Remote sensing agricultural covering identification method based on spatial feature calibration

Abstract

The invention relates to a remote sensing agricultural covering identification method based on spatial feature calibration, which adopts a multi-scale and multi-depth feature structure instead of a traditional scale change pyramid structure, so that the method has unique advantages and robustness in a point pixel level reconstruction task; according to the method, structural features of different scales and depths are generated by using multiple convolution structures, and deep feature extraction is performed, so that detail and level information of an image can be captured more comprehensively, and a rich and powerful feature basis is provided for subsequent analysis and recognition tasks.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing image pixel level classification, and in particular relates to a remote sensing agricultural cover recognition method with spatial feature calibration. Background Art

[0002] Ground film and greenhouse are two types of agricultural coverings widely used in modern agriculture. Ground film is a plastic film covering the soil surface, which is mainly used for heat preservation, moisture retention, weed inhibition and crop yield increase; greenhouse is a structure covered with transparent materials, which is used to create suitable microclimate conditions, prolong the crop growth period, and improve yield and quality. Both are of great significance in agricultural ecosystems. They can improve agricultural production efficiency and sustainability by regulating soil and air temperature and humidity, reducing pests and diseases and the use of chemical agents. However, the use of ground film and greenhouse also brings about problems such as environmental pollution and waste of resources. Remote sensing ground film land cover mapping uses satellite images and image processing technology to efficiently and accurately monitor the distribution and changes of ground film and greenhouse. This technology provides reliable data support for agricultural management and environmental protection, helps to optimize resource allocation, improve the scientific management level of agricultural production, and achieve sustainable development.

[0003] Although great progress has been made in the identification of greenhouses, the use cycle of mulch is short, usually only used for a short period of one or two months during the germination period of crops, which makes large-scale detection a challenge. Remote sensing technology is considered to be a cost-effective solution to this problem, but although commercial satellites have high spatial resolution, the cost of acquiring large-scale data is very high. The Sentinel-2 satellite provides a low-cost, high-precision remote sensing data resource that can be used for the detection of agricultural coverings. However, since greenhouses and mulch films show similar characteristics in remote sensing images, it is still a difficult problem to identify them at the same time, and relevant research is currently insufficient.

[0004] Therefore, there is an urgent need for a remote sensing agricultural cover identification method that can simultaneously identify and distinguish greenhouses and ground films to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a remote sensing agricultural cover recognition method with spatial feature calibration.

[0006] The spatial feature calibrated remote sensing agricultural cover identification method includes the following steps:

[0007] Step 1: Use multi-scale receptive field convolution to perform preliminary feature extraction on the original remote sensing data to obtain feature structures S of different sizes and depths;

[0008] Step 2: Use group convolution and point convolution to extract the feature structure S to obtain the deep feature D;

[0009] Step 3: Perform self-attention weighting on the deep feature D to obtain the weighted feature, and perform image labeling and position embedding to obtain the weighted feature A;

[0010] Step 4: After aggregating the weighted feature A, use the adaptive feature calibration structure to perform spatial feature correction on the weighted feature A, and extract the marker feature to obtain the final classification result cls.

[0011] Preferably, in step one, a variety of convolution structures are used to generate structural features of different scales and depths, and the feature structure S includes feature S1 with an original size and a number of channels of 64, feature S2 with an original size and a number of channels of 32, and feature S3 with a half size and a number of channels of 64; in step two, group convolution and point convolution of different scales are performed on S1, S2 and S3 to obtain D1, D2 and D3; in step three, feature weighting, image labeling and position embedding are performed on D1, D2 and D3 to obtain A1, A2 and A3.

[0012] Preferably, in step 1, before performing preliminary feature extraction, all bands of the original remote sensing data are resampled to 10 m.

[0013] Preferably, in step 4, A1, A2 and A3 are connected to achieve aggregation, and then the weighted feature A is spatially corrected using an adaptive feature calibration structure to obtain a final classification result cls.

[0014] Preferably, in step 4, global context information is captured by fusing weighted features and pyramid context features, and the region of interest is calibrated by capturing the global context features.

[0015] The beneficial effects of the present invention are:

[0016] 1) The present invention adopts a multi-scale and multi-depth feature structure instead of the traditional scale-varying pyramid structure, which makes the present invention have unique advantages and robustness in point pixel level reconstruction tasks; it uses a variety of convolutional structures to generate structural features of different scales and depths, and performs deep feature extraction, which can more comprehensively capture the details and hierarchical information of the image, thereby providing a rich and powerful feature foundation for subsequent analysis and recognition tasks.

