Tree damage range intelligent identification method and system based on satellite remote sensing image

Abstract

The invention provides a tree damage range intelligent identification method and system based on a satellite remote sensing image, and the method comprises the steps: determining a disaster cut image based on a satellite remote sensing image of a disaster area; fusing different spectral band data with the disaster cut image by adopting a multi-spectral band fusion technology to obtain a fused image; on the basis of the fused image and in combination with multiple indexes, a color mapping technology and a clustering algorithm are adopted for combined comparative analysis, and a post-disaster damage area preliminary screening image is obtained; inputting the post-disaster damage area preliminary screening image into a pre-constructed refined segmentation model to obtain a post-disaster tree damage detection result; performing spatial overlay analysis on the post-disaster tree damage detection result and power grid geographic data to obtain a risk identification result of a tree damage range; wherein the refined segmentation model comprises an encoder part adopting a visual architecture based on a shift window mechanism. The damage condition of trees around the power transmission line can be rapidly obtained after a disaster occurs, and the line patrol time is shortened.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent identification, and particularly relates to a method and system for intelligently identifying a damaged range of trees based on satellite remote sensing images. BACKGROUND

[0002] Damage to trees along power transmission lines can affect the safe operation of the power grid. In extreme weather (such as typhoons, blizzards, thunderstorms, and the like) and geological disasters (such as landslides, mudslides, and earthquakes), a large area of trees can be broken, bent, or uprooted, and in severe cases, the power transmission lines can be crushed, power equipment can be damaged, short circuits can occur, and even large-scale power outages can occur. Therefore, after a disaster occurs, the power department needs to quickly and accurately identify the range of damaged trees in the power transmission channel in order to carry out emergency repair, develop a clearing scheme, and restore power supply, and ensure the safe and stable operation of the power grid.

[0003] Currently, the inspection of power transmission channels in the power grid mainly relies on manual line inspection, unmanned aerial vehicle inspection, and helicopter aerial photography. However, these methods have certain limitations: manual line inspection is time-consuming and inefficient, and it is difficult to cover a large range of mountainous power transmission lines. SUMMARY

[0004] To overcome the deficiencies of the prior art, in a first aspect, the present application provides a method for intelligently identifying a damaged range of trees based on satellite remote sensing images, comprising:

[0005] Based on satellite remote sensing images of a disaster area, and in combination with the distribution characteristics of tree disaster conditions, a disaster-affected cropped image is determined;

[0006] Different spectral band data are fused with the disaster-affected cropped image using a multi-spectral band fusion technology to obtain a fused image; based on the fused image, in combination with multiple indexes, a color mapping technology and a clustering algorithm are used for joint comparative analysis to obtain a post-disaster damage area preliminary screening image;

[0007] The post-disaster damage area preliminary screening image is input into a pre-constructed refinement segmentation model to obtain a post-disaster tree damage detection result;

[0008] The post-disaster tree damage detection result is subjected to spatial overlay analysis with power grid geographic data to obtain a risk identification result of the damaged range of trees;

[0009] The refinement segmentation model includes an encoder part using a visual architecture based on a shift window mechanism.

[0010] Preferably, the input of the post-disaster damage area preliminary screening image into the pre-constructed refinement segmentation model to obtain the post-disaster tree damage detection result comprises:

[0011] determine the damage area in the image based on the post-disaster damage area preliminary screening image;

[0012] determine the damage degree index and the enhanced damage degree index according to the damage degree of the damage area;

[0013] label the damage area in the post-disaster damage area preliminary screening image based on the damage degree index and the enhanced damage degree index;

[0014] input the labeled damage area into the refined segmentation model to obtain a post-disaster tree damage detection result.

[0015] Preferably, the pre-construction process of the refined segmentation model comprises:

[0016] replace the convolutional layer of the encoder part of the deep learning segmentation model with a visual architecture based on a shift window mechanism, and add an attention mechanism module to the decoder part of the deep learning segmentation model to form an improved deep learning segmentation model;

[0017] combine the cross-entropy loss function and the semantic segmentation loss function to obtain a combined loss function of the improved deep learning segmentation model;

[0018] train the improved deep learning segmentation model based on the dataset with damage area labeling and the combined loss function to obtain a refined segmentation model.

[0019] Preferably, the post-disaster damage area preliminary screening image obtained by combining the fusion image with multiple indexes and using color mapping technology and clustering algorithm for joint comparative analysis comprises:

[0020] Based on the fusion image combined with multiple indexes, the spectrum in the fusion image is processed using color mapping technology to obtain the brightness channel value of different regions;

[0021] The brightness channel value of different regions is analyzed using a clustering algorithm to obtain a post-disaster damage area preliminary screening image.

[0022] Preferably, the brightness channel value of different regions obtained by processing the spectrum in the fusion image using color mapping technology based on the fusion image combined with multiple indexes comprises:

[0023] Based on the fusion image combined with multiple indexes, the RGB color space of the spectrum in the fusion image is converted to Lab color space using color mapping technology to form Lab color space;

[0024] The Lab color space is calculated using a brightness channel calculation formula to obtain the brightness channel value of different regions.

[0025] Preferably, the damage area is determined according to the damage degree, and the damage degree index and the enhanced damage degree index are determined, comprising:

[0026] The damage degree index is determined based on the pre-disaster normalized difference vegetation index and the post-disaster normalized difference vegetation index of the damage area.

[0027] The enhanced damage degree index is determined based on the pre-disaster enhanced vegetation index and the post-disaster enhanced vegetation index of the damage area.

[0028] Preferably, the damage degree index satisfies the following formula:

[0029] ΔNDVI = NDVI before - NDVI after

[0030] In the above formula, NDVI before is the pre-disaster normalized difference vegetation index, NDVI after is the post-disaster normalized difference vegetation index, and ΔNDVI is the damage degree index.

[0031] The enhanced damage degree index satisfies the following formula:

[0032] ΔEVI = EVI before - EVI after

[0033] In the above formula, EVI before is the pre-disaster enhanced vegetation index, EVI after is the post-disaster enhanced vegetation index, and ΔEVI is the enhanced damage degree index.

[0034] The normalized difference vegetation index satisfies the following formula:

[0035]

[0036] In the above formula, NDVI is the normalized difference vegetation index, NIR is the near-infrared band reflectance, and RED is the red light band reflectance.

[0037] The enhanced vegetation index satisfies the following formula:

[0038]

[0039] In the above formula, EVI is the enhanced vegetation index, G is the gain factor, NIR is the near-infrared band reflectance, RED is the red light band reflectance, C1 is the red light band correction coefficient, C2 is the blue light band correction coefficient, L is the atmospheric correction coefficient, and BLUE is the blue light band reflectance.

[0040] Preferably, the satellite remote sensing image based on the disaster area is combined with the distribution characteristics of the tree disaster situation to determine a disaster clipping image, including:

[0041] The satellite remote sensing image of the disaster area is analyzed by a remote sensing image analysis technology to obtain the distribution characteristics of the tree disaster situation;

[0042] Based on the distribution characteristics of the tree disaster situation, a target area is selected and clipped in the satellite remote sensing image to form a disaster clipping image.

[0043] Preferably, the post-disaster tree damage detection result is combined with the power grid geographic data for spatial superposition analysis to obtain a risk identification result of the tree damage range, including:

[0044] Based on the post-disaster tree damage detection result, the damaged tree area is determined;

[0045] The damaged tree area is combined with the power grid geographic data for spatial superposition analysis to determine the tree damage area, the tree lodging density, and the buffer high-risk area set of the tree affecting the power transmission line;

[0046] The tree damage area, the tree lodging density, and the buffer high-risk area set of the tree affecting the power transmission line are taken as the risk identification result of the tree damage range.

