Forest remote sensing image classification method and system based on convolutional neural network

Abstract

The application discloses a forest remote sensing image classification method and system based on a convolutional neural network, and relates to the technical field of remote sensing image processing. The method steps comprise the following steps: S1, a de-fogging algorithm is used to remove thin clouds and fog in a remote sensing image, and the brightness of the remote sensing image is enhanced; S2, the edge contour of a thick cloud layer shielding area in the remote sensing image is detected, and an initial mask of the thick cloud layer shielding area in the remote sensing image is obtained; S3, the thick cloud layer shielding area in the remote sensing image is replaced through global search matching, and the restoration of a ground object in the remote sensing image that is shielded by the thick cloud layer is realized; and S4, a convolutional neural network is used to extract features of the restored remote sensing image, and the detection and classification of forest types and pest conditions in the remote sensing image are realized. The application realizes the accurate detection of forest types and pest conditions by using high spatial resolution images in combination with a convolutional neural network.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a method and system for classifying forest remote sensing images based on convolutional neural networks. Background Technology

[0002] With the development of technology, high-resolution remote sensing imagery has matured. Data acquired using high-resolution remote sensing images exhibits clear textures and rich details, thus finding wide application in fields such as precision agriculture, weather forecasting, and natural disaster monitoring. Forest type identification is a crucial aspect of forest resource management and monitoring, and high-resolution remote sensing imagery is widely used for forest observation. However, current applications of remote sensing technology in forest type identification still face many challenges requiring further research and solutions: First, atmospheric clouds often appear in remote sensing images, obscuring ground features and resulting in incomplete forest images with missing usable information, significantly impacting the accuracy of forest type classification. Second, due to the vastness of forest areas, pest problems are often difficult to detect in a timely manner, interfering with the identification of forest types in remote sensing images and causing economic losses.

[0003] Chinese patent CN109034189B discloses a forest type identification method based on high-resolution remote sensing imagery. The method involves preprocessing the high-resolution remote sensing imagery to obtain panchromatic and multispectral images. Gray levels are extracted from the panchromatic imagery, and vegetation indices are extracted from the multispectral imagery. By utilizing the texture features of the Gaofen-2 imagery to obtain gray levels and vegetation indices, forest types can be identified. However, this method fails to address the issue of cloud cover in the remote sensing imagery.

[0004] Therefore, this invention proposes a forest land remote sensing image classification method and system based on convolutional neural networks. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a forest land remote sensing image classification method and system based on convolutional neural networks. This method aims to mitigate the interference of cloud cover and forest pests in remote sensing images on forest land type classification, improve the accuracy of forest land remote sensing image classification, and enable timely detection of forest pests.

[0006] The technical solution adopted in this invention is:

[0007] A forest land remote sensing image classification method based on convolutional neural networks includes the following steps:

[0008] S1. Use dehazing algorithms to remove thin clouds and fog from remote sensing images and enhance the brightness of remote sensing images;

[0009] S2. Detect the edge contour of the area obscured by thick clouds in the remote sensing image to obtain the initial mask of the area obscured by thick clouds in the remote sensing image.

[0010] S3. Replace the areas obscured by thick clouds in the remote sensing image through global search and matching, thereby restoring the ground features obscured by thick clouds in the remote sensing image.

[0011] S4. Use convolutional neural networks to extract features from the restored remote sensing images to detect and classify forest types and pest conditions in the remote sensing images.

[0012] A further improvement of the present invention is that the specific steps of replacing the areas obscured by thick clouds in the remote sensing image through global search and matching in step S3, thereby restoring the ground features obscured by thick clouds in the remote sensing image, are as follows:

[0013] S31. Obtain the edge contour of the area obscured by thick clouds in the remote sensing image based on the mask of the area obscured by thick clouds in the remote sensing image.

[0014] S32. Calculate the filling priority of the edge contour pixels, and construct a pixel template block with the pixel with the highest filling priority as the center.

[0015] S33. Search and match the area in the remote sensing image other than the area covered by thick clouds to obtain the pixel block with the highest similarity to the pixel template block, and copy the pixel block with the highest similarity to fill the area covered by thick clouds in the pixel template block.

[0016] S34. Update the mask and determine if the mask is empty. If it is not empty, proceed to step S11.

[0017] S35. Detect boundary effects in the original cloud-occluded areas of the repaired remote sensing image, and use the Telea algorithm to repair areas with boundary effects.

