A method for extracting hybrid seed maize in high-resolution remote sensing images

Abstract

The present application belongs to the field of crop remote sensing identification technology, and specifically relates to a method for extracting seed corn in high-resolution remote sensing images, comprising the following steps: obtaining high-resolution remote sensing images of seed corn in the target detection area and preprocessing; obtaining cultivated land distribution data; obtaining cultivated land plot vectors; constructing spectral features and texture features, and selecting the optimal features for three-band combination; making seed corn and field corn training data sets; inputting the training data sets into the MIAM‑Net network model for training to obtain a seed corn mask image; converting the seed corn mask image into a vector file and intersecting it with the cultivated land plot vector, and finally calculating the area ratio of seed corn on each cultivated land plot. The method for extracting seed corn in high-resolution remote sensing images of the present application can learn more robust information from the images, so that the deep learning model can make full use of the remote sensing features of seed corn and improve the extraction accuracy of seed corn.

Description

Technical Field

[0001] This application belongs to the technical field of crop remote sensing identification, and particularly relates to a method for extracting hybrid seed maize in high-resolution remote sensing images. Background Art

[0002] Hybrid seed production is an important link in maize production. Mastering the seed production area and yield of major crop seeds is the basis for balancing the seed supply and demand market. High-quality maize seeds are the most important production materials for maize cultivation, and the quality of seeds directly affects the yield, quality of maize and farmers' income. The horizontal distance between a hybrid maize seed production field and other maize pollen sources should be no less than 300 meters. The distribution of hybrid maize seed production fields should be strictly managed in accordance with the regulations formulated by enterprises to ensure the yield and purity of high-quality maize seeds.

[0003] Remote sensing technology is the key driving force to boost the development of modern agriculture. At present, remote sensing image-based hybrid seed maize identification technology has emerged. It mainly identifies the spectral, texture structure, proximity relationship, etc. of hybrid seed maize in remote sensing images according to its planting pattern and phenological characteristics, and then combines existing machine learning and deep learning model algorithms to monitor and manage hybrid seed maize bases.

[0004] However, the existing remote sensing image-based hybrid seed maize identification technology often uses machine learning algorithms such as support vector machines and random forests. However, this kind of algorithm often needs to find optimal parameters to improve the identification accuracy, which will consume a lot of time and energy. Moreover, the identified results of hybrid seed maize cultivated land are often patchy and the boundaries are incomplete, which has a great impact on subsequent analysis work.

[0005] Therefore, there is an urgent need for a technical method that can quickly and effectively obtain the distribution of hybrid seed maize to ensure the smooth progress of the hybrid seed production process. Summary of the Invention

[0006] To solve at least one technical problem existing in the prior art, this application provides a method for extracting hybrid seed maize in high-resolution remote sensing images.

[0007] This application discloses a method for extracting hybrid seed maize in high-resolution remote sensing images, including the following steps:

[0008] Step 1: Obtain high-resolution remote sensing images of hybrid seed maize in the target detection area during the key growth period;

[0009] Step 2: Preprocess the high-resolution remote sensing images;

[0010] Step 3: Obtain the global land cover product of the European Space Agency and extract cultivated land distribution data;

[0011] Step 4: Obtain the vector of cultivated land plots according to the cultivated land distribution data and the cultivated land features in the preprocessed high-resolution remote sensing image.

[0012] Step 5: Construct spectral features and texture features based on the preprocessed high-resolution remote sensing image to obtain a feature image set, and then use the Relief algorithm to select the optimal features from the feature image set for three-band combination to obtain the optimal feature image.

[0013] Step 6: Based on the vector of cultivated land plots and the sample data within the target detection area, using the optimal feature image as the base, produce the training data sets for hybrid maize and field maize.

[0014] Step 7: Input the training data sets into the MIAM-Net network model for training, and use the trained MIAM-Net network model to predict and extract the hybrid maize in the entire optimal feature image to obtain the hybrid maize mask image of the entire high-resolution image.

[0015] Step 8: Convert the hybrid maize mask image into a vector file, then perform an intersection process on this vector file and the vector of cultivated land plots obtained in Step 4. Finally, calculate the proportion ratio of the area of hybrid maize on each cultivated land plot according to the intersection result, and if the score exceeds half, it is judged as hybrid maize information.