[0047] In a second aspect, the present application also provides a tree damage range intelligent identification system based on satellite remote sensing images, including:

[0048] A clipping module is configured to determine a disaster clipping image based on a satellite remote sensing image of a disaster area in combination with the distribution characteristics of the tree disaster situation;

[0049] An initial screening image determination module is configured to fuse the different spectral band data with the disaster clipping image by using a multi-spectral band fusion technology to obtain a fusion image; based on the fusion image in combination with multiple indexes, a color mapping technology and a clustering algorithm are used for joint comparative analysis to obtain a post-disaster damage area initial screening image;

[0050] A damage detection result determination module is configured to input the post-disaster damage area initial screening image into a pre-constructed refinement segmentation model to obtain a post-disaster tree damage detection result; wherein the refinement segmentation model includes an encoder part using a visual architecture based on a shift window mechanism;

[0051] A risk identification result determination module is configured to combine the post-disaster tree damage detection result with the power grid geographic data for spatial superposition analysis to obtain a risk identification result of the tree damage range.

[0052] Preferably, the damage detection result determination module includes:

[0053] The damage area determination sub-module is configured to determine a damage area in the image based on the post-disaster damage area preliminary screening image.

[0054] The index determination sub-module is configured to determine a damage degree index and an enhanced damage degree index according to the damage degree of the damage area.

[0055] The damage area labeling sub-module is configured to label a damage area in the post-disaster damage area preliminary screening image based on the damage degree index and the enhanced damage degree index.

[0056] The damage detection result determination sub-module is configured to input the labeled damage area into the refined segmentation model to obtain a post-disaster tree damage detection result.

[0057] Preferably, the system further comprises a refined segmentation model construction module configured to:

[0058] replace the convolutional layer of the encoder part of the deep learning segmentation model with a visual architecture based on a shift window mechanism, and add an attention mechanism module to the decoder part of the deep learning segmentation model to form an improved deep learning segmentation model;

[0059] combine a cross-entropy loss function and a semantic segmentation loss function to obtain a combined loss function of the improved deep learning segmentation model;

[0060] train the improved deep learning segmentation model based on the dataset with damage area labeling and the combined loss function to obtain a refined segmentation model.

[0061] Preferably, the preliminary screening image determination module comprises:

[0062] The luminance channel value determination sub-module is configured to process the spectrum in the fusion image by using a color mapping technique based on the fusion image combined with multiple indexes to obtain luminance channel values of different regions.

[0063] The post-disaster damage area preliminary screening image determination sub-module is configured to analyze the luminance channel values of different regions by using a clustering algorithm to obtain a post-disaster damage area preliminary screening image.

[0064] Preferably, the luminance channel value determination sub-module is specifically configured to:

[0065] convert the RGB color space of the spectrum in the fusion image to Lab color space by using a color mapping technique based on the fusion image combined with multiple indexes to form a Lab color space.

[0066] For the Lab color space, the lightness channel calculation formula is used to calculate the lightness channel values of different regions.

[0067] Preferably, the index determination submodule is specifically configured to:

[0068] Determine the damage degree index based on the pre-disaster normalized difference vegetation index and the post-disaster normalized difference vegetation index of the damage area.

[0069] Determine the enhanced damage degree index based on the pre-disaster enhanced vegetation index and the post-disaster enhanced vegetation index of the damage area.

[0070] Preferably, the damage degree index satisfies the following formula:

[0071] ΔNDVI = NDVI before - NDVI after

[0072] In the above formula, NDVI before is the pre-disaster normalized difference vegetation index, NDVI after is the post-disaster normalized difference vegetation index, and ΔNDVI is the damage degree index.

[0073] The enhanced damage degree index satisfies the following formula:

[0074] ΔEVI = EVI before - EVI after

[0075] In the above formula, EVI before is the pre-disaster enhanced vegetation index, EVI after is the post-disaster enhanced vegetation index, and ΔEVI is the enhanced damage degree index.

[0076] The normalized difference vegetation index satisfies the following formula:

[0077]

[0078] In the above formula, NDVI is the normalized difference vegetation index, NIR is the near-infrared band reflectance, and RED is the red band reflectance.

[0079] The enhanced vegetation index satisfies the following formula:

[0080]

[0081] In the above formula, EVI is the enhanced vegetation index, G is the gain factor, NIR is the near-infrared band reflectance, RED is the red band reflectance, C1 is the red band correction coefficient, C2 is the blue band correction coefficient, L is the atmospheric correction coefficient, and BLUE is the blue band reflectance.

[0082] Preferably, the cutting module is specifically used to:

[0083] Analyzing satellite remote sensing images of the disaster-stricken area using remote sensing image analysis technology to obtain distribution characteristics of tree damage;

[0084] Based on the distribution characteristics of the tree damage, a target area is selected in the satellite remote sensing image and cropped to form a disaster-damaged cropped image.

[0085] Preferably, the risk identification result determination module is specifically used to:

[0086] Determining the damaged tree area based on the post-disaster tree damage detection result;

[0087] Performing spatial overlay analysis on the damaged tree area and the power grid geographic data to determine the tree damaged area, tree fall density, and a set of high-risk buffer areas where trees affect transmission lines;

[0088] The tree damaged area, the tree falling density and the buffer high-risk area where the trees affect the power transmission lines are collected as a risk identification result of the tree damaged range.

[0089] In a third aspect, the present invention also proposes an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus;

[0090] The memory is used to store one or more programs;

[0091] When the one or more programs are executed by the at least one processor, the method for intelligently identifying the damaged area of ​​trees based on satellite remote sensing images is implemented.

[0092] In a fourth aspect, the present invention application further proposes a readable storage medium having an execution program stored thereon. When the execution program is executed, the method for intelligently identifying the extent of tree damage based on satellite remote sensing images is implemented.

[0093] Compared with the closest prior art, the present invention has the following beneficial effects:

[0094] The present invention provides a method and system for intelligently identifying the scope of tree damage based on satellite remote sensing images, comprising: determining a damaged cropped image based on satellite remote sensing images of the disaster-stricken area in combination with distribution characteristics of tree damage; fusing the different spectral band data with the damaged cropped image using multispectral band fusion technology to obtain a fused image; performing joint comparative analysis based on the fused image in combination with multiple indicators using color mapping technology and a clustering algorithm to obtain a preliminary screening image of the post-disaster damaged area; inputting the preliminary screening image of the post-disaster damaged area into a pre-constructed refined segmentation model to obtain a post-disaster tree damage detection result; performing spatial overlay analysis on the post-disaster tree damage detection result and power grid geographic data to obtain a risk identification result of the tree damage scope; wherein the refined segmentation model includes an encoder part that uses a visual architecture based on a shift window mechanism. By fusing the affected cropped images and processing them using color mapping technology and clustering algorithms, the images are used as input to the refined segmentation model to obtain the post-disaster tree damage detection results, and then the risk identification results of the tree damage range are obtained. This method can be used for tree damage monitoring on a large scale. After a disaster occurs, the damage to trees around the transmission lines can be quickly obtained, which can shorten the line inspection time and significantly improve efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 A flowchart of a method for intelligently identifying tree damage range based on satellite remote sensing images provided by the present invention;

[0096] Figure 2 The specific structure of the U-Net (U-Network, U-shaped network) deep learning model in the method for intelligent identification of tree damage range based on satellite remote sensing images provided by the present invention;

[0097] Figure 3 This is an architectural diagram of an intelligent tree damage range identification system based on satellite remote sensing images provided by the present invention;

[0098] Figure 4 This is a schematic diagram of the operation of an electronic device provided in the present invention application. DETAILED DESCRIPTION

[0099] The specific implementation methods of the present invention will be further described in detail below with reference to the accompanying drawings.

