positive (False Positive, FP), false negative (False Negative, FN) and true negative (True Negative, TN) of the cultivated land plot extraction results are classified and counted, and the precision (Precision), recall (Recall), overall accuracy (Overall Accuracy, OA) and mean intersection of Union (Mean Intersection of Union, Miou) are calculated, and the model accuracy is evaluated based on these four indicators. In addition, from the perspective of farmland plots, the global over-classification error (GOC), the global under-classification error (GUC), and the global total classification error (GTC) are used to evaluate the accuracy of the prediction results of the CFPENet model.

[0187] Precision, recall, overall accuracy, and average intersection-over-union ratio are calculated based on the following formulas:

[0188] (5)

[0189] (6)

[0190] (7)

[0191] (8)

[0192] The global over-classification error GOC, the global under-classification error GUC, and the global total classification error GTC are based on the following formulas:

[0193] (9)

[0194] (10)

[0195] (11)

[0196] In the formula, GOC represents the global over-classification error, GUC represents the global under-classification error, and GTC represents the global total classification error. represents the farmland plots predicted by the model (m=0, 1, 2, 3, …, n), Indicates that in the sample label The farmland plot with the largest overlapping area, n represents the total number of farmland plots. and Respectively and The area of ​​the farmland plot, represents the intersection operation. For the multiplication sign.

[0197] Step 7: Get the actual image to be processed, and execute step 1 on the image to be processed to obtain the multispectral image to be processed, and then cut the multispectral image to be processed into multiple 256 The multispectral sub-images to be processed are of size 256, and then the multispectral sub-images to be processed are input into the CFPENet model in sequence for prediction, and the prediction results corresponding to the multispectral sub-images to be processed are obtained.

[0198] Example 2

[0199] This embodiment uses multiple domestic high-resolution 2 remote sensing images to conduct experiments, which can cover the entire area of ​​Dangyang City, Hubei Province. The acquisition time is in 2023. After preprocessing, four-band images with a resolution of 1m are obtained, specifically near-infrared, red, green, and blue. According to the distribution of cultivated land in various towns in Dangyang City, Hubei Province, four 5200 images were finally selected. 5200 pixel image and 4 2600 pixel images 2600-pixel imagery;

[0200] 5200 for 4 pieces 5200 pixel image and 4 2600 pixel images The 2600 pixel images respectively perform step 2 in Example 1 (wherein the low threshold parameter and the high threshold parameter of the edge recognition operator in step 2.3 are set to 30 and 255 respectively), and the obtained expanded sample set includes 4500 samples;

[0201] Execute step 3 in Example 1, divide the 4500 samples into 3600 training data sets and 900 validation data sets in a ratio of 8:2, and in order to ensure the correctness of the model test, select the test area separately, and generate 1500 pairs of images as the test set through image enhancement.

[0202] In this embodiment, the parameters for executing step 6 of embodiment 1 are set as follows:

[0203] This example uses the Pytorch framework to conduct experiments in the Ubuntu system, and the graphics cards are two Nvidia GeForce RTX3090 (with a total of 48GB of video memory). The initial learning rate of the model is set to 0.0001, the number of iterations is set to 100, the training batch size is set to 4, and the Adam optimizer is used for the experiment. The learning rate decay method is the Poly strategy, and the decay coefficient is 0.9.

[0204] Execute step 6 in Example 1, input the training data set of this embodiment into the CFPENet model for training, use the weight parameter of the last iteration as the final weight parameter of the CFPENet model, save the final weight parameter, obtain the trained CFPENet model of this embodiment, and then input each input sample in the test set into the trained CFPENet model for prediction, and obtain the prediction result of each input sample in the test set.

[0205] In addition, in order to verify that the effect of the present invention on the extraction of cultivated land plots is better than that of the basic deep learning model, five models, including UNet, R2UNet, ResUNet-a, BsiNet and SEANet, were selected to conduct comparative experiments with the extraction results of the CFPENet model of the present invention, and each model used the same data set and experimental parameters.

