Covered farmland remote sensing recognition method based on comparative learning

Abstract

The invention discloses a comparative learning-based remote sensing recognition method for a film-covered farmland, which is used for realizing remote sensing recognition of the film-covered farmland. The comparative learning model structure comprises four modules which are respectively a common feature extraction module of a multi-source remote sensing image, a personality feature extraction module of a multispectral remote sensing image, a personality feature extraction module of a synthetic aperture radar remote sensing image and a feature fusion recognition module; according to the method, common feature information and personalized feature information can be effectively extracted from multi-source remote sensing images including a multispectral remote sensing image and a synthetic aperture radar remote sensing image, and a film mulching farmland remote sensing recognition result is finally obtained through information fusion. According to the method, the characteristics of the multi-source remote sensing image data are fully utilized, a comparative learning model structure design is developed, intelligent film mulching farmland remote sensing recognition is realized, the automation and accuracy of farmland film mulching monitoring are improved, and technical support is provided for green agriculture development in China.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and remote sensing image interpretation, and in particular to a remote sensing identification method for film-covered farmland based on contrast learning. Background Art

[0002] Plastic-filmed farmland is widely distributed in the arid areas of northwest my country. Due to the dry climate and rapid evaporation of surface water, agricultural land in arid areas generally adopts the method of laying plastic mulch for planting. In recent years, with the continuous deepening of my country's green agricultural concept, the current demand for the rational use and scientific management of plastic mulch has become increasingly prominent. Therefore, the mapping of plastic-filmed farmland plays an important data support role in improving the full life cycle monitoring of the use and recycling of plastic mulch.

[0003] With the rapid development of space exploration technology, earth observation satellites equipped with different types of sensors have emerged. Among them, they mainly include passive earth observation remote sensing satellites represented by multispectral remote sensing satellites and active earth observation remote sensing satellites represented by synthetic aperture radar satellites. Both types of satellites have their own advantages and disadvantages. Multispectral remote sensing image data has the advantages of high spatial resolution and wide spectral band coverage, but it is easily affected by clouds and fog, and will lose surface spectral information in areas blocked by clouds and fog. Synthetic aperture radar remote sensing data has the advantage of strong cloud and fog penetration and is not easily affected by clouds and fog, but the spatial resolution is relatively low. From the perspective of multi-source remote sensing image interpretation, the joint analysis of the two multi-source remote sensing image data will effectively improve the efficiency of remote sensing identification of film-covered farmland.

[0004] At the same time, the existing remote sensing identification of film-covered farmland mainly relies on manual identification. That is, it requires manual labeling by professionals, which not only has high manpower and material costs, but also has the possibility of misjudgment due to subjective factors. Therefore, it is urgent to develop a more efficient and accurate remote sensing identification method for film-covered farmland.

[0005] Currently, artificial intelligence technologies represented by deep learning have received widespread attention and have been applied in various fields such as agriculture, industry, and commerce. However, the existing deep learning model architecture for multi-source remote sensing image interpretation is still in the early stages of exploration. The standard multi-source remote sensing image deep learning framework is mainly based on a series of neural network structures, and there is room for improvement in the analysis of data in multiple scales, multiple perspectives, and other dimensions.

[0006] In view of the problems of low manual recognition efficiency in existing remote sensing recognition methods for film-covered farmland and the lack of multi-scale and multi-perspective analysis strategies for the interpretation of multi-source remote sensing images in standard deep learning frameworks, this paper proposes a remote sensing recognition method for film-covered farmland based on contrastive learning. By designing and constructing a contrastive learning model structure, intelligent remote sensing recognition of film-covered farmland is achieved, thereby improving the automation and accuracy of film-covered farmland monitoring. Summary of the invention

[0007] The technical problem to be solved by the present invention is the low efficiency of manual recognition of existing remote sensing recognition of film-covered farmland, and the lack of multi-scale and multi-perspective analysis strategies for the interpretation of multi-source remote sensing images in the standard deep learning framework. Therefore, a remote sensing recognition method for film-covered farmland based on contrastive learning is proposed to realize intelligent remote sensing recognition of film-covered farmland and improve the automation and accuracy of farmland film covering monitoring.

[0008] The technical solution adopted by the present invention to solve the above technical problems is: a method for remote sensing identification of farmland mulch film based on contrast learning, which realizes the process of identifying mulch-film farmland from multi-source remote sensing images. The technical solution process of the present invention is as follows Figure 1 The method of the present invention comprises the following steps:

[0009] 1) Multi-source remote sensing image reading:

[0010] First, multi-source remote sensing image samples including multispectral remote sensing images and synthetic aperture radar remote sensing images are read in, and the true values ​​of the film-covered farmland in the multi-source remote sensing image sample set are annotated to provide samples and true values ​​for the next step of model training.