[0100] Example 1:

[0101] like Figure 1 As shown, the present invention proposes a method for intelligently identifying the damaged area of ​​trees based on satellite remote sensing images, comprising the following steps:

[0102] Step 1: based on satellite remote sensing images of the disaster area, combined with the distribution characteristics of tree disaster situation, determine the disaster clipping image;

[0103] Step 2: using multispectral band fusion technology to fuse the different spectral band data with the disaster clipping image, get the fusion image; based on the fusion image combined with multiple indexes, using color mapping technology and clustering algorithm for joint comparative analysis, get the post-disaster damage area preliminary screening image;

[0104] Step 3: input the post-disaster damage area preliminary screening image into the pre-constructed refinement segmentation model, get the post-disaster tree damage detection result;

[0105] Step 4: spatial overlay analysis of the post-disaster tree damage detection result and power grid geographic data, get the risk identification result of tree damage range; wherein the refinement segmentation model includes an encoder part using a visual architecture based on a shift window mechanism.

[0106] In step 1, before determining the disaster clipping image based on the satellite remote sensing images of the disaster area combined with the distribution characteristics of tree disaster situation, it includes:

[0107] Radiometric correction or / and geometric correction is performed on the satellite remote sensing image to obtain a corrected image;

[0108] The corrected image is processed using a histogram equalization enhancement method to obtain a new pixel image;

[0109] The new pixel image is processed using a Gaussian filter method or a median filter method to remove noise in the new pixel image.

[0110] Radiometric correction and geometric correction are performed on the satellite remote sensing image to form a corrected image to ensure the uniformity between different images and reduce the influence caused by light changes or differences in sensors used by the satellite to collect. Then, the histogram equalization enhancement method is used to improve the distinguishability of the corrected image, making the tree area more clear and distinguishable, as shown in the following formula. At the same time, the Gaussian filter method or the median filter method is used to remove noise in the new pixel image to avoid the influence of artifacts on subsequent segmentation accuracy.

[0111] The above-mentioned pixels of the new pixel image satisfy the following formula:

[0112]

[0113] In the above formula, s g is the new pixel value corresponding to the gth gray level in the new pixel image, k is the total number of gray levels, p r (r g ) is the probability of the gth gray level.

[0114] Further, in step 1 above, the satellite remote sensing image of the disaster area is combined with the distribution characteristics of tree disaster situation to determine the disaster clipping image, including the following steps:

[0115] Step 1.1: Analyze the satellite remote sensing image of the disaster area by remote sensing image analysis technology to obtain the distribution characteristics of tree disaster situation.

[0116] Step 1.2: Based on the distribution characteristics of tree disaster situation, select the target area in the satellite remote sensing image and clip it to form a disaster clipping image.

[0117] In step 1.1 above, high-resolution satellite remote sensing images are used to collect images of strong convective disaster trees in a wide range of disaster areas, ensuring the comprehensiveness and timeliness of disaster information. By analyzing the satellite remote sensing image through remote sensing image analysis technology, the distribution characteristics of tree disaster situation can be obtained, providing scientific support for subsequent disaster loss assessment and emergency repair.

[0118] In step 1.2 above, based on the distribution characteristics of tree disaster situation, the target area is selected and the satellite remote sensing image is clipped to a specific range to form a clipping image, reducing the amount of data and improving the calculation efficiency. For clipping images that require super large area, grid cutting is performed to ensure the consistency of the input size of the subsequent refined segmentation model, while avoiding the omission of small area targets. In addition, geographic information data can be combined to accurately match the forest area that needs to be identified, ensuring the accuracy of the processing target.

[0119] In step 2 above, to improve the consistency of the data, after the multi-spectral band fusion technology is fused, the pixel values of the fused image can be normalized to the range of 0 to 1 to form a normalized fused image, reducing the impact of light and sensor parameter differences. In addition, atmospheric correction algorithm is used to further adjust the spectral reflectance of the normalized fused image, making the data have better physical meaning and providing more stable input for the subsequent refined segmentation model. The pixel value of the normalized fused image satisfies the following formula:

[0120]

[0121] In the above formula, x is the pixel value of the fused image, x' is the pixel value of the normalized fused image, min(x) is the minimum value of the pixel in the fused image, and max(x) is the maximum value of the pixel in the fused image.

[0122] Through a series of data preprocessing and image enhancement operations, the quality of the satellite remote sensing image is improved, and the influence of light changes, noise interference and geometric distortion on the subsequent segmentation results is reduced.

[0123] Further, in the step 2, the post-disaster damage area preliminary screening image is obtained by combining multiple indexes based on the fusion image, using color mapping technology and clustering algorithm for joint comparative analysis, including the following steps:

[0124] Step 2.1: Based on the fusion image combining multiple indexes, the luminance channel values of different regions are obtained by using color mapping technology to process the spectrum in the fusion image.

[0125] Step 2.2: The clustering algorithm is used to analyze the luminance channel values of different regions to obtain the post-disaster damage area preliminary screening image.

[0126] Further, in the step 2.1, the luminance channel values of different regions are obtained by using color mapping technology to process the spectrum in the fusion image based on the fusion image combining multiple indexes, including:

[0127] Step 2.1.1: Based on the fusion image combining multiple indexes, the RGB (red, green and blue) color space of the spectrum in the fusion image is converted to Lab (Lab color space) color space using color mapping technology to form Lab color space.

[0128] Step 2.1.2: The luminance channel values of different regions are obtained by using the luminance channel calculation formula to calculate the Lab color space.

[0129] After the fusion image is standardized and normalized, the normalized fusion image is converted from RGB color space to Lab color space, mainly to reduce the influence of light changes on color features in the fusion image.

[0130] The luminance channel value satisfies the following formula:

[0131] L = 116f(Y / Y n )-16, a = 500(f(X / X n )-f(Y / Y n )),

[0132] b = 200(f(Y / Y n )-f(Z / Z n ))

[0133] In the above formula, L is the luminance channel value, a is the color channel of Lab color space, and b is the color channel different from a; X, Y, and Z are three components of Lab color space, and X n , Y n , and Z n are the coordinate values of the reference white point in Lab color space.

[0134] The above is that, through multi-spectral band fusion, the distribution characteristics of the tree disaster situation are enhanced, and stable and high-quality input data are provided for subsequent feature extraction and segmentation.

[0135] In the above step 2.1, the color mapping technology is used to process the spectrum in the fusion image, and then the color threshold method can be used to preliminarily extract the tree region, and the morphological processing (erosion, expansion) is combined to remove noise. In view of the color difference of different tree species, the brightness channel value of different regions is obtained;

[0136] In the above step 2.2, the clustering algorithm is used to further optimize the brightness channel value of different regions, and the discrimination between the background and the vegetation is improved. In order to accurately measure the brightness information of the image in the Lab color space, the calculation formula of the brightness channel value L is introduced, the brightness of different regions is quantitatively analyzed, and the separability of the tree region under uneven illumination is improved.

[0137] Since different spectral bands reflect different vegetation, multi-band data such as visible light, near-infrared and short-wave infrared can be used for fusion to enhance the contrast of the tree region and improve the recognizability of the damaged region. The spectral, color and texture features are used to identify the health status of the tree, and through the preliminary segmentation of the damage region, the accurate identification of the damaged range of the tree is realized. Through the damage classification based on the spectral and texture features, and the color mapping technology, the regions with different damage degrees are visualized, which provides a scientific basis for disaster assessment and recovery decision-making, and also prepares high-quality label data for the training of subsequent refined segmentation model, which plays an auxiliary role.