[0206] In order to verify the promoting effect of the contour-texture module of the present invention on the extraction of cultivated land plots, based on the CFPENet model, the influence of adding the contour-texture module on the plot extraction results was discussed. The parameters and data sets of each experiment were the same.

[0207] The experimental results of this embodiment are evaluated and analyzed:

[0208] In order to verify the extraction accuracy of the CFPENet model for cultivated land plots, the cultivated land plot extraction model proposed in the present invention is used to train the sample data, and then the training weight data of the model is used to predict the test set, and the test results are qualitatively and quantitatively compared with the sample data of cultivated land plots manually labeled in the corresponding area.

[0209] Qualitative comparison of experimental results:

[0210] Figure 7 (a)~(h) in Figure 8 (a) to (m), and Fig. 9 (a) to (m) in the figure give the experimental results of farmland plots of all comparison models and the experimental results of whether the contour-texture module is added to the CFPENet model. Figure 7 (a)~(h) in the figure help to compare the extraction results of each model. Figure 8 The color changes from blue to red in (a) to (m), indicating that the model's attention in this area changes from weak to strong, which can more intuitively show the promotion effect of the contour-texture module on the model. Fig. 9 (a)~(m) in the figure can more intuitively show the extraction effect of whether the contour-texture module is added to the CFPENet model.

[0211] from Figure 7 In (a) to (h), the farmland plot boundaries extracted by the UNet and R2UNet models have obvious adhesion phenomenon, low integrity, and small boundaries inside the boundaries, indicating that there are many holes inside the extracted farmland plots. The integrity of the farmland boundary extracted by ResUNet-a is improved compared with the previous two models, and the hole phenomenon is reduced. BsiNet further improves the integrity of the farmland plot boundary. This is because it adopts a multi-task learning strategy and strengthens the learning of farmland plot boundaries. By comparison, it can be seen that the farmland plot boundaries extracted by the CFPENet model are more consistent with the actual plot boundaries, and there are almost no holes inside the farmland, indicating that this model can better characterize the farmland plot boundaries, highlighting the superior performance of the CFPENet model in farmland plot extraction.

[0212] from Figure 8 As can be seen from (a) to (m) in the figure, before adding the contour-texture module, the CFPENet model's focus was relatively messy, with a strong focus on the interior of fields with regular shapes and large areas, while being unable to recognize fields with smaller areas. After adding the contour-texture module, the CFPENet model can reduce the impact of field area and effectively focus on all fields. More importantly, the contour-texture module strengthens the focus on field boundaries, allowing the model to focus more on the scope and boundaries of farmland plots, thereby improving the accuracy of the results.

[0213] from Fig. 9 It can be seen from (a) to (m) in the figure that after adding the contour-texture module, the CFPENet model can effectively reduce the adhesion between farmland plots, especially those with smaller areas, thereby more accurately depicting the boundaries of farmland plots.

[0214] The quantitative comparison of the experimental results is:

[0215] Table 1 is a comparison table of the accuracy of cultivated land plot extraction results of each model. It can be seen from Table 1 that the precision, recall, overall accuracy, and average intersection-over-union ratio of the CFPENet model proposed in the present invention are significantly improved compared with other models. The higher the values ​​of these four indicators, the better the model effect. In addition, the three indicators of global over-classification error, global under-classification error, and global total classification error of the CFPENet model of the present invention are significantly reduced compared with other models. The lower the values ​​of these three indicators, the better the model effect. Overall, the CFPENet model of the present invention has significant improvements in all aspects compared with other models.