[0017] 2) The present invention improves the model's ability to process deep features and understand image spatial structures through weighted processing, image labeling and position embedding in the spatial perception visual coding module; the global context feature capture module enhances the model's ability to capture global context information by fusing weighted features and pyramid context feature extraction.

[0018] 3) The present invention introduces an accurate region of interest calibration structure, which calibrates the region of interest by capturing global context features, thereby improving the model's attention and recognition ability on the target region, and helping to improve the accuracy and robustness of the model in pixel-level classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of remote sensing agricultural cover identification method calibrated for spatial features;

[0020] Figure 2 Distribution diagram of the collected training and test samples;

[0021] Figure 3 This is the result map of identifying a certain plot of land using this method;

[0022] Figure 4 This is the result map of identifying a certain plot of land using this method;

[0023] Figure 5 This is the result map of identifying a certain plot of land using this method;

[0024] Figure 6 This is the result map of identifying a certain plot of land using this method. DETAILED DESCRIPTION

[0025] The present invention is further described below in conjunction with embodiments. The description of the following embodiments is only used to help understand the present invention. It should be noted that for ordinary persons in the art, without departing from the principle of the present invention, the present invention can also be modified in some ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0026] Embodiment 1

[0027] As an embodiment, the remote sensing agricultural cover recognition method of spatial feature calibration is as follows: Figure 1 As shown, this embodiment is also configured with a corresponding spatial feature calibrated remote sensing agricultural cover identification system and a computer storage medium, in which a computer program is stored; when the computer program is run on a computer, the computer executes this spatial feature calibrated remote sensing agricultural cover identification method.

[0028] The remote sensing agricultural cover identification method with spatial feature calibration includes the following steps:

[0029] Step 1: In this embodiment, Google Earth Engine is used to download Sentinal-2 image data, and the non-10m bands in Sentinal-2 are resampled to 10m. The mathematical expression is:

[0030]

[0031] I_R represents the resampled image, resampling represents the resampling operation, I represents the input Sentinel-2 image, band i Represents the i-th band of the image.

[0032] The original remote sensing data is initially extracted by using multi-scale receptive field convolution in composite convolution to obtain different types of features. The resampled image I_R is firstly extracted by using the multi-scale receptive field convolution structure. The mathematical expression is:

[0033] S1=f RFB1 (I_R)

[0034] S2=f RFB2 (I_R)

[0035] S3=f RFB3 (I_R)

[0036] Among them, S represents the feature structure obtained by using the multi-scale receptive field structure, S1 is the feature with the original size and the number of channels of 64, S2 is the feature with the original size and the number of channels of 32, S3 is the feature with half the size and the number of channels of 64, and f RFB1 、f RFB2 、f RFB3 Represents different multi-scale receptive field structures.

[0037] Many existing models use deeper and more complex structures for preliminary feature extraction, such as dense convolutional network structure, residual dense convolutional network structure, etc., but the computational cost of such methods is huge and is not conducive to practical application. Multi-scale receptive field structure, which uses three different scale receptive field structures to improve feature extraction performance. The specific formula can be expressed as:

[0038] f RFB1 (I_R)=cat(conv3(I R ),conv5(I R ),conv7(I R ))

[0039] Conv3, conv3, and conv5 are convolution operations with receptive field sizes of 3, 5, and 7, respectively.

[0040] Step 2: Use the group convolution and point convolution in the composite convolution to perform group convolution and point convolution operations on the obtained feature S to extract deep features;

[0041] D1=f GDcov1 (S1)

[0042] D2=f GDcov2 (S2)

[0043] D3=f GDcov3 (S3)

[0044] Among them, f GDcov1 、f GDcov2 、f GDcov3 represents group convolution and point convolution structures of different scales, and D represents the extracted deep features.

[0045] Step 3: Use the self-attention module to weight the obtained features, and perform image labeling and position embedding on the image blocks to obtain the weighted feature A. The mathematical expression is:

[0046] A1=f SP1 (D1)

[0047] A2=f SP2 (D2)

[0048] A3=f SP3 (D3)

[0049] Among them, f SP1 、f SP2 、f SP3 Represents spatial perception visual encoding modules of different scales. The main difference between them is the size of weight coefficients. By setting weight coefficients of different sizes, structures of the same size can be generated.

[0050] Step 4: Use the adaptive feature calibration structure to perform spatial feature correction and aggregation on weighted features of different sizes, and obtain the final classification result by labeling features. The mathematical expression is:

[0051] cls=f FC (concat(A1,A2,A3))

[0052] Among them, concat represents the connection operation, f FC represents the adaptive feature calibration module, and cls represents the final classification result.