[0138] Further, in the above step 3, the post-disaster damage region preliminary screening image is input into the pre-constructed refined segmentation model to obtain the post-disaster tree damage detection result, including the following steps:

[0139] Step 3.1: determining the damage region in the image based on the post-disaster damage region preliminary screening image;

[0140] Step 3.2: determining the damage degree index and the enhanced damage degree index according to the damage degree of the damage region;

[0141] Step 3.3: labeling the damage region in the post-disaster damage region preliminary screening image based on the damage degree index and the enhanced damage degree index;

[0142] Step 3.4: inputting the labeled damage region into the refined segmentation model to obtain the post-disaster tree damage detection result.

[0143] In the step 3.1, the spectral information and the texture feature of the post-disaster damage area preliminary screening image are combined, a feature selection method is used to extract the saliency feature of the healthy area and the damaged area of the tree, and a damage area is obtained. The feature selection method can reduce the redundant information and improve the recognition accuracy of the damage area. The texture feature satisfies the following formula:

[0144]

[0145] In the above formula, Energy is the energy of the texture feature, P(i,j) 2 is the value of the (i,j) 2 term in the gray level co-occurrence matrix in the post-disaster damage area preliminary screening image, and i and j are two-dimensional coordinate points in the post-disaster damage area preliminary screening image.

[0146] Then, the spectral feature difference between the pre-disaster and post-disaster remote sensing images is calculated, the damage degree of the tree is quantified, and the damage area is classified according to the set threshold. NDVI (Normalized Difference Vegetation Index, damage degree index) and EVI (Enhanced Vegetation Index, enhanced vegetation index) are introduced as key indexes for vegetation damage recognition. NDVI is more suitable for rapid response detection in areas with severe damage and severe vegetation destruction in the disaster area; EVI is more suitable for areas with dense forests, high humidity or significant atmospheric interference. Both of them can extract features in parallel during image segmentation and classification, and can be input into the refined segmentation model through feature-level fusion to improve the overall recognition accuracy and robustness. For example, a slight damage area may show a slight decrease in NDVI, while a severe damage area shows a significant decrease in vegetation index. In addition, supervised classification algorithms such as random forest or SVM can be used in combination with spectral features to achieve more detailed damage area classification.

[0147] Further, in the step 3.2, the damage degree index and the enhanced damage degree index are determined according to the damage degree of the damage area, including:

[0148] The damage degree index is determined based on the pre-disaster normalized difference vegetation index and the post-disaster normalized difference vegetation index of the damage area.

[0149] The enhanced damage degree index is determined based on the pre-disaster enhanced vegetation index and the post-disaster enhanced vegetation index of the damage area.

[0150] Further, the damage degree index satisfies the following formula:

[0151] ΔNDVI = NDVI before - NDVI after

[0152] In the above formula, NDVI before NDVI is a pre-disaster normalized difference vegetation index, NDVI after NDVI is a post-disaster normalized difference vegetation index, and ΔNDVI is a damage degree index.

[0153] The enhanced damage degree index satisfies the following formula:

[0154] ΔEVI = EVI before EVI after

[0155] In the above formula, EVI before EVI is a pre-disaster enhanced vegetation index, EVI after EVI is a post-disaster enhanced vegetation index, and ΔEVI is an enhanced damage degree index.

[0156] The normalized difference vegetation index satisfies the following formula:

[0157]

[0158] In the above formula, NDVI is a normalized difference vegetation index, NIR is a near-infrared band reflectance, and RED is a red band reflectance.

[0159] The enhanced vegetation index satisfies the following formula:

[0160]

[0161] In the above formula, EVI is an enhanced vegetation index, G is a gain factor, NIR is a near-infrared band reflectance, RED is a red band reflectance, C1 is a red band correction coefficient, C2 is a blue band correction coefficient, L is an atmospheric correction coefficient, and BLUE is a blue band reflectance.

[0162] The above NDVI, EVI, and other vegetation indexes are calculated to quantitatively evaluate the health of trees, and then the data dimension of the vegetation indexes can be reduced by principal components analysis (PCA), weighted average method, and the like, so that the subsequent refined segmentation model can more effectively extract tree damage features.

[0163] In step 3.3, the damage area can be divided into three types of damage segmentation results, i.e., light (yellow), moderate (orange) and severe (red), according to the damage degree, and color identification can be performed in the post-disaster damage area preliminary screening image to obtain the labeled damage area. In order to ensure the continuity of the segmentation result, morphological filtering can also be used to remove isolated small areas, and the damage area boundary can be optimized by region growing algorithm to make it more consistent with the actual situation. Finally, the damage area can be generated into a high-precision tree damage distribution map, and in the region growing process, the pixel with high confidence damage area is used as a seed point, and when the following similarity criteria are met, it is classified into the same region:

[0164] |I(x,y)-I(x0,y0)|<T

[0165] wherein I(x,y) is the current pixel gray value, I(x0,y0) is the seed pixel gray value, and T is the growth similarity threshold value for controlling the gray consistency standard of region expansion.

[0166] Further, in step 3.4, the pre-construction process of the refined segmentation model includes:

[0167] The convolutional layer of the encoder part of the deep learning segmentation model is replaced by a visual architecture based on a shift window mechanism, and an attention mechanism module is added to the decoder part of the deep learning segmentation model to form an improved deep learning segmentation model.

[0168] The cross-entropy loss function and the semantic segmentation loss function are combined to obtain a combined loss function of the improved deep learning segmentation model.

[0169] Based on the dataset with damage area annotation and the combined loss function, the improved deep learning segmentation model is trained to obtain the refined segmentation model.

[0170] As shown in the above, Figure 2 the refined segmentation model can use a U-Net deep learning model. Through the automatic feature extraction and pixel-level segmentation capability of the U-Net deep learning model, the refined segmentation model can realize fine detection of tree damage areas. Combined with the pre-trained U-Net deep learning model and the optimized network structure, the segmentation accuracy and robustness are improved, and through the evaluation of the U-Net deep learning model, the generalization ability of the U-Net deep learning model under different remote sensing data and environmental conditions is ensured, thereby providing an efficient and accurate tree damage range identification scheme. Figure 2In the formula, Input is the input of the U-Net deep learning model, usually a multi-channel remote sensing image, used for subsequent feature extraction and segmentation; Output is the final output of the U-Net deep learning model, usually a class probability or segmentation label for each pixel. Conv 3x3 (3x3 convolution): a standard 2D convolution layer, usually using a 3x3 convolution kernel to extract local features. Conv 1x1 (1x1 convolution) is a 1x1 convolution kernel used by the U-Net deep learning model to linearly transform the channel information of each pixel. Max Pooling 2x2 (2x2 max pooling) is the down-sampling operation of the U-Net deep learning model, usually using a 2x2 window to compress the spatial size of the feature map, reduce resolution, expand the receptive field, and extract higher-level abstract features. Up Conv 2x2 (2x2 up-sampling convolution) is the up-sampling operation of the U-Net deep learning model. Copy and Crop (copy and crop) is the feature map of the corresponding layer of the encoder, which is copied and cropped and then spliced to the current layer of the decoder.

[0171] The U-Net deep learning model is trained based on remote sensing data to reduce training time and improve the generalization ability of the U-Net deep learning model. In the initialization stage, transfer learning is used with public datasets to enable the U-Net deep learning model to have initial image segmentation capabilities, and further training is performed with self-owned data to adapt to different environmental conditions.