[0018] A further improvement of the present invention is that the calculation formula for the filling priority of the edge contour pixels in step S32 is as follows:

[0019]

[0020] Where P(p) is the fill priority of pixel p, Ψ p For a pixel block constructed with pixel point p as the center, |Ψ p | For pixel block Ψ p The number of pixels in the image is Φ, which represents the area in the remotely sensed image excluding thick cloud cover. C(q) is the confidence score of pixel q, which is 0 when pixel q is located in the area to be repaired, and 1 otherwise. Used to measure the number of known pixels contained in a pixel block constructed centered on pixel p; Let n be the direction vector of the isoluminance line for pixel p. p Let be the normal direction vector of pixel p, and λ be the normalization factor. Let α be the structural information of pixel p; α and β are their weights.

[0021] A further improvement of the present invention is that the similarity measurement method described in step S33 is specifically by measuring the color difference and chromaticity difference between two pixel blocks. The smaller the difference, the higher the similarity between the two pixel blocks.

[0022] Pixel block Ψ p and pixel block Ψ q Color difference C(Ψ) p ,Ψ q The formula for calculating ) is:

[0023]

[0024] Where a and b are pixel blocks Ψ p and pixel block Ψ q In the RGB color space, R(a), G(a), and B(a) are the R, G, and B values ​​of pixel a, respectively, and R(b), G(b), and B(b) are the R, G, and B values ​​of pixel b, respectively, in the RGB color space.

[0025] Pixel block Ψ p and pixel block Ψ q Color difference H(Ψ) p ,Ψ q The formula for calculating ) is:

[0026]

[0027] Where a and b are pixel blocks Ψ p and pixel block Ψ q In the image, S(a) and S(b) represent the chromaticity information of pixel a and pixel b, respectively.

[0028] Pixel block Ψ p and pixel block Ψ q The formula for calculating similarity is:

[0029] D(Ψ p ,Ψ q ) = w c C(Ψ p ,Ψ q )+w h H(Ψ p ,Ψ q ),

[0030] Among them, w c and w hAs the weight for color difference and chromaticity difference, w c >w h .

[0031] A further improvement of this invention lies in the following steps: Step S4 involves using a convolutional neural network to extract features from the repaired remote sensing image, thereby enabling the detection and classification of forest land types and pest conditions in the remote sensing image.

[0032] S41. Collect forest remote sensing images at different resolutions, and label and crop them. The labeled categories are: arbor forest, insect pests_arbor forest, bamboo forest, insect pests_bamboo forest, special shrub forest, insect pests_special shrub forest, sparse forest, insect pests_sparse forest, shrub forest, and insect pests_shrub forest. Crop the remote sensing images to 600×600 pixels according to the categories. Divide the remote sensing images that have been labeled and cropped into training and testing sets for training convolutional neural network classification models at a ratio of 4:1.

[0033] S42. Train the convolutional neural network classification model and iteratively update the network parameters using the Adam optimizer until a preset number of training iterations or the change in the loss function is less than a threshold is reached to obtain the optimal convolutional neural network classification model.

[0034] S43. Upsample and downsample the remote sensing image to be measured to obtain three remote sensing images at different scales, and subtract them from the mean remote sensing image to obtain a multi-scale remote sensing image to be measured.

[0035] S44. Input the multi-scale remote sensing images to be tested into the trained convolutional neural network classification model, and use the forward propagation method to extract the features of the fully connected layer of the remote sensing images as the global features of the remote sensing images to input the linear SVM classifier for scene classification, so as to realize the detection and classification of forest land type and pest status in the remote sensing images.

[0036] This invention also provides a forest remote sensing image classification system based on convolutional neural networks. The system includes: a thin cloud removal module, a ground feature restoration module, and a forest classification module. The thin cloud removal module is used to remove thin clouds and fog from remote sensing images and enhance the brightness of remote sensing images. The ground feature restoration module is used to restore ground features in remote sensing images that are obscured by thick cloud layers. The forest classification module is used to perform fine division of forest areas in remote sensing images and pest detection.

[0037] A further improvement of the present invention is that the feature restoration module includes a pixel determination unit, a search and matching unit, and a pixel restoration unit; the pixel determination unit is used to determine the pixel with the highest filling priority among the edge contour pixels; the search and matching unit is used to search and match the pixel block with the highest similarity to the pixel template block in the remote sensing image except for areas obscured by thick clouds; the pixel restoration unit is used to restore the pixel with the highest filling priority using the search and matching pixel block with the highest similarity.

