A deep learning-based distribution network planning regional feature classification method and system

Abstract

The application relates to a kind of based on deep learning's regional feature classification method and system of distribution network planning, belong to artificial intelligence technical field.The method includes the following steps: using unmanned aerial vehicle to collect distribution network planning area orthophoto and label to form data set;The main network of DeeplabV3+ is replaced by MobileNetV3, the global pooling operation in ASPP module is replaced by NSCT, and EMA is introduced at the end of the encoder to obtain an improved model;The model is trained based on the data set, and then the orthophoto of the region to be classified is input into the trained model to obtain a pixel-by-pixel feature classification prediction result map, wherein the improved model includes an encoder and a decoder, each part uses a specific structure and operation, such as NSCT for feature decomposition and fusion, EMA for multi-scale attention processing, and the decoder performs upsampling, splicing and other operations to output the final result, solving the limitations of manual visual interpretation of orthophoto.

Description

Technical Field

[0001] This invention relates to a method and system for classifying land cover in power distribution network planning areas based on deep learning, belonging to the field of artificial intelligence technology. Background Technology

[0002] The demand for ground feature information identification in power distribution network planning and design projects is increasing. Currently, the main method relies on manual visual interpretation of UAV orthophotos. This traditional method has significant limitations: manual operation is inefficient, time-consuming, and labor-intensive, and is easily affected by subjective factors, resulting in insufficient accuracy of classification results. At the same time, existing technologies lack automated classification schemes, which cannot efficiently handle the complex and ever-changing ground feature scenes in the power distribution network planning area, and cannot meet the real-time and accuracy requirements of large-scale power distribution network planning.

[0003] For example, Chinese invention patent application CN114913436A discloses a method, device, electronic device, and medium for land cover classification based on a multi-scale attention mechanism, including the steps of: acquiring an input image; inputting the input image into a deep convolutional neural network to generate high-level semantic features and low-level semantic features; inputting the high-level semantic features into a dilated pyramid pooling module to generate a feature map; inputting the low-level semantic features into a second dual attention mechanism module to generate a feature map; and generating a predicted image based on the feature map and the feature map. However, its dilated pyramid pooling module relies on dilated convolution and global pooling operations, which, although capable of capturing multi-scale context, can lead to image edge and contour features being easily smoothed or lost during the decomposition process. At the same time, its attention mechanism only generates weights by stacking convolutional layers, lacking dynamic interaction across channels, which limits the robustness of the model in complex backgrounds.

[0004] In summary, there is an urgent need for a land cover classification method that can automate the processing of ground cover information; reduce the number of model parameters and accelerate model training while maintaining accuracy in segmentation tasks; maximize the preservation of edge and contour features of images, enhance the representation ability of image spatial structure, and improve the segmentation effect of the model; and capture key semantic information under different receptive fields to further improve the segmentation effect. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method and system for classifying land cover in power distribution network planning areas based on deep learning.

[0006] The technical solution of the present invention is as follows:

[0007] On the one hand, this invention proposes a deep learning-based method for classifying land cover in power distribution network planning areas, comprising the following steps:

[0008] UAVs were used to collect orthophotos within the power grid planning area, and image annotation tools were used to annotate the land cover types in the orthophotos to form a dataset;

[0009] Using the DeeplabV3+ model as a prototype, the Xception backbone network was replaced with MobileNetV3, the global pooling operation in the hollow spatial pyramid pooling ASPP module was replaced with non-subsampled contourlet transform (NSCT), and an efficient multi-scale attention mechanism (EMA) was introduced at the end of the encoder to obtain an improved DeeplabV3+ model.

[0010] The improved DeeplabV3+ model is trained based on the dataset to obtain the trained improved DeeplabV3+ model.

[0011] Obtain orthophotos of the distribution network planning area where the land cover classification is to be performed, input the orthophotos into the trained improved DeeplabV3+ model, and obtain the pixel-by-pixel land cover classification prediction result map.

[0012] Preferably, the improved DeeplabV3+ model includes an encoder and a decoder, specifically:

[0013] The encoder includes MobileNetV3 and ASPP modules and an efficient multi-scale attention mechanism EMA. The ASPP module includes 5 parallel branches, 1 projective convolution operation unit, and the first 4 branches of the 5 parallel branches are the same as the first 4 branches in the ASPP module of the DeeplabV3+ model. The 5th branch includes a non-subsampled contourlet transform (NSCT) unit and a 1×1 convolution operation channel stacking unit.

[0014] The decoder includes a 1×1 convolution operation unit, a 4x upsampling unit, a channel stacking unit, and a 3×3 convolution operation unit.