[0016] In an alternative embodiment, in Step 1, the high-resolution remote sensing image includes 2-meter resolution blue, green, red, and near-infrared multi-spectral bands and 0.5-meter resolution panchromatic band.

[0017] In an alternative embodiment, in Step 2, the preprocessing of the high-resolution remote sensing image includes:

[0018] Radiometric calibration, atmospheric correction, geometric correction, and image fusion processing.

[0019] In an alternative embodiment, in Step 4, the vector of cultivated land plots is obtained by combining visual interpretation methods.

[0020] In an alternative embodiment, in Step 5, if the spectral feature refers to vegetation index feature, then constructing spectral features based on the preprocessed high-resolution remote sensing image includes:

[0021] Using spectral band operations to calculate vegetation indices for the remote sensing image, and the vegetation indices include Enhanced Vegetation Index (EVI), Normalized Difference Vegetation Index (NDVI), Difference Vegetation Index (DVI), and Ratio Vegetation Index (RVI).

[0022] Among them, the calculation formula for the Enhanced Vegetation Index (EVI) is:

[0023]

[0024] The calculation formula of the normalized difference vegetation index NDVI is as follows:

[0025]

[0026] The calculation formula of the difference vegetation index DVI is as follows:

[0027] DVI = P 1 -P 2 ;

[0028] The calculation formula of the ratio vegetation index RVI is as follows:

[0029]

[0030] In the formula, P 1 is the reflectance in the near-infrared band, P 2 is the reflectance in the red light band, and P 3 is the reflectance in the blue light band.

[0031] In an alternative embodiment, in step five, constructing the texture features based on the preprocessed high-resolution remote sensing image includes:

[0032] Calculating the texture features using the gray-level co-occurrence matrix, and selecting three eigenvalues, namely angular second moment, contrast, and entropy, from the gray-level co-occurrence matrix to quantify the texture features;

[0033] Specifically, calculating the angular second moment texture feature, the formula is as follows:

[0034]

[0035] Calculating the contrast texture feature, the formula is as follows:

[0036]

[0037] Calculating the entropy texture feature, the formula is as follows:

[0038]

[0039] Among them, ASM represents the angular second moment, Con represents the contrast, and Entropy represents the entropy; L is the number of gray levels; P(i,j) is the image after converting to L levels; i is the gray value of the pixel; j is the gray value of the pixel with a fixed length step distance from i.

[0040] In an alternative embodiment, in step five, using the Relief algorithm to select the optimal features from the feature image set for three-band combination to obtain the optimal feature image includes:

[0041] For each sample in the feature image set, a neighboring sample is randomly selected, and then the difference between the class of the selected neighboring sample and the class of the current sample is judged. If the difference between the features is greater, the weight of this feature is increased; conversely, if the difference between the features is smaller, the weight of this feature is decreased. Repeat the above steps until similar weight adjustments have been made for all samples;

[0042] Perform synthesis processing on the top three optimal features ranked by weight to obtain the optimal feature image.

[0043] In an alternative embodiment, step six further includes:

[0044] Set the overlap degree of adjacent slices to 50%, and segment the synthesized optimal feature image to obtain a plurality of image data sets with a size of 512×512;

[0045] Correspondingly, in step seven, the image data set obtained after segmentation is applied through the trained MIAM-Net network model to obtain the corresponding seed production maize mask image, and then stitching is performed to finally obtain the seed production maize mask image of the entire high-resolution image.