[0038] A further improvement of this invention is that the forest land classification module includes a remote sensing image processing unit, a classification model training unit, and a remote sensing image classification unit; the remote sensing image processing unit is used to classify and crop forest land remote sensing images of different resolutions, and divide them proportionally into a training set and a test set for training a convolutional neural network classification model; the classification model training unit is used to train the convolutional neural network classification model to obtain the optimal convolutional neural network classification model; the remote sensing image classification unit is used to process the remote sensing image to be tested and input it into the trained convolutional neural network classification model to detect and classify forest land types and pest conditions in the remote sensing image.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. Construct a pixel template block centered on the pixel with the highest priority for filling the edge contour of the occluded area; in the remote sensing image, except for the area occluded by thick clouds, search and match based on color similarity and chromaticity similarity to obtain the pixel block with the highest similarity to the pixel template block and copy it to fill the area occluded by thick clouds in the pixel template block. Through iterative processing of this process, the restoration of land features occluded by thick clouds in the remote sensing image is achieved, and the accuracy of forest type classification in the remote sensing image is improved.

[0041] 2. A convolutional neural network classification model is trained using forest remote sensing images at different resolutions; multi-scale transformation is performed on the restored remote sensing images to be tested; the trained convolutional neural network classification model is used to extract features from the multi-scale remote sensing images to be tested, and the features of its fully connected layers are used as global features of the remote sensing images to input into a linear SVM classifier for scene classification, thereby realizing the detection and classification of forest types and pest conditions in remote sensing images. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the forest remote sensing image classification method based on convolutional neural networks of the present invention.

[0043] Figure 2 This is a schematic diagram illustrating the process of replacing areas obscured by thick cloud cover in remote sensing images in the forest remote sensing image classification method based on convolutional neural networks of the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the process of using convolutional neural networks to detect and classify forest types and pest conditions in remote sensing images in the forest remote sensing image classification method based on convolutional neural networks of the present invention.

[0045] Figure 4 This is a schematic diagram illustrating the principle framework of the forest remote sensing image classification system based on convolutional neural networks of the present invention.

[0046] Figure 5 This is a schematic diagram of the land cover restoration module of the forest land remote sensing image classification system based on convolutional neural networks of the present invention;

[0047] Figure 6 This is a schematic diagram of the forest land classification module in the forest land remote sensing image classification system based on convolutional neural networks of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] In this embodiment, as Figure 1 As shown, a forest remote sensing image classification method based on convolutional neural networks is provided, including the following steps:

[0051] S1. Use dehazing algorithms to remove thin clouds and fog from remote sensing images and enhance the brightness of remote sensing images;

[0052] S2. Detect the edge contour of the area obscured by thick clouds in the remote sensing image to obtain the initial mask of the area obscured by thick clouds in the remote sensing image.

[0053] S3. Replace the areas obscured by thick clouds in the remote sensing image through global search and matching, thereby restoring the ground features obscured by thick clouds in the remote sensing image.

[0054] S4. Use convolutional neural networks to extract features from the restored remote sensing images to detect and classify forest types and pest conditions in the remote sensing images.

[0055] In this embodiment, as Figure 2 As shown, step S3, which involves replacing areas obscured by thick clouds in the remote sensing image through global search and matching to restore features obscured by thick clouds, specifically involves the following steps:

[0056] S31. Obtain the edge contour of the area obscured by thick clouds in the remote sensing image based on the mask of the area obscured by thick clouds in the remote sensing image.

[0057] S32. Calculate the filling priority of the edge contour pixels, and construct a pixel template block with the pixel with the highest filling priority as the center.

[0058] S33. Search and match the area in the remote sensing image other than the area covered by thick clouds to obtain the pixel block with the highest similarity to the pixel template block, and copy the pixel block with the highest similarity to fill the area covered by thick clouds in the pixel template block.

[0059] S34. Update the mask and determine if the mask is empty. If it is not empty, proceed to step S11.

[0060] S35. Detect boundary effects in the original cloud-occluded areas of the repaired remote sensing image, and use the Telea algorithm to repair areas with boundary effects.