[0015] Preferably, the MobileNetV3 includes 3×3 convolutional layers, multiple inverse residual bottleneck blocks, and 1×1 convolutional layers, wherein:

[0016] The 3×3 convolutional layer is used to perform a 3×3 convolution operation with a stride of 2 on the pixel values ​​of each channel, row, and column in the input image of MobileNetV3, and output the original feature map; and the original feature map is normalized and processed by the h-swish activation function to output the preliminary feature map.

[0017] The multi-layer inverse residual bottleneck block is used to perform a 1×1 convolution operation on the preliminary feature map to increase its dimensionality and output the increased dimensionality preliminary feature map; a depthwise separable convolution is performed on the increased dimensionality preliminary feature map, and a channel reweighting operation is performed through the SE channel attention mechanism; the increased dimensionality preliminary feature map after the channel reweighting operation is reduced by 1×1 convolution operation to obtain the stage feature map and the low-level feature map.

[0018] The 1×1 convolutional layer is used to perform projective convolution on the final stage feature map of the stage feature map to obtain a high-dimensional feature map.

[0019] Preferably, the NSCT unit performs a pyramid decomposition on the high-dimensional feature map output from MobileNetV3 using a non-downsampled pyramidal filter bank to obtain high-frequency and low-frequency subbands; decomposes the high-frequency subband into multiple directional subbands using a non-downsampled directional filter bank; fuses the multi-directional high-frequency and low-frequency subbands to obtain corresponding high-frequency and low-frequency feature maps; fuses the obtained high-frequency and low-frequency feature maps and performs a 1×1 convolution operation to reduce the dimensionality of the fused image, outputting the NSCT feature map; fuses the NSCT feature map with the feature maps output from the first four branches and performs a 1×1 convolution operation to reduce the dimensionality, obtaining a multi-scale context feature map.

[0020] Preferably, the efficient multi-scale attention mechanism (EMA) unit utilizes an efficient multi-scale attention mechanism to divide the multi-scale context feature map along the channel dimension. Each sub-feature is divided into several sub-features, and a parallel two-branch operation is performed on each sub-feature. The parallel two-branch operation includes a first branch operation and a second branch operation, wherein:

[0021] The first branch operation specifically involves encoding the channels of the sub-features using two one-dimensional global average pooling operations in the horizontal and vertical directions; performing a 1×1 convolution operation on the resulting one-dimensional pooling results along the horizontal and vertical directions to output linear response maps in the horizontal and vertical directions; activating the linear response maps in the horizontal and vertical directions using the sigmoid function; performing a clustering operation after activation to obtain a spatial attention map; and multiplying the spatial attention map with the sub-features channel by channel to reweight them, outputting the reweighted multi-scale context feature map. Each sub-feature is normalized and subjected to two-dimensional average pooling. Then, the sub-features after normalization and two-dimensional average pooling are subjected to Softmax and 3×3 convolution to obtain the normalized channel attention weight vector and the multi-scale context feature map. The output sub-feature map is obtained by matrix multiplication of the normalized channel attention weight vector and the output sub-feature map;

[0022] The second branch encodes the channels of the sub-features using two-dimensional average pooling. The result of the two-dimensional average pooling is then processed by Softmax to obtain a normalized channel attention weight vector. This normalized channel attention weight vector is combined with the normalized sub-features from the first branch to generate a second spatial attention weight map. The first and second spatial attention weight maps are then added together and input into the Sigmoid activation function to obtain the final spatial attention weight map. The spatial attention weight map is then multiplied pixel by pixel with the sub-features to perform spatial reweighting on the current sub-features, resulting in a weighted feature map of the current sub-feature after EMA processing.

[0023] Preferably, the decoder is used to upsample the weighted feature map by 4 times; the weighted feature map after 4 times upsampling is concatenated with the low-level feature map to obtain a fused feature map; the fused feature map is subjected to a 3×3 convolution operation and 4 times upsampling to obtain a final feature map; the final feature map is subjected to class mapping to output a final pixel-by-pixel class score map.

[0024] On the other hand, this invention also proposes a deep learning-based land cover classification system for power distribution network planning areas, comprising the following modules:

[0025] Data acquisition and preprocessing module: Use drones to collect orthophotos within the power distribution network planning area, and use image annotation tools to label the types of ground features in the orthophotos;

[0026] The ground cover classification prediction module is internally deployed with a trained improved DeeplabV3+ model. Orthophotos are input into the trained improved DeeplabV3+ model to obtain pixel-by-pixel ground cover classification prediction results. The improved DeeplabV3+ model is based on the DeeplabV3+ model, replacing the Xception backbone network with MobileNetV3, replacing the global pooling operation in the Hollow Spatial Pyramid Pooling (ASPP) module with Non-Subsampled Profilotransform (NSCT), and introducing an efficient multi-scale attention mechanism (EMA) at the end of the encoder.

[0027] In another aspect, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in any embodiment of the present invention.