[0046] In an alternative embodiment, the MIAM-Net network model includes a multi-information attention module, and the multi-information attention module has three convolutional layers with different kernel sizes, a channel self-attention module, and a spatial self-attention module;

[0047] Correspondingly, in step seven, inputting the training data set into the MIAM-Net network model for training includes:

[0048] Step 701: Parallel process the input feature map through the three convolutional layers in the multi-information attention module to obtain the following three feature maps with different receptive fields:

[0049]

[0050] Among them, F i represents the input feature map, represents a dilated convolution matrix with a dilation rate of 4 under a 3×3 convolution kernel size, represents a dilated convolution matrix with a dilation rate of 3 under a 5×5 convolution kernel size, represents a dilated convolution matrix with a dilation rate of 4 under a 3×3 convolution kernel size, F D3 represents the feature map captured by the 3×3 dilated convolution layer, F D5 represents the feature map captured by the 5×5 dilated convolution layer;

[0051] Step 702: Compress the combined feature maps of F D5 and F D3 into a new feature map F G with a size of 1×1 through global average pooling operation. The obtained F G feature map is expressed as:

[0052]

[0053] where GAP represents global average pooling;

[0054] Step 703: The F G feature map is input into the fully connected layer, and passes through the batch normalization layer and the ReLu layer to generate a new feature map The obtained feature map is expressed as:

[0055]

[0056] where W fc represents the matrix of the fully connected layer, B(·) represents batch normalization, and σ r (·) represents the ReLu activation operation. It undergoes another fully connected operation to obtain a new feature map which can be expressed as:

[0057]

[0058] Step 704: Perform the sigmoid activation function operation on the feature map to obtain the following channel self-attention map α:

[0059]

[0060] Step 705: Obtain the calibrated feature maps of F and F respectively through D3 and F D5 where α is the channel self-attention map of F D3 , α ′ is the channel self-attention map of F D5 , and α ′ is 1 - α;

[0061] Step 706: Perform convolution operations on F S1 =W 1×1 ·F D3 and on F D3 , where W 1×1 represents a 1×1 convolution operation;

[0062] Step 707: Fuse F S1 and feature maps, perform ReLu activation, then perform 1×1 convolution operation and sigmoid activation function operation to obtain the spatial self-attention map β, and the spatial self-attention map β is expressed as:

[0063]

[0064] where σ s represents the sigmoid activation function, and σ r is the ReLU activation function;

[0065] Step 708: Obtain the output result of the multi-information attention module through , where W 1×1 represents performing 1×1 convolution operation, and F S2 and are the feature maps mapped through β and β ′ , and are calculated through and respectively, and β ′ is 1-β.

[0066] In an optional implementation manner, in the eighth step, calculating the area ratio of the hybrid maize on each cultivated land plot according to the intersection result includes:

[0067] Calculate the area of the hybrid maize in the intersection part, and then calculate the area ratio of the hybrid maize according to the following formula:

[0068]

[0069] In the formula, S in is the area ratio of the intersection part of the hybrid maize and the cultivated land, S sd is the area of the hybrid maize, and S f is the area of the cultivated land.

[0070] The present application has at least the following beneficial technical effects:

[0071] The method for extracting hybrid maize in high-resolution remote sensing images of the present application constructs multiple vegetation indices and texture features, uses a feature selection algorithm to select the features that can best reflect hybrid maize, and then adds them to the MIAM-Net algorithm for automatic classification. MIAM-Net can learn more robust information from the images through channel dimension and spatial dimension constraints, enabling the deep learning model to make full use of the remote sensing features of hybrid maize, improving the extraction accuracy of hybrid maize. At the same time, according to the calculation and discrimination of the area ratio of hybrid maize on the cultivated land plots, not only the retention of the relatively complete hybrid maize plot structure is realized, but also the refined management and effective monitoring of hybrid maize are ensured. Description of the Drawings

[0072] Figure 1 is a flowchart of the method for extracting hybrid maize seeds in high-resolution remote sensing images according to the present application;

[0073] Figure 2 is a model structure diagram of the multi-information attention module in the method for extracting hybrid maize seeds in high-resolution remote sensing images according to the present application. Detailed Embodiments

[0074] To make the purpose, technical solutions, and advantages of the implementation of the present application clearer, the technical solutions in the embodiments of the present application will be described in more detail below in conjunction with the embodiments of the present application and the accompanying drawings. The described embodiments are some, but not all, of the embodiments of the present application. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application and should not be construed as limiting the present application.

[0075] The present application discloses a method for extracting hybrid maize seeds in high-resolution remote sensing images, and the specific steps are as Figure 1 shown. Below, a specific example will be used to describe the method of the present application in detail.