[0061] In this embodiment, the formula for calculating the filling priority of edge contour pixels in step S32 is as follows:

[0062]

[0063] Where P(p) is the fill priority of pixel p, Ψ p For a pixel block constructed with pixel point p as the center, |Ψ p | For pixel block Ψ p The number of pixels in the image is Φ, which represents the area in the remotely sensed image excluding thick cloud cover. C(q) is the confidence score of pixel q, which is 0 when pixel q is located in the area to be repaired, and 1 otherwise. Used to measure the number of known pixels contained in a pixel block constructed with pixel p as the center; Let n be the direction vector of the isoluminance line for pixel p. p Let be the normal direction vector of pixel p, and λ be the normalization factor. Let α be the structural information of pixel p; α and β are their weights.

[0064] In this embodiment, the similarity measurement method described in step S33 is specifically by measuring the color difference and chromaticity difference between two pixel blocks. The smaller the difference, the higher the similarity between the two pixel blocks.

[0065] Pixel block Ψ p and pixel block Ψ q Color difference C(Ψ) p ,Ψ q The formula for calculating ) is:

[0066]

[0067] Where a and b are pixel blocks Ψ p and pixel block Ψ q In the RGB color space, R(a), G(a), and B(a) are the R, G, and B values ​​of pixel a, respectively, and R(b), G(b), and B(b) are the R, G, and B values ​​of pixel b, respectively, in the RGB color space.

[0068] Pixel block Ψ p and pixel block Ψ q Color difference H(Ψ) p ,Ψ q The formula for calculating ) is:

[0069]

[0070] Where a and b are pixel blocks Ψ p and pixel block Ψ q In the image, S(a) and S(b) represent the chromaticity information of pixel a and pixel b, respectively.

[0071] Pixel block Ψ p and pixel block Ψ q The formula for calculating similarity is:

[0072] D(Ψ p ,Ψ q ) = w c C(Ψ p ,Ψ q )+w h H(Ψ p ,Ψ q ),

[0073] Among them, w c and w h As the weight for color difference and chromaticity difference, w c >w h .

[0074] Convolutional Neural Networks (CNNs) are end-to-end classification models that utilize supervised learning to directly learn class labels from raw images. They eliminate the need for manual feature design and can learn feature representations using large amounts of training data. A typical CNN model's basic structure includes convolutional layers, nonlinearization layers, pooling layers, and fully connected layers. Feature extraction primarily occurs through convolutional layers, which perform convolution operations with the input feature map using kernels to obtain the output feature map. Through successive convolutional layers, features are extracted from shallow to deep layers of the image. Nonlinearization layers apply nonlinear transformations to each pixel of the input feature map using activation functions, increasing the network's expressive power. Pooling layers reduce the spatial dimensionality of the feature map through downsampling. Fully connected layers fully connect all neurons in the previous layer, transforming the two-dimensional feature map output after multiple convolutions, nonlinearizations, and pooling into a one-dimensional feature vector, learning the global features of the image. The final fully connected layer is typically a softmax layer, which acts as a classifier, converting the input features into class probabilities, thus mapping the previously learned features to the sample label space and completing the classification task.

[0075] In this embodiment, as Figure 3 As shown, the specific steps in step S4, which involve using a convolutional neural network to extract features from the repaired remote sensing image to detect and classify forest types and pest conditions in the remote sensing image, are as follows:

[0076] S41. Collect forest remote sensing images at different resolutions, and label and crop them. The labeled categories are: arbor forest, insect pests_arbor forest, bamboo forest, insect pests_bamboo forest, special shrub forest, insect pests_special shrub forest, sparse forest, insect pests_sparse forest, shrub forest, and insect pests_shrub forest. The specific category settings and sample number for each category of forest remote sensing images are shown in Table 1. The remote sensing images are cropped to 600×600 pixels according to the categories. The remote sensing images that have been labeled and cropped are divided into training set and test set for training convolutional neural network classification model at a ratio of 4:1.

[0077] Table 1. Classification and Sample Quantity of Forest Land Remote Sensing Images

[0078] Category Sample size Category Sample size Arbor forest 210 Pests in arbor forests 167 Bamboo Forest 302 Pests in bamboo forests 205 Special shrubland 238 Pests - Special shrubland 148 sparse woodland 333 Pests - Sparse Woodland 190 shrubland 230 Pests - Shrubland 200

[0079] S42. Train the convolutional neural network classification model and iteratively update the network parameters using the Adam optimizer until a preset number of training iterations or the change in the loss function is less than a threshold is reached to obtain the optimal convolutional neural network classification model.

[0080] S43. Upsample and downsample the remote sensing image to be measured to obtain three remote sensing images at different scales, and subtract them from the mean remote sensing image to obtain a multi-scale remote sensing image to be measured.