[0011] 2) Contrastive learning model training:

[0012] Figure 2 This is a schematic diagram of the proposed contrastive learning model structure, which includes four modules: Figure 3 The common feature extraction module of multi-source remote sensing images shown in Figure 4 The individual feature extraction module of the multi-spectral remote sensing image shown, Figure 5 The individual feature extraction module of the synthetic aperture radar remote sensing image shown, and Figure 6 The feature fusion recognition module shown is composed of

[0013] The common feature extraction module of multi-source remote sensing images consists of convolutional layers, pooling layers, etc. Multispectral remote sensing images and synthetic aperture radar remote sensing images have undergone multi-level pooling operations to obtain high-dimensional features. During the training process, the calculated Euclidean space distance is used as the auxiliary loss function of the module for training. The specific training rules are: when the multispectral remote sensing image samples and synthetic aperture radar remote sensing image samples input at the same time are samples of the same geographical coordinate area, the Euclidean space distance obtained by step-by-step pooling is required to be the smallest, that is, the auxiliary loss function at this time aims to reduce the information feature space distance between multi-source remote sensing images with the same content information; when the multispectral remote sensing image samples and synthetic aperture radar remote sensing image samples input at the same time are samples of different geographical coordinate areas, the Euclidean space distance obtained by step-by-step pooling is required to be the largest, that is, the auxiliary loss function at this time aims to expand the information feature space distance between multi-source remote sensing images with different content information. It is necessary to use multi-source remote sensing image samples to train this module separately. After completing the training of this module, during the overall training of the comparative learning model structure, the module can output common features A and common features B, a total of two feature matrices.

[0014] The individual feature extraction module of multispectral remote sensing images and the individual feature extraction module of synthetic aperture radar remote sensing images adopt a similar structural design, both of which are composed of convolutional layers, pooling layers, etc. After training, when the multi-source remote sensing image test sample to be analyzed is input, individual feature output 1, individual feature output 2, individual feature output 3, individual feature output 4, individual feature output 5, and individual feature output 6 can be output respectively.

[0015] The feature fusion recognition module consists of convolution layer, upsampling layer, merging layer, output layer, etc. The final output result is obtained by fusing the common feature output and individual feature output of other modules.

[0016] based on Figure 2 The contrastive learning model structure shown in the figure uses multi-source remote sensing image samples including multispectral remote sensing images and synthetic aperture radar remote sensing images as training samples to train the model. The loss function of the training is binary cross entropy, the optimizer is Adam, the number of iterations is 200, the number of batches is 40, and the learning rate is 0.01. Through the above process, a trained contrastive learning model is obtained.

[0017] 3) Remote sensing identification of film-covered farmland based on contrastive learning model:

[0018] The multi-source remote sensing images to be analyzed are input into the trained comparative learning model to obtain the remote sensing identification results of film-covered farmland.

[0019] Compared with the traditional manual identification of film-covered farmland and the multi-source remote sensing image identification method based on the standard deep learning framework, the advantages of the present invention are: the manual remote sensing identification method of film-covered farmland has the problems of high manpower and material costs and many subjective misjudgments; the multi-source remote sensing image identification method based on the standard deep learning framework lacks the multi-scale, multi-perspective and other dimensional analysis of multi-source data, so the performance is unstable. In contrast, the film-covered farmland remote sensing identification method based on contrastive learning proposed in the present invention can effectively extract common feature information and individual feature information from multi-source remote sensing images including multi-spectral remote sensing images and synthetic aperture radar remote sensing images, and finally obtain the film-covered farmland remote sensing identification results through information fusion, realize intelligent film-covered farmland remote sensing identification, and improve the automation and accuracy of farmland film monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention;

[0021] Figure 2 It is a schematic diagram of the proposed contrastive learning model structure;

[0022] Figure 3 It is the common feature extraction module of the proposed multi-source remote sensing images;

[0023] Figure 4 It is the proposed individual feature extraction module of multispectral remote sensing images;

[0024] Figure 5 It is the proposed individual feature extraction module of synthetic aperture radar remote sensing image;

[0025] Figure 6 It is the proposed feature fusion recognition module. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below with reference to the accompanying drawings.

[0027] The present invention is aimed at remote sensing identification of film-covered farmland in multi-source remote sensing images, and designs a remote sensing identification method of film-covered farmland based on contrastive learning to realize intelligent remote sensing identification of film-covered farmland. In particular, in view of the problem that the existing multi-source remote sensing image target recognition method does not fully consider multi-perspective analysis, the Euclidean space distance measurement for image content is introduced into the design and training process of the deep learning model. By designing a common feature extraction module for multi-source remote sensing images, the contrastive learning model is strengthened in its feature matrix extraction capability for remote sensing images collected by two different sensor categories, multi-spectral and synthetic aperture radar. At the same time, by effectively fusing the common features and individual features of the extracted multi-source remote sensing images based on the feature fusion recognition module, the highly robust remote sensing identification of film-covered farmland in multi-source remote sensing images is innovatively realized.