[0216] Table 1

[0217]

[0218] Table 2 is a comparison table of experimental results with and without adding the contour-texture module to the CFPENet model. CFPENet (HB) means adding the contour-texture module to the CFPENet model. NT (HB) indicates that the contour-texture module is not added to the CFPENet model. As shown in Table 2, the CFPENet model with the contour-texture module is higher than the CFPENet model without the contour-texture module in terms of precision, recall, overall accuracy, and average intersection-over-union ratio, indicating that the contour-texture module can improve the accuracy of the CFPENet model. This is because both contour and texture features can assist in the task of extracting farmland plots. The contour-texture module can extract these two features of farmland at the same time, thereby helping the model to effectively extract farmland plots. In addition, the CFPENet model with the contour-texture module is lower than the CFPENet model without the contour-texture module in terms of global over-classification error, global under-classification error, and global total classification error. Among them, the global under-classification error index is reduced the most, indicating that the contour-texture module can effectively reduce the phenomenon of missed extraction of farmland plots. This is because the contour-texture module can assist the CFPENet model to more accurately identify the boundaries of plots, which can increase the extraction rate of farmland plots with irregular shapes and small areas, thereby improving the accuracy of the CFPENet model.

[0219] Table 2

[0220]

[0221] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0222] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0223] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0224] It should be noted that the embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the described embodiments or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A remote sensing extraction method for cultivated land plots driven by contour and texture features, characterized in that: The following steps are involved: Step 1: Obtain images of the existing study area and preprocess the images to obtain multispectral images; Step 2, according to the multispectral image, a corresponding black-and-white binary image and a contour-texture feature map are obtained, and the multispectral image, the black-and-white binary image, and the contour-texture feature map are respectively cropped to a set size, and a plurality of multispectral sub-images and black-and-white binary images and contour-texture feature maps of the corresponding areas of each multispectral sub-image are obtained and used as input samples, a first true label, and a second true label, respectively, and each input sample and the first true label and the second true label of the corresponding area are taken as a sample, and all samples constitute a sample set, and then the sample set is expanded to obtain an expanded sample set; Step 3: Divide the expanded sample set into a training data set and a validation data set in proportion, then reselect the study area as the test area, obtain the image of the test area, and then repeat step 2 to obtain multiple samples corresponding to the test area to form a test set; Step 4: Build the CFPENet model; Step 5: Construct a composite loss function L; Step 6: Input the training data set into the CFPENet model for training, use the weight parameter of the last iteration as the final weight parameter of the CFPENet model, save the final weight parameter, obtain the trained CFPENet model, and then input each input sample in the test set into the trained CFPENet model for prediction, and obtain the prediction result of each input sample in the test set; Step 7: Obtain the image to be processed, and execute step 1 for the image to be processed to obtain the multispectral image to be processed, then crop the multispectral image to be processed into multiple multispectral sub-images to be processed of a set size, and then input the multiple spectral sub-images to be processed into the CFPENet model in sequence for prediction, and obtain the prediction results corresponding to each multispectral sub-image to be processed. The CFPENet model in step 4 includes I deep residual shrinkage modules, I contour-texture guidance modules, J feature search modules, J level information fusion modules, and a contour-texture attention module. The input sample passes through a CBR module with a convolution kernel size of 3×3 and a convolution number of C to obtain feature F1; The input of the i-th deep residual shrinkage module is feature F i , the output of the i-th deep residual contraction module is feature F i+1 ; The input of the j-th feature search module is the feature F output by the j-th deep residual shrinkage module j+1 , the output of the j-th feature search module is feature FC j ; Features FC1~FC output by each feature search module J They are input into the contour-texture attention module respectively, and the output of the contour-texture attention module is the feature FCT; The input of the i-th contour-texture guidance module is the feature FCT output by the contour-texture attention module and the feature F output by the i-th deep residual shrinkage module. i+1 , the output of the i-th contour-texture guided module is the feature FCTG i ; When j is 1, 2, ..., J-1, the input of the j-th level information fusion module is the feature FCTG output by the j-th contour-texture guidance module j And the feature FRH output by the j+1th level information fusion module j+1 ; When j is J, the input of the j-th level information fusion module is the output FCTG of the j-th contour-texture guidance module j and the output feature FCTG of the j+1th contour-texture guidance module j+1 ; The output of the first level information fusion module is also the output of the CFPENet model; Among them, i and j are both serial numbers, serial number i is 1, 2, ..., I, and serial number j is 1, 2, ..., J.