[0172] First, the network structure of the U-Net deep learning model is fine-tuned;

[0173] The U-Net is used as a segmentation network, and the encoder part of the U-Net deep learning model is optimized, replacing the original convolution layer with a Swin Transformer (visual architecture based on shift window mechanism) to improve feature extraction capability. In addition, an attention mechanism (Attention Gate) module is added to the decoder part to enhance the sensitivity of the U-Net deep learning model to small-scale damage areas, reduce misclassification, and improve the accuracy of boundary recognition. This process can be expressed as follows:

[0174]

[0175] In the above formula: Q is the Query matrix, specifically the query matrix, which represents the representation of information that the U-Net deep learning model currently needs to pay attention to; K is the Key matrix, specifically the key matrix, which represents the labels or keywords of all input information, used to match the query matrix and determine relevance; T is the transpose of the matrix; V is the Value matrix, specifically the value matrix, which represents the specific content of all input information, and is the part that is finally selected and combined for output. d is the vector dimension in the U-Net deep learning model, B is the position encoding bias in the U-Net deep learning model, and Attention is the output feature weight result of the U-Net deep learning model; Softmax represents the function in the U-Net deep learning model.

[0176] After that, the improved deep learning segmentation model is trained and optimized; the data set with damage area annotation is used to train the improved deep learning segmentation model, and the combination of cross entropy loss (Cross Entropy Loss) and semantic segmentation loss function (Dice Loss) is used to optimize the segmentation effect. During the training process, the Adam (Adaptive Moment Estimation, combined with momentum) optimizer and the learning rate scheduling strategy are used to accelerate the convergence and prevent overfitting, which is expressed as follows:

[0177]

[0178] In the above formula: P is the predicted segmentation area; G is the real label area, |P∩G| is the intersection area of prediction and reality, |P| is the total area of the predicted segmentation area, and |G| is the total area of the real label area.

[0179] The combined loss function is expressed as follows:

[0180]

[0181] In the above formula: α and β are weight coefficients, is the cross entropy loss, is the loss of the semantic segmentation loss function, is the final total loss of the combined loss function;

[0182] At the same time, during the training process, the training process of the improved deep learning segmentation model is evaluated as follows:

[0183] The cross-validation method is used to evaluate the performance of the model, and the Intersection over Union (IoU), semantic segmentation loss coefficient, precision, recall and other indicators are used to measure the segmentation quality. At the same time, the generalization ability of the model is tested under different light, season and resolution conditions of remote sensing images to ensure its stability in different environments.

[0184] wherein, IoU is the Intersection over Union, IoU>0.7 usually indicates that the segmentation effect of the trained improved deep learning segmentation model is very good; between 0.5-0.7 is in the acceptable range; if IoU<0.5, it means that the trained improved deep learning segmentation model has poor segmentation quality, which may need to be optimized or retrained.

[0185] The semantic segmentation loss coefficient reflects the degree of overlap between the predicted area and the true area. Semantic segmentation loss coefficient>0.85 indicates that the segmentation quality of the trained improved deep learning segmentation model is excellent, 0.75-0.85 is at a medium level, and if <0.75, the segmentation stability of the trained improved deep learning segmentation model is low.

[0186] If the precision is very high but the recall is very low, it means that the trained improved deep learning segmentation model is very conservative and only detects the most obvious tree lodging area, resulting in a large number of missed detections, which may miss the targets that threaten the power transmission facilities in actual application. If the recall is very high but the precision is low, it means that the trained improved deep learning segmentation model is too aggressive and will produce more false positives. In order to minimize missed detection, the cost of manual review is increased. Precision>0.80 is better, and precision<0.70 has more false positives; recall>0.85 is ideal, and recall<0.75 may have serious risk of missed detection.

[0187] Further, in the above step 4, the spatial overlay analysis of the post-disaster tree damage detection result and the power grid geographic data is performed to obtain the risk identification result of the tree damage range, which includes:

[0188] Step 4.1: determining the damaged tree area based on the post-disaster tree damage detection result;

[0189] Step 4.2: performing spatial overlay analysis on the damaged tree area and the power grid geographic data to determine the tree damage area, tree lodging density and buffer high-risk area set of trees affecting power transmission lines;

[0190] Step 4.3: taking the tree damage area, tree lodging density and buffer high-risk area set of trees affecting power transmission lines as the risk identification result of the tree damage range.

[0191] The classification and recognition of different levels of lodging, breaking, and slight damage can be achieved in step 4.

[0192] The disaster impact analysis of tree damage area and lodging density is to evaluate the potential threat of tree damage to power transmission lines. By calculating the tree damage area and lodging density, the impact of the disaster area on power facilities can be more accurately evaluated, and the specific steps are as follows:

[0193] (1) Tree damage area calculation:

[0194] Using the tree damage area data in the post-disaster tree damage detection results, combined with the pixel resolution information of the damaged tree area. By identifying the damaged tree area and combining multispectral and texture features, the tree damage area is quantified, and an accurate tree damage area map is generated. The actual damage area can be calculated as follows:

[0195] A = N damaged · R 2

[0196] In the above formula, A is the actual damage area, N damaged is the number of damaged pixels, and R is the image resolution of the damaged tree area (m / pixel);

[0197] (2) Tree lodging density calculation:

[0198] For the "lodging" phenomenon in the tree disaster type, combined with the texture direction, shadow projection, and geometric shape features in the image, the spatial distribution density of the lodging trees is extracted. The sliding window counting or object-based detection method is used to count the number of lodging trees per unit area, so as to quantify the lodging density and evaluate the direct threat of lodging trees to power transmission lines. The tree lodging density is calculated as follows:

[0199]

[0200] In the above formula, D is the tree lodging density, and N fallen is the number of lodging trees.

[0201] (3) Calculation of the buffer high-risk area set of trees affecting power transmission lines:

[0202] The tree influence power transmission line buffer high-risk area set is mainly used for potential threat assessment of the power transmission line, on the basis of the tree damage area, further combining the GIS (Geographic Information System) spatial data of the power transmission line, analyzing the geometric relationship between the tree barrier area and the power transmission line. Through the means of constructing a spatial buffer, calculating the overlapping area, and evaluating the minimum distance, the "high-risk contact area" is identified, the damaged grades and spatial positions are fused, the power grid risk influence atlas based on geographic distribution is constructed, and accurate support is provided for the power department to make repair strategies and resource scheduling.

[0203] R risk ={x∈A∣dist(x,L)<r}

[0204] In the above formula, R risk is the tree influence power transmission line buffer high-risk area set, x is any spatial point, L is the power transmission line path, and r is the risk buffer radius.

[0205] Meanwhile, the area calculation function can be used to calculate the overlap ratio of the tree damage area and the power transmission line buffer area, and the overlap ratio is included in the risk identification result of the tree damage range. The calculation process of the overlap ratio satisfies the following formula:

[0206]

[0207] In the above formula, Overlap_Ratio is the overlap ratio, R damaged is the tree damage area, r buffer is the power transmission line buffer area, and Area(·) is the area calculation function.

[0208] The method provided by the present application realizes accurate identification and loss assessment of tree toppling after strong convective weather or extreme natural disasters (such as typhoon, thunderstorm, gale, earthquake, etc.) by collecting and analyzing high-resolution satellite remote sensing images, using computer vision and machine learning technology. This method can quickly and automatically detect the tree toppling situation in a large area, optimize the disaster emergency response strategy, improve the efficiency of rescue and disaster relief, and provide scientific basis for post-disaster recovery and reconstruction. The main innovation points of the present application are:

[0209] (1) Fusion of U-Net deep learning model and visual architecture based on shift window mechanism for post-disaster tree damage detection:

[0210] This invention is the first to fuse the U-Net deep learning model with a visual architecture based on a shifting window mechanism for the precise identification of damaged features such as fallen trees in post-disaster satellite remote sensing images. The U-Net deep learning model can effectively extract local spatial features of images and is suitable for precise segmentation; while the visual architecture based on a shifting window mechanism has strong global modeling capabilities and can capture semantic association information in images on a larger scale. By embedding the visual architecture based on a shifting window mechanism into the encoder part of the U-Net deep learning model, the model's ability to recognize fallen trees in complex backgrounds is significantly improved, making it particularly suitable for target detection in cluttered post-disaster scenes. This structure takes into account both spatial resolution and semantic extraction depth, and has good application prospects in the field of remote sensing identification of forest disasters.