[0076] The method for extracting hybrid maize seeds in high-resolution remote sensing images according to the present application includes the following steps:

[0077] Step 101: Obtain high-resolution remote sensing images of hybrid maize seeds in the target detection area during the critical growth period.

[0078] In this embodiment, high-resolution images obtained by the domestic Beijing-3A satellite during the critical growth period of hybrid maize seeds are selected. The satellite adopts a sun-synchronous polar orbit with a revisit period of 3-5 days. The images include multi-spectral bands of blue, green, red, and near-infrared with a resolution of 2 meters and a panchromatic band with a resolution of 0.5 meters.

[0079] Step 102: Preprocess the high-resolution remote sensing images.

[0080] Among them, the preprocessing method can be appropriately selected according to needs. In this embodiment, it is preferable to perform radiometric calibration, atmospheric correction, geometric correction, and image fusion processing on the high-resolution remote sensing images.

[0081] Step 103: Obtain the global land cover product of the European Space Agency and extract the cultivated land distribution data.

[0082] In this embodiment, the type number of the extracted cultivated land distribution data is 40.

[0083] Step 104: According to expert experience, combined with the reference of the cultivated land distribution data and the cultivated land characteristics in the preprocessed high-resolution remote sensing image, and then combined with the visual interpretation method, extract the vector of cultivated land plots with regular boundaries, and try to ensure that the plots are relatively independent.

[0084] Step 105: Based on the preprocessed high-resolution remote sensing image, construct spectral features and texture features to obtain a set of feature images. Then, use the Relief algorithm to select the optimal features from the set of feature images for three-band combination to obtain the optimal feature image.

[0085] Among them, the spectral feature refers to the vegetation index feature. Correspondingly, the construction of spectral features based on the preprocessed high-resolution remote sensing image in this step 105 specifically includes:

[0086] Use spectral band operation to calculate the vegetation index for the remote sensing image. The vegetation index includes Enhanced Vegetation Index (EVI), Normalized Difference Vegetation Index (NDVI), Difference Vegetation Index (DVI), and Ratio Vegetation Index (RVI);

[0087] Among them, the calculation formula for the Enhanced Vegetation Index (EVI) is:

[0088]

[0089] The calculation formula for the Normalized Difference Vegetation Index (NDVI) is:

[0090]

[0091] The calculation formula for the Difference Vegetation Index (DVI) is:

[0092] DVI = P 1 - P 2 ;

[0093] The calculation formula for the Ratio Vegetation Index (RVI) is:

[0094]

[0095] In the formula, P 1 is the reflectance of the near-infrared band, P 2 is the reflectance of the red light band, and P 3 is the reflectance of the blue light band.

[0096] Furthermore, the construction of texture features based on the preprocessed high-resolution remote sensing image in this step 105 includes:

[0097] The Gray Level Co-occurrence Matrix (GLCM) is used to calculate texture features, and three eigenvalues, namely angular second moment, contrast, and entropy, are selected from the gray level co-occurrence matrix to quantify texture features, constructing a feature image set;

[0098] Specifically, the angular second moment texture feature is calculated using the gray level co-occurrence matrix, and the formula is as follows:

[0099]

[0100] The contrast texture feature is calculated, and the formula is as follows:

[0101]

[0102] The entropy texture feature is calculated, and the formula is as follows:

[0103]

[0104] Among them, ASM represents the angular second moment, Con represents the contrast, and Entropy represents the entropy; L is the number of gray levels; P(i, j) is the image after being converted to L levels; i is the gray value of the pixel; j is the gray value of the pixel at a fixed length step away from i.

[0105] It should be noted that the angular second moment can reflect the evenness of the image gray distribution and the coarseness of the texture; the contrast can reflect the clarity of the image and the depth of the texture grooves. The clearer the texture and the greater the contrast, the greater the contrast; the entropy can reflect the randomness contained in the image and represents the complexity of the image; therefore, the above three eigenvalues can better represent the texture features of the hybrid corn in the image.