2. According to claim 1, a remote sensing extraction method for cultivated land plots driven by contour and texture features is characterized in that: The preprocessing of the image to obtain a multispectral image in step 1 specifically includes the following processes: first, performing radiometric calibration, atmospheric correction and geometric correction on the multispectral band of the image; second, performing radiometric calibration and orthorectification on the panchromatic band of the image; finally, using an image fusion method to fuse the processed panchromatic band and multispectral band to obtain a multispectral image of a set resolution.

3. According to claim 2, a remote sensing extraction method for cultivated land plots driven by contour and texture features is characterized in that: The step 2 specifically includes the following steps: Step 2.1, using the vector surface construction method to draw the vector data of the cultivated land based on the multispectral image, without drawing the non-cultivated land area, to obtain the vector data of the cultivated land; Step 2.2: Use the vector-to-raster tool to convert the drawn cultivated land vector data into a black-and-white binary image, with cultivated land in white and non-cultivated land in black. Step 2.3, using edge recognition operator to calculate the contour feature map based on the black and white binary image; Step 2.4, using the gray level co-occurrence matrix to calculate the variance based on the multispectral image to obtain the texture feature map; Step 2.5, adding the contour feature map and the texture feature map to obtain a contour-texture feature map; Step 2.6, crop the multispectral image into multiple multispectral sub-images of set sizes, each of which is an input sample; crop the black-and-white binary image into multiple black-and-white binary sub-images of set sizes, each of which is a first true label; crop the contour-texture feature image into multiple contour-texture feature sub-images of set sizes, each of which is a second true label; each multispectral sub-image corresponds to a black-and-white binary sub-image and a contour-texture feature sub-image of the same area, and take each input sample and the corresponding first true label and second true label as a sample, and all samples constitute a sample set; Step 2.7: Expand the sample set in step 2.6 to obtain an expanded sample set.

4. According to claim 1, a remote sensing extraction method for cultivated land plots driven by the coordinated use of contour and texture features, characterized in that: The i-th deep residual shrinkage module is used to convert the input feature F i The specific processing steps include: Step 4.1.1: Set feature F i The feature F is obtained by sequentially passing through two CBR modules with a convolution kernel size of 3×3 and a convolution number of C. i，1 ; Step 4.1.2: Feature F i，1 After taking the absolute value and adaptive average pooling operations, we get the feature F i，2 ; Step 4.1.3: Set feature F i，2 After being calculated through full connection, batch normalization-Relu function-full connection, and Sigmoid function, the feature F is obtained. i，3 ; Step 4.1.4: Set feature F i，2 With feature F i，3 Multiply to get the threshold ; Step 4.1.5: Set the threshold With feature F i，1 Perform soft thresholding to obtain feature F i，4 ; The soft thresholding operation is calculated based on the following formula: , Where y is the characteristic F i，4 , x is the feature F i，1 ; Step 4.1.6: Set feature F i，4 With feature F i After adding, we get the feature F output by the i-th deep residual shrinkage module i+1 .

5. According to claim 4, a remote sensing extraction method for cultivated land plots driven by contour and texture features is characterized in that: The j-th feature search module performs an input feature F j+1 The specific processing steps include: Step 4.2.1: Set feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 3×1, a convolution kernel size of 1×3, and a convolution kernel size of 3×3 with a dilation rate of 3. j+1，1 ; Step 4.2.2: Feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 5×1, a convolution kernel size of 1×5, and a convolution kernel size of 3×3 with a dilation rate of 5. j+1，2 ; Step 4.2.3: Set feature F j+1 The feature F is obtained by sequentially passing through the CBR module with a convolution kernel size of 7×1, a convolution kernel size of 1×7, and a convolution kernel size of 3×3 with a dilation rate of 7. j+1，3 ; Step 4.2.4: Set feature F j+1 Through the CBR module with a convolution kernel size of 3×3, the feature F is obtained j+1，4 ; Step 4.2.5: Use the Concat operation to convert feature F j+1，1 , Feature F j+1，2 , and feature F j+1，3 The convolution kernel is stacked in the channel dimension and passes through the CBR module with a convolution kernel size of 1×1, and then combined with F j+1，4 Add them together to get the feature FC output by the j-th feature search module j .