[0028] In actual use, the computer will call the program based on the method flow of the present invention to complete the specific remote sensing identification work of film-covered farmland based on comparative learning. Figure 1 The flowchart of the method of the present invention is as follows:

[0029] Step 1: Read in multi-source remote sensing images:

[0030] First, multi-source remote sensing image samples including multispectral remote sensing images and synthetic aperture radar remote sensing images are read in, and the true values ​​of the film-covered farmland in the multi-source remote sensing image sample set are annotated to provide samples and true values ​​for the next step of model training.

[0031] Step 2: Contrastive learning model training:

[0032] Figure 2 This is a schematic diagram of the proposed contrastive learning model structure, which includes four modules: Figure 3 The common feature extraction module of multi-source remote sensing images shown in Figure 4 The individual feature extraction module of the multi-spectral remote sensing image shown, Figure 5 The individual feature extraction module of the synthetic aperture radar remote sensing image shown, and Figure 6 The feature fusion recognition module shown is composed of

[0033] The common feature extraction module of multi-source remote sensing images is composed of convolutional layers, pooling layers, etc. Figure 3As shown. In particular, since the film width of the film-covered farmland is generally 30 cm to 100 cm, which is smaller than the spatial resolution of common multispectral remote sensing images, there is a mixed pixel problem of "film-soil-vegetation" in the remote sensing image. In response to the mixed pixel problem, the contrastive learning model designed by the present invention specifically constructs a progressive feature extraction framework based on "small convolution kernel convolution-large convolution kernel convolution-pooling". Specifically, in the individual feature extraction module, the image input is first input into a small convolution kernel convolution layer with a convolution kernel size of 1×1 for matrix transformation, and then enters a large convolution kernel convolution layer with a convolution kernel size of 3×3 for matrix transformation, and finally enters a pooling layer with a pooling window of 2×2 for downsampling. After two progressive convolution and downsampling processes, the module can extract the global features of multispectral remote sensing images and synthetic aperture radar remote sensing images respectively, and better realize the extraction of common features of multi-source remote sensing images in the film-covered farmland area. In addition, during the training process, the calculated Euclidean space distance is used as the auxiliary loss function of the module for training. The specific training rules are: when the multispectral remote sensing image samples and the synthetic aperture radar remote sensing image samples input at the same time are samples of the same geographical coordinate area, the Euclidean space distance obtained by step-by-step pooling is required to be the smallest, that is, the auxiliary loss function at this time aims to reduce the information feature space distance between multi-source remote sensing images with the same content information; when the multispectral remote sensing image samples and the synthetic aperture radar remote sensing image samples input at the same time are samples of different geographical coordinate areas, the Euclidean space distance obtained by step-by-step pooling is required to be the largest, that is, the auxiliary loss function at this time aims to expand the information feature space distance between multi-source remote sensing images with different content information. It is necessary to use multi-source remote sensing image samples to train this module separately. After completing the training of this module, when the overall training of the comparative learning model structure is completed, the module can output common features A and common features B, a total of two feature matrices.

[0034] The individual feature extraction module of multispectral remote sensing images and the individual feature extraction module of synthetic aperture radar remote sensing images adopt a similar structural design, both of which are composed of convolutional layers, pooling layers, etc. Figure 4 and Figure 5As shown. In particular, the network layer design of the individual feature extraction module adopts a multi-level feature extraction framework based on "convolution-pooling". Specifically, the image input is first input into a small convolution kernel convolution layer with a convolution kernel size of 1×1 for matrix transformation, and then enters a large convolution kernel convolution layer with a convolution kernel size of 3×3 for matrix transformation, and obtains the feature matrix of the current resolution scale. Subsequently, the feature matrix of the current resolution scale is input into a pooling layer with a pooling window of 2×2 for downsampling, and the above-mentioned convolution and pooling operations are further repeated to obtain remote sensing image feature matrices at three resolution scales. These feature matrices of different resolution scales will provide multi-source remote sensing image feature information of different resolution perspectives for the feature fusion recognition module designed by the present invention. After the training is completed, when the multi-source remote sensing image test sample to be analyzed is input, individual feature output 1, individual feature output 2, individual feature output 3, individual feature output 4, individual feature output 5, and individual feature output 6 can be output respectively.