[0106] Furthermore, in the specific steps of using the Relief algorithm to select the optimal features from the feature image set for three-band combination to obtain the optimal feature image in this step 105, the calculation method of the Relief algorithm is as follows:

[0107]

[0108] Among them, is the weight value of the p-th dimension feature in the i-th iteration, n is the number of samples, diff(p, x, H(x)) represents the difference between sample x and the nearest neighbor of the same class sample in the p-th dimension feature, and diff(p, x, M(x)) represents the difference between sample x and the nearest neighbor of the different class sample in the p-th dimension feature.

[0109] Among them, diff(p, I 1 , I 2 ) is to calculate two samples (i.e., I 1 , I 2)A function for the differences between the features of each dimension. Usually, the absolute error distance calculation is adopted, and the formula is as follows;

[0110]

[0111] Specifically, for the selected 7 feature data (i.e., the enhanced vegetation index, normalized vegetation index, difference vegetation index, ratio vegetation index, angular second moment, contrast, and entropy mentioned above), there may be redundancy among the features, which affects the accuracy of the model algorithm. Therefore, it is necessary to screen these 7 feature data.

[0112] In this embodiment, the sample data includes a total of 63 seed production corn sample points and 24 non-seed production corn sample points. The 7 feature data are extracted to the sample points through the multi-value extraction to point tool of Arcgis, and then the Relief algorithm is implemented by using python to screen the feature data. Among them, the basic idea of the Relief algorithm is to evaluate the importance of features by comparing the feature values of the sample data. The basic steps of the algorithm are as follows: for each sample in the dataset, a neighboring sample is randomly selected; according to the difference between the category of the selected neighboring sample and the category of the current sample, the features of the current sample are adjusted in weight; if the difference between the features is greater, it means that the feature contributes more to distinguishing different categories of samples, so the weight of this feature will increase accordingly; on the contrary, if the difference between the features is smaller, the weight of this feature will decrease; repeat the above steps until similar weight adjustments are made for all samples.

[0113] Then, the features are sorted according to the weights of the features. The features with higher weights are considered more important, and finally, the features with higher rankings are selected as the final feature subset; in this embodiment, the features with the top three weights in Table 1 below are taken (i.e., taking RVI, EVI, and Entropy):

[0114] Table 1 Data analysis table for calculating relief feature weights

[0115] Feature Weight RVI (Ratio Vegetation Index) 0.7931 EVI (Enhanced Vegetation Index) 0.7565 Entropy 0.1711 DVI (Difference Vegetation Index) 0.0851 ASM (Angular Second Moment) 0.0255 NDVI (Normalized Difference Vegetation Index) 0.0221 CON (Contrast) 0.0216

[0116] Finally, the above-selected top three optimal features are synthesized to obtain the optimal feature image.

[0117] Step 106: Based on the vector of the cultivated land plot and the sample data in the target detection area, using the optimal feature image as the base, a training dataset for seed production corn and field corn is made.

[0118] Furthermore, the following steps may also be included in this step 106:

[0119] Set the overlap degree of adjacent slices to 50%, and segment the synthesized optimal feature image to obtain multiple image datasets of size 512×512; among them, the purpose of setting the overlap degree of adjacent slices to 50% is to reduce the influence of tomograms in the results extracted by the subsequent model.

[0120] Step 107: Input the training dataset into the MIAM-Net network model for training, generate a model file of the hybrid maize seeds (i.e., the trained MIAM-Net network model) through training, and predict and extract the hybrid maize seeds in the entire optimal feature image through the trained MIAM-Net network model to obtain the hybrid maize seed mask image of the entire high-resolution image.

[0121] Correspondingly, when the segmentation processing step is included in step 106, in this step 107, the trained MIAM-Net network model is applied to the image dataset obtained after the above segmentation to obtain the corresponding hybrid maize seed mask image, and then stitched to finally obtain the hybrid maize seed mask image of the entire high-resolution image.

[0122] Among them, the MIAM-Net network is an improvement based on the U-Net network structure, with the same core structure as the U-Net network, mainly including four downsamplings, four upsamplings and four splicing layers; different from the U-Net, a multi-information attention module (MIAM) is introduced, which mainly consists of a channel self-attention module and a spatial self-attention module to replace the traditional convolution operation.