6. According to claim 5, a remote sensing extraction method for cultivated land plots driven by the coordinated use of contour and texture features is characterized in that: The contour-texture attention module specifically includes the following processing steps for the input: Step 4.3.1: First, for features FC2~FC J Up-sample so that the feature FC1~FC J The size remains the same, and the updated FC2~FC J , Step 4.3.2, concatenate feature FC1, updated feature FC2, and updated feature FC3 in the channel dimension to obtain feature FT; Step 4.3.3: Pass the feature FT through two CBR modules with a convolution kernel size of 3×3, a 1×1 convolution, and a sigmoid function to obtain the feature FCT output by the contour-texture attention module.

7. The remote sensing extraction method of cultivated land plots driven by contour and texture features according to claim 6, characterized in that: The i-th contour-texture guidance module is used to input the feature F i+1 The feature FCT specifically includes the following processing steps: Step 4.4.4: The contour-texture guidance module performs 2 i-1 Upsample by times to get the feature FCT i+1 ; Step 4.4.5: Convert feature FCT i+1 With feature F i+1 Multiply them to get the feature FCT′ i+1 ; Step 4.4.6: Feature FCT i+1 、F i+1 and FCT′ i+1 Add them together to get the feature FH i ; Step 4.4.7: Set feature FH i Through two CBR modules with a convolution kernel size of 3×3, the feature FHC is obtained i ; Step 4.4.8: Feature FHC i After the average pooling operation, multiply it with itself to get the feature FHC′ i , and finally the feature FHC′ i With feature FH i Add them together to get the feature FCTG output by the i-th contour-texture guidance module i .

8. The remote sensing extraction method of cultivated land plots driven by contour and texture features according to claim 7 is characterized in that: The j-th level information fusion module specifically includes the following processing steps for the input features: Step 4.5.

1. Characteristic FCTG j+1 or characteristic FRH j+1 The feature size is restored by deconvolution operation, and the feature FCTG j Keep the size of the same, and stack them in the channel dimension to get the feature FG j ; Step 4.5.2: Feature FG j Group by channel to get feature FG j，1 , FG j，2 , FG j，3 and FG j，4 ; Step 4.5.3: FG j，1 , FG j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 1 j，1 ; Step 4.5.4: Feature FGH j，1 , FG j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 2 j，2 ; Step 4.5.5: Feature FGH j，1 , FGH j，2 , FG j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a convolution kernel size of 3×3 and a dilation rate of 3 j，3 ; Step 4.5.6: Feature FGH j，1 , FGH j，2 , FGH j，3 and FG j，4 After addition, the feature FGH is obtained through a CBR module with a dilation rate of 4 and a convolution kernel size of 3×3 j，4 ; Step 4.5.7: Feature FGH j，1 , FGH j，2 , FGH j，3 and FGH j，4 Stacking is performed on the channel dimension to obtain the feature FGH j ; Step 4.5.8: Feature FGH j After a 1×1 convolution operation, it is combined with the feature FG j Add them together and then perform a 3×3 convolution operation to obtain the fused feature FRH j .

9. The method for extracting cultivated land remote sensing by collaboratively driving contour and texture features according to claim 8, characterized in that: The composite loss function L in step 5 is based on the following formula: , Where L is the composite loss function of the CFPENet model, is the loss function of contour-texture features, is the loss function of the plot range; Loss function of contour-texture features The definition is as follows: , In the formula, is the loss function of contour-texture features, is the Dice loss function, P is the prediction result of the CFPENet model, G is the second true label, To perform the intersection operation, To take the union operation; Plot-wide loss function The definition is as follows: , In the formula, is the loss function of the plot range, is the weighted binary cross entropy loss function, is the weighted intersection-over-union loss function.

Images