[0081] When extracting features from fully connected layers as global features, convolutional neural network classification models require the input image to be of a fixed size. However, adjusting the size of the remote sensing image to be tested will lead to a loss of spatial resolution and some detailed information of the remote sensing image to be tested. Therefore, the remote sensing image to be tested is upsampled and downsampled separately, and subtracted from the mean image respectively to obtain a multi-scale remote sensing image to be tested, thereby reducing the loss of spatial information of the remote sensing image to be tested.

[0082] S44. Input the multi-scale remote sensing images to be tested into the trained convolutional neural network classification model, and use the forward propagation method to extract the features of the fully connected layer of the remote sensing images as the global features of the remote sensing images. Input the linear SVM (Support Vector Machine) classifier to classify the scene, so as to realize the detection and classification of forest land type and pest status in the remote sensing images.

[0083] SVM is a generalized linear classifier that performs binary classification of data using supervised learning. Standard SVM cannot directly handle multi-class classification problems. For multi-class classification problems, this embodiment utilizes standard SVM to systematically construct multiple decision boundaries to achieve multi-class classification of samples.

[0084] Example 2

[0085] In this embodiment, as Figure 4 As shown, a forest remote sensing image classification system based on convolutional neural networks is provided. The system includes: a thin cloud removal module, a ground feature restoration module, and a forest classification module. The thin cloud removal module is used to remove thin clouds and fog from the remote sensing image and enhance the brightness of the remote sensing image. The ground feature restoration module is used to restore ground features in the remote sensing image that are obscured by thick cloud layers. The forest classification module is used to perform fine division of forest areas in the remote sensing image and pest detection.

[0086] In this embodiment, as Figure 5 As shown, the feature restoration module includes a pixel determination unit, a search and matching unit, and a pixel restoration unit. The pixel determination unit is used to determine the pixel with the highest filling priority among the edge contour pixels. The search and matching unit is used to search and match the pixel block with the highest similarity to the pixel template block in the remote sensing image, except for areas obscured by thick clouds. The pixel restoration unit is used to restore the pixel with the highest filling priority using the searched and matched pixel block with the highest similarity.

[0087] In this embodiment, as Figure 6As shown, the forest land classification module includes a remote sensing image processing unit, a classification model training unit, and a remote sensing image classification unit. The remote sensing image processing unit is used to label and crop forest land remote sensing images of different resolutions, and divide them proportionally into training and testing sets for training convolutional neural network classification models. The classification model training unit is used to train convolutional neural network classification models to obtain the optimal convolutional neural network classification model. The remote sensing image classification unit is used to process the remote sensing images to be tested and input them into the trained convolutional neural network classification model to perform accurate classification of forest land types and pest conditions in the remote sensing images.

[0088] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A forest land remote sensing image classification method based on convolutional neural networks, characterized in that, The method includes the following steps: S1. Use dehazing algorithms to remove thin clouds and fog from remote sensing images and enhance the brightness of remote sensing images; S2. Detect the edge contour of the area obscured by thick clouds in the remote sensing image to obtain the initial mask of the area obscured by thick clouds in the remote sensing image. S3. Replace the areas obscured by thick clouds in the remote sensing image through global search and matching, thereby restoring the ground features obscured by thick clouds in the remote sensing image. S4. Use convolutional neural networks to extract features from the restored remote sensing images to detect and classify forest types and pest conditions in the remote sensing images. Step S3, which involves replacing areas obscured by thick clouds in remote sensing images through global search and matching to restore features obscured by thick clouds, consists of the following specific steps: S31. Obtain the edge contour of the area obscured by thick clouds in the remote sensing image based on the mask of the area obscured by thick clouds in the remote sensing image. S32. Calculate the filling priority of the edge contour pixels, and construct a pixel template block with the pixel with the highest filling priority as the center. S33. Search and match the area in the remote sensing image other than the area covered by thick clouds to obtain the pixel block with the highest similarity to the pixel template block, and copy the pixel block with the highest similarity to fill the area covered by thick clouds in the pixel template block. S34. Update the mask and determine if the mask is empty. If it is not empty, go to step S31. S35. Detect boundary effects in the original cloud-occluded areas of the restored remote sensing image, and use the Telea algorithm to restore areas with boundary effects. The formula for calculating the filling priority of edge contour pixels in step S32 is as follows: ， in, The fill priority for pixel p. A pixel block constructed with pixel point p as the center. For pixel blocks Number of pixels This refers to the area in the remote sensing image excluding the area obscured by thick clouds. Let be the confidence score of pixel q, when pixel q is located in the area to be repaired. It is 0 otherwise it is 1. The number of known pixels contained in the pixel block constructed with pixel p as the center; Let be the isoluminance line direction vector of pixel p. Let p be the normal direction vector. As the normalization factor, The structural information content of pixel p; and It is the weight of the two; The similarity measurement method described in step S33 is specifically by measuring the color difference and chromaticity difference between two pixel blocks. The smaller the difference, the higher the similarity between the two pixel blocks. Pixel Blocks and pixel blocks Color differences The calculation formula is: ， Where a and b are pixel blocks respectively. and pixel blocks In the RGB color space, R(a), G(a), and B(a) are the R, G, and B values ​​of pixel a, respectively, and R(b), G(b), and B(b) are the R, G, and B values ​​of pixel b, respectively, in the RGB color space. Pixel Blocks and pixel blocks Color difference The calculation formula is: ， Where a and b are pixel blocks respectively. and pixel blocks In the image, S(a) and S(b) represent the chromaticity information of pixel a and pixel b, respectively. Pixel Blocks and pixel blocks The formula for calculating similarity is: ， in, and Weights for color difference and chromaticity difference. > .