[0123] As Figure 2 shown, the multi-information attention module is constructed based on three convolutional layers (the three convolutional layers are 3×3 dilated convolution, 5×5 dilated convolution and 3×3 dilated convolution respectively, and the dilation rates of the dilated convolution are 4, 3, 4 respectively) and combined with the channel self-attention module and the spatial self-attention module;

[0124] Correspondingly, in the step 107, inputting the training dataset into the MIAM-Net network model for training includes:

[0125] Step 701: Parallel process the input feature map through the three convolutional layers in the multi-information attention module to obtain the following feature maps with different receptive fields:

[0126]

[0127] Among them, F i represents the input feature map, represents the dilated convolution matrix with a dilation rate of 4 under a 3×3 convolution kernel size, represents the dilated convolution matrix with a dilation rate of 3 under a 5×5 convolution kernel size, Denote the dilated convolution matrix with a dilation rate of 4 under a 3×3 convolution kernel size as F D3 Denote the feature map captured by the 3×3 dilated convolution layer as F D5 Denote the feature map captured by the 5×5 dilated convolution layer;

[0128] Step 702, compress the combined feature maps of F D5 and F D3 into a new feature map F G with a size of 1×1 through global average pooling operation. The obtained F G feature map is expressed as:

[0129]

[0130] where ⊕ represents element-wise addition of two feature maps, and GAP represents global average pooling;

[0131] Step 703, input the F G feature map into the fully connected layer, and generate a new feature map after passing through the batch normalization layer and the ReLu layer The obtained feature map is expressed as:

[0132]

[0133] where W fc represents the matrix of the fully connected layer, B(·) represents batch normalization, and σ r (·) represents the ReLu activation operation, and then go through another fully connected operation to obtain a new feature map which can be expressed as:

[0134]

[0135] Step 704, perform the sigmoid activation function operation on the feature map to obtain the following channel self-attention map α:

[0136]

[0137] Step 705, respectively obtain the calibrated feature maps of F and through D3 and F D5 (these two channel attention maps can help us adaptively extract more representative feature maps from receptive fields of different scales), where α is the channel self-attention map of F D3 , α ′ is the channel self-attention map of F D5 , and α ′ is 1-α;

[0138] Step 706: Respectively, through F S1 = W 1×1 ·F D3 and perform convolution operations on F D3 , , where W 1×1 represents a 1×1 convolution operation;

[0139] Step 707: Fuse and ReLu activate the feature maps of F S1 and , then perform a 1×1 convolution operation and a sigmoid activation function operation to obtain the spatial self-attention map β. The spatial self-attention map β is expressed as:

[0140]

[0141] where σ s represents the sigmoid activation function, and σ r is the ReLU activation function;

[0142] It should be noted that the channel self-attention module in the spatial self-attention module focuses on the categories of features, and the spatial self-attention module focuses on the positions of features. The spatial self-attention module can further improve the robustness of the network in representing features. The feature maps obtained from the 3×3 convolution layer and the channel self-attention module are used as the input to the spatial self-attention module. Additionally, to refine the position information of the target, we first perform a 1×1 convolution operation on the input feature map.

[0143] Step 708: Obtain the output result of the multi-information attention module through ;

[0144] where W 1×1 represents a 1×1 convolution operation, and the values of β and β ′ (β ′ is 1-β) respectively represent the importance of the channel information of the corresponding elements in and F S1 . F S2 and F C S2 are the feature maps mapped through β and β ′ , and are calculated through and respectively. Then, connect F S2 and to perform a convolution operation to obtain the final output result F out of the spatial self-attention module. F out is also the output result of the entire multi-information attention module (MIAM).

[0145] Step 108: Convert the hybrid maize masking image into a vector file, then perform an intersection process on this vector file and the cultivated land plot vector obtained in Step 104. Finally, calculate the area proportion ratio of hybrid maize on each cultivated land plot according to the intersection result, and if the score exceeds half, it is judged as hybrid maize information.

[0146] Specifically, in this Step 108, calculating the area proportion ratio of hybrid maize on each cultivated land plot according to the intersection result includes:

[0147] In Arcgis, use the intersection tool to apply the hybrid maize vector file and the cultivated land plot vector obtained in Step 104 to obtain the intersection result, then calculate the area of hybrid maize in the intersection part, and then calculate the area proportion of hybrid maize according to the following formula:

[0148]

[0149] In the formula, S in is the area proportion of the intersection part of hybrid maize and cultivated land, S sd is the area of hybrid maize, and S f is the area of cultivated land.