[0211] (2) Introducing multispectral NDVI / EVI joint change analysis for preliminary screening of damaged areas after disasters: The present invention introduces multi-temporal remote sensing vegetation index change analysis as a pre-processing module for tree fall identification. Through joint comparative analysis of multiple indicators such as NDVI and EVI, it can effectively determine whether there are significant changes in tree growth conditions before and after the disaster. This method not only improves the accuracy of detecting tree fall areas, but also significantly reduces the false detection rate and computational overhead of the model. The spatial distribution characteristics of the index change rate map can be used to quickly locate and roughly screen potential damaged areas, providing accurate candidate areas for subsequent high-precision image segmentation, effectively improving the overall recognition efficiency and system practicality.

[0212] (3) Innovative introduction of tree damage classification mechanism and spatial location correlation analysis

[0213] Based on the traditional image recognition of tree damage levels, the present invention further constructs a tree damage classification mechanism, combining indicators such as color, shape, and edge clarity of damaged areas in remote sensing images to achieve classification and identification of different levels such as fallen, broken, and slightly damaged trees. On this basis, spatial position association analysis is further introduced. By extracting the geographic coordinate information of fallen trees and performing spatial overlay analysis with the power grid GIS data, a spatial overlapping relationship map between tree barrier locations and transmission lines is constructed. At the same time, combined with the spatial distance and orientation relationship between different tree barriers and power grid facilities, a risk level assessment mechanism based on the impact radius is established to achieve quantitative spatial assessment of the impact on power grid operation from image recognition. This model provides data support for rapid post-disaster decision-making and line maintenance priority sorting.

[0214] The objects of the present invention are:

[0215] (S1) Improve the efficiency of power line inspection: Traditional manual inspection and unmanned aerial vehicle inspection have problems such as long operation period, limited coverage, and great influence from terrain and weather. The application uses high-resolution satellite image data combined with a refined segmentation model to automatically identify tree damage in the power corridor in all-weather and large-scale, reducing manual input and improving inspection efficiency.

[0216] (S2) Improve disaster emergency response capability: When natural disasters such as typhoons, snowstorms, thunderstorms, and strong winds occur, the damage to trees along the power line will seriously threaten power safety. The application can use multi-temporal satellite images for change detection, quickly locate the damaged tree area, generate a damaged area assessment report, and provide scientific decision-making basis for the power department, improving the efficiency of post-disaster repair and emergency response.

[0217] (S3) Accurately identify damaged trees and reduce the risk of power transmission: Through multi-source remote sensing data fusion, the application can distinguish between normal trees and damaged trees, accurately identify fallen and broken trees, and automatically calculate the potential threat of damaged trees to the line, providing accurate data support for power grid dispatching, operation and maintenance, and repair.

[0218] (S4) Intelligent, automated, and highly adaptable: The application method is based on remote sensing images, deep learning, and change detection algorithms, and can be applied to different regions (plain, mountainous, forest) and different weather conditions (sunny, rainy, night), overcoming the limitations of traditional optical inspection by cloud cover, and enabling long-term automated monitoring, providing intelligent management means for power grid operation.

[0219] As a person skilled in the art would know, unmanned aerial vehicles are not suitable for large-scale applications due to weather, endurance, and operational difficulty; helicopter inspection is costly and limited by flight height and terrain, making it difficult to ensure full coverage. Therefore, in the context of power post-disaster emergency scenarios, there is an urgent need for an efficient, automated, and widely applicable tree damage monitoring method to improve the efficiency of power line inspection and disaster assessment.

[0220] Satellite remote sensing technology can provide large-scale, high-time-efficiency, and all-weather data acquisition capabilities, enabling rapid acquisition of tree damage around power lines after a disaster occurs. Combined with deep learning and computer vision, the application can automatically extract damaged tree areas in the power corridor, analyze damage extent, and generate a risk warning map to assist power grid maintenance personnel in quickly developing repair and obstacle removal plans, improving the power grid's ability to respond to disasters.

[0221] The method provided by the application realizes accurate identification of damaged trees along a power transmission channel by combining high-resolution satellite images and a refined segmentation model. Through the collected high-resolution satellite remote sensing images, the method can quickly evaluate the damage range of trees, identify the threat of fallen trees to power lines, and provide a scientific basis for post-disaster repair. The application can be widely applied to power grid inspection, post-disaster assessment, and transmission corridor management, improving the resilience and recovery capability of power grids in the face of natural disasters, and ensuring the safe and stable operation of power transmission lines.

[0222] Embodiment 2:

[0223] As shown in Figure 3 The application also provides an intelligent tree damage range identification system based on satellite remote sensing images, which comprises:

[0224] A clipping module is configured to determine a disaster-affected clipping image based on satellite remote sensing images of a disaster-affected area and in combination with the distribution characteristics of tree disaster conditions.

[0225] An initial screening image determination module is configured to fuse the different spectral band data with the disaster-affected clipping image by using a multispectral band fusion technology to obtain a fused image, and to obtain an initial screening image of a post-disaster damage area by using color mapping technology and clustering algorithms for joint comparative analysis based on the fused image in combination with multiple indexes.

[0226] A damage detection result determination module is configured to input the initial screening image of the post-disaster damage area into a pre-constructed refined segmentation model to obtain a post-disaster tree damage detection result, wherein the refined segmentation model includes an encoder part using a visual architecture based on a shift window mechanism.

[0227] A risk identification result determination module is configured to perform spatial overlay analysis on the post-disaster tree damage detection result and power grid geographic data to obtain a risk identification result of the tree damage range.

[0228] Further, the damage detection result determination module comprises:

[0229] A damage area determination sub-module is configured to determine a damage area in the image based on the initial screening image of the post-disaster damage area.

[0230] An index determination sub-module is configured to determine a damage degree index and an enhanced damage degree index for the damage area according to the damage degree.

[0231] A damage area labeling sub-module is configured to label a damage area in the initial screening image of the post-disaster damage area based on the damage degree index and the enhanced damage degree index.

[0232] The damage detection result determination submodule is configured to input the labeled damage region into the refined segmentation model to obtain a post-disaster tree damage detection result.

[0233] Further, the system further comprises a refined segmentation model construction module configured to:

[0234] The convolutional layer of the encoder part of the deep learning segmentation model is replaced by a visual architecture based on a shift window mechanism, and an attention mechanism module is added to the decoder part of the deep learning segmentation model to form an improved deep learning segmentation model.

[0235] The cross-entropy loss function and the semantic segmentation loss function are combined to obtain a combined loss function of the improved deep learning segmentation model.

[0236] Based on the dataset with damage region labeling and the combined loss function, the improved deep learning segmentation model is trained to obtain a refined segmentation model.

[0237] Further, the preliminary screening image determination module comprises:

[0238] The luminance channel value determination submodule is configured to process the spectrum in the fusion image based on the fusion image combined with multiple indicators using a color mapping technique to obtain luminance channel values of different regions.

[0239] The post-disaster damage region preliminary screening image determination submodule is configured to analyze the luminance channel values of different regions using a clustering algorithm to obtain a post-disaster damage region preliminary screening image.

[0240] Further, the luminance channel value determination submodule is specifically configured to:

[0241] Based on the fusion image combined with multiple indicators, the RGB color space of the spectrum in the fusion image is converted to Lab color space using a color mapping technique to form Lab color space.

[0242] The Lab color space is calculated using a luminance channel calculation formula to obtain luminance channel values of different regions.

[0243] Further, the index determination submodule is specifically configured to:

[0244] Based on the pre-disaster normalized difference vegetation index and the post-disaster normalized difference vegetation index of the damage region, a damage degree index is determined.