[0035] The feature fusion recognition module consists of convolutional layer, upsampling layer, merging layer, output layer, etc. Figure 6 As shown. In particular, for the multi-source remote sensing image feature matrix of different resolution scales output by the individual feature extraction module, a step-by-step feature fusion structure design based on "merging-upsampling" is adopted, that is, firstly, the individual feature output and the common feature output obtained after two pooling are fused based on merging and upsampling, and then the fused feature matrix is ​​fused with the individual feature output of other resolutions in turn to obtain a feature matrix containing multi-source remote sensing image feature information of multiple resolutions. Subsequently, the matrix feature information is calculated from global to local through a convolution layer with a large convolution kernel size of 3×3 and a convolution layer with a small convolution kernel size of 1×1, and the final recognition result of film-covered farmland is obtained.

[0036] based on Figure 2 The contrastive learning model structure shown in the figure uses multi-source remote sensing image samples including multispectral remote sensing images and synthetic aperture radar remote sensing images as training samples to train the model. The loss function of the training is binary cross entropy, the optimizer is Adam, the number of iterations is 200, the number of batches is 40, and the learning rate is 0.01. Through the above process, a trained contrastive learning model is obtained.

[0037] Step 3: Remote sensing identification of film-covered farmland based on contrastive learning model:

[0038] The multi-source remote sensing images to be analyzed are input into the trained comparative learning model to obtain the remote sensing identification results of film-covered farmland.

Claims

1. A remote sensing identification method for film-covered farmland based on contrastive learning, characterized in that: The following steps are involved: Step 1) Read in multi-source remote sensing images: Read in multi-source remote sensing image samples including multispectral remote sensing images and synthetic aperture radar remote sensing images, and annotate the true values ​​of the film-covered farmland in the multi-source remote sensing image sample set to provide samples and true values ​​for the next step of comparative learning model training; Step 2) Contrastive learning model training: The contrastive learning model structure consists of four modules, namely, the common feature extraction module of multi-source remote sensing images, the individual feature extraction module of multispectral remote sensing images, the individual feature extraction module of synthetic aperture radar remote sensing images, and the feature fusion recognition module; The common feature extraction module of the multi-source remote sensing image is composed of a convolution layer and a pooling layer; the multispectral remote sensing image and the synthetic aperture radar remote sensing image are respectively subjected to multi-level pooling operations to obtain high-dimensional features; The individual feature extraction module of multispectral remote sensing images and the individual feature extraction module of synthetic aperture radar remote sensing images are both composed of convolutional layers and pooling layers. After training, when the multi-source remote sensing image test sample to be analyzed is input, individual feature output 1, individual feature output 2, individual feature output 3, individual feature output 4, individual feature output 5, and individual feature output 6 are output respectively. The feature fusion recognition module consists of a convolution layer, an upsampling layer, a merging layer, and an output layer. By fusing the common feature outputs and individual feature outputs of other modules, the final output result is obtained and the trained comparative learning model is completed. Step 3) Remote sensing extraction of farmland mulch film based on multimodal semantic learning model: The multi-source remote sensing images to be analyzed are input into the trained comparative learning model to obtain the remote sensing identification results of film-covered farmland.

2. The remote sensing identification method of film-covered farmland based on contrastive learning according to claim 1 is characterized in that: In the training process of the common feature extraction module of multi-source remote sensing images, the calculated Euclidean space distance is used as the auxiliary loss function of the common feature extraction module of multi-source remote sensing images for training; the specific training rules are: when the multi-spectral remote sensing image samples and the synthetic aperture radar remote sensing image samples input at the same time are samples of the same geographical coordinate area, the Euclidean space distance obtained by step-by-step pooling is required to be the smallest, that is, the auxiliary loss function at this time aims to reduce the information feature space distance between multi-source remote sensing images with the same content information; when the multi-spectral remote sensing image samples and the synthetic aperture radar remote sensing image samples input at the same time are samples of different geographical coordinate areas, the Euclidean space distance obtained by step-by-step pooling is required to be the largest, that is, the auxiliary loss function at this time aims to expand the information feature space distance between multi-source remote sensing images with different content information; the common feature extraction module of multi-source remote sensing images is trained separately using multi-source remote sensing image samples; after completing the training, in the overall training of the comparative learning model structure, the common feature extraction module of multi-source remote sensing images outputs common features A and common features B, a total of two feature matrices.

3. The remote sensing identification method of film-covered farmland based on contrastive learning according to claim 1 is characterized in that: Based on the contrastive learning model structure, multi-source remote sensing image samples including multispectral remote sensing images and synthetic aperture radar remote sensing images are used as training samples to train the model. The training loss function is binary cross entropy, the optimizer is Adam, the number of iterations is 200, the batch number is 40, and the learning rate is 0.

01. Through the above process, a trained contrastive learning model is obtained.

4. The remote sensing identification method of film-covered farmland based on contrastive learning according to claim 1 is characterized in that: In the step 2) and step 3), a contrastive learning model is used to perform model training and remote sensing identification of film-covered farmland.

Images