[0150] Finally, in this embodiment, the cultivated land plots with the area proportion of hybrid maize exceeding 50% are selected as the final hybrid maize extraction result.

[0151] The above is only the specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for extracting seed corn from high-resolution remote sensing images, characterized in that: The steps include: Step 1: Obtain high-resolution remote sensing images of seed corn in the target detection area during the key growth period; Step 2: preprocessing the high-resolution remote sensing image; Step 3: Obtain ESA's global land cover products and extract cultivated land distribution data; Step 4: obtaining a cultivated land plot vector according to the cultivated land distribution data and the cultivated land features in the pre-processed high-resolution remote sensing image; Step 5: construct spectral features and texture features based on the preprocessed high-resolution remote sensing image to obtain a feature image set, and then use the Relief algorithm to select the optimal features from the feature image set to perform three-band combination to obtain the optimal feature image; Step 6: Based on the cultivated land plot vector and the sample data in the target detection area, and taking the optimal feature image as the base, a seed corn and field corn training data set is prepared, wherein the sample data includes a plurality of seed corn sample points and a plurality of non-seed corn sample points; Step 7, inputting the training data set into the MIAM-Net network model for training, and predicting and extracting the seed corn in the entire optimal feature image through the trained MIAM-Net network model to obtain the seed corn mask image of the entire high-resolution image, wherein the MIAM-Net network model is based on the U-Net network structure and introduces a multi-information attention module, and the multi-information attention module has a channel self-attention module and a spatial self-attention module; Step eight, converting the seed corn mask image into a vector file, and then intersecting the vector file with the cultivated land plot vector obtained in step four, and finally calculating the area ratio of seed corn on each cultivated land plot based on the intersection result, and determining the information of seed corn if the score exceeds half.

2. The method for extracting seed corn from high-resolution remote sensing images according to claim 1, characterized in that: In the step 1, the high-resolution remote sensing image includes 2-meter resolution blue, green, red, and near-infrared multispectral bands and 0.5-meter resolution panchromatic band.

3. The method for extracting seed corn from high-resolution remote sensing images according to claim 1, characterized in that: In the step 2, preprocessing the high-resolution remote sensing image includes: Radiation calibration, atmospheric correction, geometric correction and image fusion processing.

4. The method for extracting seed corn from high-resolution remote sensing images according to claim 1, characterized in that: In the step 4, the vector of the cultivated land plot is obtained by combining the visual interpretation method.

5. The method for extracting seed corn from high-resolution remote sensing images according to claim 1, characterized in that: In the step 5, the spectral feature refers to the vegetation index feature, and constructing the spectral feature based on the preprocessed high-resolution remote sensing image includes: The vegetation index is calculated from the remote sensing image by using spectral band operation, and the vegetation index includes enhanced vegetation index EVI, normalized difference vegetation index NDVI, difference vegetation index DVI and ratio vegetation index RVI; Among them, the calculation formula of the enhanced vegetation index EVI is: The calculation formula of normalized difference vegetation index NDVI is: The calculation formula of the difference vegetation index DVI is: DVI = P1-P2; The calculation formula of the ratio vegetation index RVI is: Wherein, P1 is the reflectivity of the near-infrared band, P2 is the reflectivity of the red light band, and P3 is the reflectivity of the blue light band.

6. The method for extracting seed corn from high-resolution remote sensing images according to claim 5, characterized in that: In the step 5, constructing texture features based on the preprocessed high-resolution remote sensing image includes: The texture features are calculated by using a gray level co-occurrence matrix, and three eigenvalues, namely, angular second moment, contrast and entropy, are selected from the gray level co-occurrence matrix to quantify the texture features; Specifically, the angular second-order moment texture feature is calculated using the following formula: Calculate the contrast texture feature, the formula is as follows: Calculate the entropy texture feature, the formula is as follows: Among them, ASM stands for angular second moment, Con stands for contrast, and Entropy stands for entropy; L is the number of gray levels; P(i,j) is the image after conversion to level L; i is the gray value of the pixel; j is the gray value of the pixel with a fixed length step distance from i.