[0245] Based on the pre-disaster enhanced vegetation index and the post-disaster enhanced vegetation index of the damage region, an enhanced damage degree index is determined.

[0246] Further, the damage degree index satisfies the following formula:

[0247] ΔNDVI = NDVI before -NDVI after

[0248] In the above formula, NDVI before is a normalized difference vegetation index before disaster, NDVI after is a normalized difference vegetation index after disaster, and ΔNDVI is a damage degree index;

[0249] The enhanced damage degree index satisfies the following formula:

[0250] ΔEVI = EVI before -EVI after

[0251] In the above formula, EVI before is an enhanced vegetation index before disaster, EVI after is an enhanced vegetation index after disaster, and ΔEVI is an enhanced damage degree index;

[0252] The normalized difference vegetation index satisfies the following formula:

[0253]

[0254] In the above formula, NDVI is a normalized difference vegetation index, NIR is a near-infrared band reflectivity, and RED is a red light band reflectivity;

[0255] The enhanced vegetation index satisfies the following formula:

[0256]

[0257] In the above formula, EVI is an enhanced vegetation index, G is a gain factor, NIR is a near-infrared band reflectivity, RED is a red light band reflectivity, C1 is a red light band correction coefficient, C2 is a blue light band correction coefficient, L is an atmospheric correction coefficient, and BLUE is a blue light band reflectivity.

[0258] Further, the clipping module is specifically configured to:

[0259] analyze satellite remote sensing images of the disaster area through remote sensing image analysis technology to obtain tree disaster distribution characteristics;

[0260] select a target area in the satellite remote sensing images based on the tree disaster distribution characteristics and clip the target area to form a disaster clipping image.

[0261] Further, the risk identification result determination module is specifically configured to:

[0262] Based on the post-disaster tree damage detection result, the damaged tree area is determined;

[0263] The damaged tree area is spatially overlaid and analyzed with the power grid geographic data to determine a tree damage area, a tree down density, and a buffer high-risk area set of trees affecting power transmission lines;

[0264] The tree damage area, the tree down density, and the buffer high-risk area set of trees affecting power transmission lines are taken as the risk identification result of the tree damage range.

[0265] The present application faces the needs of power grid disaster early warning and post-disaster assessment, and proposes a system that integrates remote sensing image processing and refined segmentation model tree down identification and disaster impact analysis technology. After strong convective weather or extreme natural disasters, the system can quickly and accurately identify the tree damage range in a large area, quantify the tree damage area and down density, and assess the potential threat to power transmission lines. Through pre-processing, multi-spectral fusion, texture feature extraction, and refined segmentation of the U-Net deep learning model of high-resolution remote sensing images, high-precision identification of tree damage areas is achieved. Further spatial analysis of damaged areas and down density in combination with power line spatial information effectively improves disaster perception efficiency and power line risk assessment capability. The method has strong adaptability and high automation, and can be widely applied to power transmission channel inspection, disaster monitoring and post-disaster recovery assessment, effectively enhancing the intelligent operation level and resilience protection capability of the power grid system.

[0266] Embodiment 3:

[0267] As shown in Figure 4 The present application also provides an electronic device, which can be a computer device, a single-chip microcomputer device, a smart mobile device, etc. The electronic device in this embodiment can include a processor, a memory, a transceiver component, etc. The memory, the processor and the transceiver component are connected through a bus; the memory can be used to store an execution program, and the exemplary execution program can include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be called and / or modified when the instructions are executed.

[0268] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, and the like, which are a computing core and a control core of the terminal, and are suitable for implementing one or more instructions, and are suitable for loading and executing one or more instructions in the storage medium to implement a corresponding method process or a corresponding function, so as to implement the steps of the satellite remote sensing image tree damage range intelligent identification method in the above embodiment.

[0269] Embodiment 4:

[0270] Based on the same inventive concept, the application further provides a readable storage medium, specifically an electronic device readable storage medium (Memory). The electronic device readable storage medium is a memory device in the electronic device, and is used for storing programs and data. It can be understood that the storage medium herein can include a built-in storage medium in the electronic device, and of course can also include an expansion storage medium supported by the electronic device. The storage medium provides a storage space, and the storage space stores an operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more execution programs (including program codes). It should be noted that the storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, and the steps of the satellite remote sensing image tree damage range intelligent identification method in the above embodiment can be implemented.

[0271] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0272] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0273] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0274] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0275] Finally, it should be noted that the above examples are merely used to illustrate the technical solutions of the present application, but not to limit the protection scope thereof, and although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that after reading the present application, they can make various changes, modifications or equivalent replacements to the specific embodiments of the present application, but these changes, modifications or equivalent replacements all fall within the protection scope of the claims of the present application.

Claims

1. A method for intelligently identifying tree damage range based on satellite remote sensing images, characterized in that: include: Based on satellite remote sensing images of the disaster-stricken areas and the distribution characteristics of tree damage, the disaster-stricken cropped images are determined; The multispectral band fusion technology is used to fuse the different spectral band data with the disaster-affected cropped image to obtain a fused image; based on the fused image and combined with multiple indicators, a color mapping technology and a clustering algorithm are used to perform joint comparative analysis to obtain a preliminary screening image of the post-disaster damaged area; Inputting the pre-screened image of the post-disaster damaged area into a pre-built refinement segmentation model to obtain a post-disaster tree damage detection result; Performing spatial overlay analysis on the post-disaster tree damage detection results and power grid geographic data to obtain risk identification results of the tree damage range; The refined segmentation model includes an encoder part that adopts a visual architecture based on a shift window mechanism.

2. The method according to claim 1, characterized in that The post-disaster tree damage detection results are obtained by inputting the pre-constructed refined segmentation model into the pre-screened image of the post-disaster damaged area, including: Based on the preliminary screening image of the post-disaster damaged area, determining the damaged area in the image; For the damaged area, determining a damage degree index and an enhanced damage degree index according to the damage degree; Based on the damage degree index and the enhanced damage degree index, marking the damaged area in the primary screening image of the post-disaster damaged area; The marked damaged area is input into the refined segmentation model to obtain a post-disaster tree damage detection result.

3. The method according to claim 1 or 2, characterized in that The pre-construction process of the refined segmentation model includes: Replacing the convolutional layers of the encoder part of the deep learning segmentation model with a visual architecture based on a shifted window mechanism, and adding an attention mechanism module to the decoder part of the deep learning segmentation model to form an improved deep learning segmentation model; The cross entropy loss function and the semantic segmentation loss function are combined to obtain the combined loss function of the improved deep learning segmentation model; Based on the data set with damaged area annotations and the combined loss function, the improved deep learning segmentation model is trained to obtain a refined segmentation model.

4. The method according to claim 1, wherein The fusion image is combined with multiple indicators, and a color mapping technology and a clustering algorithm are used for joint comparative analysis to obtain a preliminary screening image of the post-disaster damaged area, including: Based on the fused image and a plurality of indicators, a color mapping technique is used to process the spectrum in the fused image to obtain brightness channel values ​​of different regions; The brightness channel values ​​of the different areas are analyzed using a clustering algorithm to obtain a preliminary screening image of the post-disaster damaged area.

5. The method according to claim 4, characterized in that The method of processing the spectrum in the fused image by using a color mapping technique based on the fused image in combination with multiple indicators to obtain brightness channel values ​​of different regions includes: Based on the fused image and a plurality of indicators, the RGB color space of the spectrum in the fused image is converted into the Lab color space by using a color mapping technology to form a Lab color space; The Lab color space is calculated using a lightness channel calculation formula to obtain lightness channel values ​​for different areas.