2. The forest land remote sensing image classification method based on convolutional neural networks as described in claim 1, characterized in that, Step S4, which describes using a convolutional neural network to extract features from the restored remote sensing image to detect and classify forest types and pest conditions in the remote sensing image, involves the following specific steps: S41. Collect forest remote sensing images at different resolutions, and label and crop them according to their categories. The labeled categories are: arbor forest, insect pests_arbor forest, bamboo forest, insect pests_bamboo forest, special shrub forest, insect pests_special shrub forest, sparse forest, insect pests_sparse forest, shrub forest, and insect pests_shrub forest. Crop the remote sensing images to 600×600 pixels according to the categories. Divide the remote sensing images that have been labeled and cropped into training set and test set at a ratio of 4:1 for training the convolutional neural network classification model. S42. Train the convolutional neural network classification model and iteratively update the network parameters using the Adam optimizer until a preset number of training iterations or a change in the loss function is less than a threshold, thus obtaining the optimal convolutional neural network classification model. S43. Upsample and downsample the remote sensing image to be measured to obtain three remote sensing images at different scales, and subtract them from the mean remote sensing image to be measured to obtain a multi-scale remote sensing image to be measured. S44. Input the multi-scale remote sensing images to be tested into the trained convolutional neural network classification model, and use forward propagation to extract the features of the fully connected layer of the remote sensing images as global features of the remote sensing images. Input the features into the linear SVM classifier for scene classification, so as to realize the detection and classification of forest land type and pest status in the remote sensing images.

3. A forest land remote sensing image classification system based on convolutional neural networks, implemented based on the forest land remote sensing image classification method based on convolutional neural networks as described in any one of claims 1-2, characterized in that, The system includes: a thin cloud removal module, a ground feature restoration module, and a forest land classification module; the thin cloud removal module is used to remove thin clouds and fog from remote sensing images and enhance the brightness of remote sensing images; the ground feature restoration module is used to restore ground features in remote sensing images that are obscured by thick cloud layers; the forest land classification module is used to perform fine division of forest areas in remote sensing images and pest detection.

4. The forest land remote sensing image classification system based on convolutional neural networks as described in claim 3, characterized in that, The feature restoration module includes a pixel determination unit, a search and matching unit, and a pixel restoration unit. The pixel determination unit is used to determine the pixel with the highest filling priority among the edge contour pixels. The search and matching unit is used to search and match the pixel block with the highest similarity to the pixel template block in the remote sensing image, except for areas obscured by thick clouds. The pixel restoration unit is used to restore the pixel with the highest filling priority using the search and matching pixel block with the highest similarity.

5. The forest land remote sensing image classification system based on convolutional neural networks as described in claim 3, characterized in that, The forest land classification module includes a remote sensing image processing unit, a classification model training unit, and a remote sensing image classification unit. The remote sensing image processing unit is used to classify and crop forest land remote sensing images of different resolutions, and divide them proportionally into training and testing sets for training convolutional neural network classification models. The classification model training unit is used to train convolutional neural network classification models to obtain the optimal convolutional neural network classification model. The remote sensing image classification unit is used to process the remote sensing images to be tested and input them into the trained convolutional neural network classification model to perform fine classification of forest land types and pest conditions in the remote sensing images.

Images