7. The method for extracting seed corn from high-resolution remote sensing images according to claim 6, characterized in that: In step 5, the optimal feature is selected from the feature image set by using the Relief algorithm to perform three-band combination, and the optimal feature image obtained includes: For each sample in the feature image set, a neighboring sample is randomly selected, and the difference between the category of the selected neighboring sample and the category of the current sample is determined. If the difference between the features is greater, the weight of the feature is increased. Conversely, if the difference between the features is smaller, the weight of the feature is reduced. The above steps are repeated until similar weight adjustments are made to all samples. The top three optimal features ranked by weight are synthesized to obtain the optimal feature image.

8. The method for extracting seed corn from high-resolution remote sensing images according to claim 1, characterized in that: The step six also includes: The overlap of adjacent slices is set to 50%, and the synthesized optimal feature image is segmented to obtain multiple image data sets of 512×512 size; Correspondingly, in step seven, the trained MIAM-Net network model is used to apply the image data set obtained after segmentation to obtain the corresponding seed corn mask image, which is then spliced ​​to finally obtain the seed corn mask image of the entire high-resolution image.

9. The method for extracting seed corn from high-resolution remote sensing images according to claim 8, characterized in that: The multi-information attention module has three convolutional layers with different kernel sizes, a channel self-attention module and a spatial self-attention module; Accordingly, in step seven, inputting the training data set into the MIAM-Net network model for training includes: Step 701: Process the input feature map in parallel through the three convolutional layers in the multi-information attention module to obtain the following feature maps with different receptive fields: Among them, F i represents the input feature map, represents the dilated convolution matrix with a dilation rate of 4 under a 3×3 convolution kernel size. represents the dilated convolution matrix with a dilation rate of 3 under a 5×5 convolution kernel size. represents the dilated convolution matrix with a dilated rate of 4 under a 3×3 convolution kernel size, F D3 represents the feature map captured by the 3×3 hole convolution layer, F D5 Represents the feature map captured by the 5×5 hole convolution layer; Step 702: Use global average pooling to transform F D5 and F D3 The combined feature map is compressed into a new feature map F of size 1×1 G , the obtained F G The feature map is represented as: Among them, GAP represents global average pooling; Step 703, F G The feature map is input into the fully connected layer, and a new feature map is generated through the batch normalization layer and the ReLu layer. What you get The feature map is represented as: Among them, W fc represents the matrix of the fully connected layer, B(·) represents batch normalization, σ r (·) represents the ReLu activation operation, which undergoes the full connection operation again to obtain a new feature map It can be expressed as: Step 704: feature map Perform the sigmoid activation function operation to obtain the following channel self-attention map α: Step 705: respectively through and Get F D3 and F D5 The calibrated feature map, where α is F D3 The channel self-attention map, α ′ F D5 The channel self-attention map, α ′ is 1-α; Step 706: respectively pass F S1 =W 1×1 ·F D3 and F D3 , Perform convolution operation, where W 1×1 Indicates a 1×1 convolution operation; Step 707: F S1 and The feature map is fused and activated with ReLu, and then a 1×1 convolution operation and a sigmoid activation function operation are performed to obtain the spatial self-attention map β, which is expressed as: Among them, σ s represents the sigmoid activation function, σ r is the ReLU activation function; Step 708: Pass The output result of the multi-information attention module is obtained, where W 1×1 Indicates a 1×1 convolution operation, F S2 and For the β and β ′ The mapped feature maps are respectively and Calculated, β ′ is 1-β.

10. The method for extracting seed corn from high-resolution remote sensing images according to claim 5, characterized in that: In the step eight, calculating the area ratio of seed corn on each cultivated land plot according to the intersection result includes: Calculate the seed corn area of ​​the intersecting part, and then calculate the percentage of seed corn area according to the following formula: In the formula, S in is the proportion of the area where seed corn and cultivated land intersect, S sd is the area of ​​seed corn, S f The area of ​​arable land.

Images