6. The method according to claim 2, characterized in that Determining a damage degree index and an enhanced damage degree index for the damaged area according to the damage degree includes: Determining a damage degree index based on a pre-disaster normalized difference vegetation index and a post-disaster normalized difference vegetation index of the damaged area; An enhanced damage degree index is determined based on the pre-disaster enhanced vegetation index and the post-disaster enhanced vegetation index of the damaged area.

7. The method according to claim 6, characterized in that The damage degree index satisfies the following formula: ΔNDVI=NDVI before -NDVI after In the above formula, NDVI before NDVI is the normalized difference vegetation index before the disaster after is the normalized difference vegetation index after the disaster, and ΔNDVI is the damage degree index; The enhanced damage degree index satisfies the following formula: ΔEVI=EVI before -HOUSE after In the above formula, EVI before EVI is the enhanced vegetation index before the disaster after is the enhanced vegetation index after the disaster, and ΔEVI is the enhanced damage index; The normalized difference vegetation index satisfies the following formula: In the above formula, NDVI is the normalized difference vegetation index, NIR is the near-infrared band reflectance, and RED is the red band reflectance; The enhanced vegetation index satisfies the following formula: In the above formula, EVI is the enhanced vegetation index, G is the gain factor, NIR is the near-infrared band reflectance, RED is the red band reflectance, C1 is the red band correction coefficient, C2 is the blue band correction coefficient, L is the atmospheric correction coefficient, and BLUE is the blue band reflectance.

8. The method according to claim 1, characterized in that The method of determining the disaster-stricken cropped image based on the satellite remote sensing image of the disaster-stricken area and the distribution characteristics of the tree damage includes: Analyzing satellite remote sensing images of the disaster-stricken area using remote sensing image analysis technology to obtain distribution characteristics of tree damage; Based on the distribution characteristics of the tree damage, a target area is selected in the satellite remote sensing image and cropped to form a disaster-damaged cropped image.

9. The method according to claim 1, characterized in that The post-disaster tree damage detection results are spatially overlaid with the power grid geographic data to obtain risk identification results of the tree damage range, including: Determining the damaged tree area based on the post-disaster tree damage detection result; Performing spatial overlay analysis on the damaged tree area and the power grid geographic data to determine the tree damaged area, tree fall density, and a set of high-risk buffer areas where trees affect transmission lines; The tree damaged area, the tree falling density and the buffer high-risk area where the trees affect the power transmission lines are collected as a risk identification result of the tree damaged range.

10. An intelligent identification system for tree damage range based on satellite remote sensing images, characterized in that: include: A cropping module is used to determine the cropped image of the disaster-stricken area based on satellite remote sensing images of the disaster-stricken area and the distribution characteristics of tree damage; A primary screening image determination module is configured to fuse the data of different spectral bands with the cropped image of the disaster-affected area using multispectral band fusion technology to obtain a fused image; based on the fused image and combined with multiple indicators, a color mapping technology and a clustering algorithm are used for joint comparative analysis to obtain a primary screening image of the damaged area after the disaster; a damage detection result determination module, configured to input the pre-screened image of the post-disaster damaged area into a pre-built refined segmentation model to obtain a post-disaster tree damage detection result; wherein the refined segmentation model includes an encoder portion using a visual architecture based on a shifting window mechanism; The risk identification result determination module is used to perform spatial overlay analysis on the post-disaster tree damage detection results and the power grid geographic data to obtain risk identification results of the tree damage range.

11. The system according to claim 10, wherein: The damage detection result determination module includes: A damage area determination submodule, configured to determine the damage area in the image based on the primary screening image of the post-disaster damage area; An index determination submodule, configured to determine a damage degree index and an enhanced damage degree index for the damaged area according to the damage degree; a damaged area marking submodule, configured to mark damaged areas in the primary screening image of the post-disaster damaged area based on the damage extent index and the enhanced damage extent index; The damage detection result determination submodule is used to input the marked damage area into the refined segmentation model to obtain the post-disaster tree damage detection result.

12. The system according to claim 10 or 11, characterized in that Also includes: Refinement segmentation model building module for: Replacing the convolutional layers of the encoder part of the deep learning segmentation model with a visual architecture based on a shifted window mechanism, and adding an attention mechanism module to the decoder part of the deep learning segmentation model to form an improved deep learning segmentation model; The cross entropy loss function and the semantic segmentation loss function are combined to obtain the combined loss function of the improved deep learning segmentation model; Based on the data set with damaged area annotations and the combined loss function, the improved deep learning segmentation model is trained to obtain a refined segmentation model.

13. The system according to claim 10, wherein: The primary screening image determination module includes: a luminance channel value determination submodule, configured to process the spectrum in the fused image using a color mapping technique based on the fused image in combination with multiple indicators to obtain luminance channel values ​​of different regions; The post-disaster damaged area preliminary screening image determination submodule is used to analyze the brightness channel values ​​of the different areas using a clustering algorithm to obtain a post-disaster damaged area preliminary screening image.

14. The system according to claim 13, wherein: The brightness channel value determination submodule is specifically used to: Based on the fused image and a plurality of indicators, the RGB color space of the spectrum in the fused image is converted into the Lab color space by using a color mapping technology to form a Lab color space; The Lab color space is calculated using a lightness channel calculation formula to obtain lightness channel values ​​for different areas.

15. The system according to claim 11, wherein: The indicator determination submodule is specifically used to: Determining a damage degree index based on a pre-disaster normalized difference vegetation index and a post-disaster normalized difference vegetation index of the damaged area; An enhanced damage degree index is determined based on the pre-disaster enhanced vegetation index and the post-disaster enhanced vegetation index of the damaged area.

16. The system according to claim 15, wherein: The damage degree index satisfies the following formula: ΔNDVI=NDVI before -NDVI after In the above formula, NDVI before NDVI is the normalized difference vegetation index before the disaster after is the normalized difference vegetation index after the disaster, and ΔNDVI is the damage degree index; The enhanced damage degree index satisfies the following formula: ΔEVI=EVI before -HOUSE after In the above formula, EVI before EVI is the enhanced vegetation index before the disaster after is the enhanced vegetation index after the disaster, and ΔEVI is the enhanced damage index; The normalized difference vegetation index satisfies the following formula: In the above formula, NDVI is the normalized difference vegetation index, NIR is the near-infrared band reflectance, and RED is the red band reflectance; The enhanced vegetation index satisfies the following formula: In the above formula, EVI is the enhanced vegetation index, G is the gain factor, NIR is the near-infrared band reflectance, RED is the red band reflectance, C1 is the red band correction coefficient, C2 is the blue band correction coefficient, L is the atmospheric correction coefficient, and BLUE is the blue band reflectance.

17. The system according to claim 10, wherein: The cutting module is specifically used for: Analyzing satellite remote sensing images of the disaster-stricken area using remote sensing image analysis technology to obtain distribution characteristics of tree damage; Based on the distribution characteristics of the tree damage, a target area is selected in the satellite remote sensing image and cropped to form a disaster-damaged cropped image.

18. The system according to claim 1, wherein: The risk identification result determination module is specifically used to: Determining the damaged tree area based on the post-disaster tree damage detection result; Performing spatial overlay analysis on the damaged tree area and the power grid geographic data to determine the tree damaged area, tree fall density, and a set of high-risk buffer areas where trees affect transmission lines; The tree damaged area, the tree falling density and the buffer high-risk area where the trees affect the power transmission lines are collected as a risk identification result of the tree damaged range.

19. An electronic device, characterized in that: include: at least one processor and memory; The memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the method for intelligently identifying the damaged area of ​​trees based on satellite remote sensing images as described in any one of claims 1 to 9 is implemented.

20. A readable storage medium, characterized in that An execution program is stored thereon, and when the execution program is executed, an intelligent identification method for tree damage range based on satellite remote sensing images as described in any one of claims 1 to 9 is implemented.

Images