[0053] f FC By combining self-attention, MLP feature conversion, and deep separable convolution techniques, the effective extraction and fusion of image features are achieved, allowing the model to more accurately capture key information when processing image-related tasks. Through operations such as residual connections and scaling layers, the model can maintain the stability of the training process, thereby achieving good performance in classification and recognition.

[0054] Embodiment 2

[0055] As another embodiment, this embodiment 3 proposes, based on the embodiment 1, a more specific remote sensing agricultural cover recognition method with spatial feature calibration:

[0056] In order to make full use of the spatial-spectral characteristics of remote sensing image pixels and adjacent pixels, in this embodiment, a weighted attention structure is designed to enhance the model's ability to extract contextual information. Different from the traditional method, this embodiment uses adaptive weights of different sizes to achieve weighting of multiple scale features, so that the model can better capture the spatial relationship and spectral information at different scales. SP The specific structure is shown in Table 1.

[0057] Table 1 Pseudo code of spatial perception visual encoding module

[0058]

[0059] It should be noted that the parts in this embodiment that are the same or similar to those in the first embodiment can be referenced to each other and will not be described in detail in this application.

[0060] Embodiment 3

[0061] As another embodiment, this embodiment three proposes, on the basis of embodiments one and two, a remote sensing agricultural cover recognition method using this spatial feature calibration, and by adopting three different settings, compares the classification results of this method with the current four mainstream deep learning methods, namely morphFormer, DHKCLS, MVAHN, VIT, and three machine learning methods, namely support vector machine SVM, random forest RF, and artificial neural network ANN.

[0062] The comparison results are shown in Tables 2 to 4:

[0063] Table 2 Quantitative results 1 (5% training samples, sample block size 16*16)

[0064]

[0065] Table 3 Quantitative results 2 (20% training samples, sample block size 16*16)

[0066]

[0067] Table 4 Quantitative results 3 (20% training samples, sample block size 8*8)

[0068]

[0069] As can be seen from Tables 2 to 4, this method is significantly superior to other methods in quantitative indicators such as the accuracy of each object, average classification accuracy ACA, overall classification accuracy OCA, and Kappa coefficient. By utilizing multi-scale and multi-depth feature structures, unlike the traditional scale-varying pyramid structure, this method can more comprehensively capture the details and hierarchical information of the image, thereby providing a rich and powerful feature basis. Combining the spatial perception visual encoding module and the global context feature capture module, the present invention enhances the model's ability to understand the spatial structure and global context information of the image. In addition, the introduced precise region of interest calibration structure further improves the model's attention to and recognition ability of the target area. In general, the present invention effectively improves the accuracy and robustness in remote sensing agricultural cover recognition tasks.

[0070] It should be noted that the parts in this embodiment that are the same or similar to those in the first embodiment can be referenced to each other and will not be described in detail in this application.

[0071] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

Claims

1. A remote sensing agricultural cover identification method based on spatial feature calibration, characterized in that: The following steps are involved: Step 1: Use multi-scale receptive field convolution to perform preliminary feature extraction on the original remote sensing data to obtain feature structures S of different sizes and depths; Step 2: Use group convolution and point convolution to extract the feature structure S to obtain the deep feature D; Step 3: Perform self-attention weighting on the deep feature D to obtain the weighted feature, and perform image labeling and position embedding to obtain the weighted feature A; Step 4: After aggregating the weighted feature A, use the adaptive feature calibration structure to perform spatial feature correction on the weighted feature A, and extract the marker feature to obtain the final classification result cls.

2. The spatial feature calibrated remote sensing agricultural cover identification method according to claim 1, characterized in that: In step one, a variety of convolution structures are used to generate structural features of different scales and depths. The feature structure S includes feature S1 with an original size and 64 channels, feature S2 with an original size and 32 channels, and feature S3 with half the size and 64 channels. In step two, group convolution and point convolution of different scales are performed on S1, S2, and S3 to obtain D1, D2, and D3. In step three, feature weighting, image labeling, and position embedding are performed on D1, D2, and D3 to obtain A1, A2, and A3.

3. The spatial feature calibrated remote sensing agricultural cover identification method according to claim 1, characterized in that: In step 1, all bands of the original remote sensing data were resampled to 10 m before preliminary feature extraction.

4. The spatial feature calibrated remote sensing agricultural cover identification method according to claim 2, characterized in that: In step 4, A1, A2 and A3 are connected to achieve aggregation, and then the weighted feature A is spatially corrected using the adaptive feature calibration structure to obtain the final classification result cls.

5. The spatial feature calibrated remote sensing agricultural cover identification method according to claim 1, characterized in that: In step 4, the global context information is captured by fusing weighted features and pyramid context feature extraction, and the region of interest is calibrated by capturing the global context features.

Images