[0028] In another aspect, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0029] The present invention has the following beneficial effects:

[0030] (1) This invention is a deep learning-based distribution network planning area land cover classification system. By replacing the backbone network of the DeeplabV3+ model with MobileNetV3, the feature extraction process is reconstructed using inverse residual bottleneck blocks. MobileNetV3 replaces standard convolution with depthwise separable convolution, decoupling the convolution operation into channel-wise spatial feature extraction and 1×1 cross-channel integration, greatly simplifying the computation graph. At the same time, combined with the "expansion-compression" strategy, after enhancing the feature expression capability in the dimensionality-up phase, the parameter scale is controlled by the dimensionality-down operation. This design significantly reduces the computational load of the model, shortens the training cycle, and successfully achieves real-time inference on the UAV mobile terminal. The distribution network planning system thereby achieves end-to-end automation from image acquisition to land cover analysis.

[0031] (2) This invention is a deep learning-based distribution network planning area land cover classification system. By replacing the global average pooling of the original ASPP module in the DeeplabV3+ model with non-downsampled contour wave transform, the high-frequency and low-frequency components of the image are decomposed using a tower filter, and then the high-frequency multi-directional decomposition is performed by a directional filter bank. The downsampling operation is avoided, the complete spatial resolution is preserved, and the geometric features of the gradient change region are accurately maintained. In the distribution network land cover classification, the improved DeeplabV3+ model has significantly improved the ability to describe complex edges such as the corners of high-voltage towers, the outlines of manhole covers, and the intersection of field ridges. The boundary continuity of the segmentation results is highly consistent with the real land cover morphology, solving the extra cost of manual edge correction.

[0032] (3) This invention is a deep learning-based distribution network planning area land cover classification system. By introducing an efficient multi-scale attention mechanism and a channel grouping parallel processing mechanism into the DeeplabV3+ model, local details and global context are extracted simultaneously. The horizontal and vertical pooling of 1×1 convolution branches capture long-range dependencies, the 3×3 convolution branches enhance local feature responses, and cross-channel weight fusion achieves adaptive feature enhancement. This mechanism improves the robustness of the identification of key equipment in the distribution network. Even if the equipment is obscured by vegetation or is at the edge of the image, the target subject can still be accurately segmented, overcoming the previous problem of missed detection caused by small target size or background interference. Attached Figure Description

[0033] Figure 1 This is a flowchart of the land feature classification method provided in Embodiment 1 of the present invention;

[0034] Figure 2 This is a schematic diagram of the improved DeeplabV3+ network structure provided in Embodiment 1 of the present invention;

[0035] Figure 3 This is a flowchart of the non-subsampled contour wave transform algorithm provided in Embodiment 1 of the present invention;

[0036] Figure 4 The flowchart is for the multi-scale attention mechanism provided in Embodiment 1 of the present invention. Detailed Implementation

[0037] 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.

[0038] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0039] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0040] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0041] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0042] Example 1:

[0043] See Figure 1 This embodiment provides a deep learning-based method for classifying land cover in power distribution network planning areas, including the following steps:

[0044] S100. Based on the power distribution network planning and design project, plan the flight route of the UAV within the corresponding survey area;

[0045] The orthophoto of the survey area is obtained by using UAV photogrammetry to collect image data and other relevant data from the UAV.

[0046] Use image annotation tools to label the types of land features in the image, such as buildings, roads, woodlands, gardens, paddy fields, terraced fields, etc., and record their location, shape and other information to form structured data;

[0047] In this embodiment, to improve the generalization ability of the model, image enhancement processing is performed on the labeled images, such as rotation, scaling, color adjustment, and noise addition, and the output is a dataset containing the enhanced images.

[0048] S200, Participation Figure 2 Based on the DeeplabV3+ model, the Xception backbone network was replaced with MobileNetV3, the global pooling operation in the Spatial Pyramid Pooling (ASPP) module was replaced with Non-Subsampled Contour Transform (NSCT), and an efficient multi-scale attention mechanism (EMA) was introduced at the encoder end to obtain an improved DeeplabV3+ model. The improved DeeplabV3+ model includes an encoder and a decoder, specifically:

[0049] The encoder includes MobileNetV3 and ASPP modules. The ASPP module includes 5 parallel branches, 1 projection convolution operation unit, and an efficient multi-scale attention mechanism EMA unit. The first 4 branches of the 5 parallel branches are the same as the first 4 branches in the ASPP module of the DeeplabV3+ model. The 5th branch includes a non-subsampled contourlet transform (NSCT) unit and a 1×1 convolution operation channel stacking unit.

[0050] The decoder includes a 1×1 convolution operation unit, a 4x upsampling unit, a channel stacking unit, and a 3×3 convolution operation unit.

[0051] S300. Train the improved DeeplabV3+ model based on the training and validation sets. The MobileNetV3 model performs 3×3 convolutions with a stride of 2 on the images in the training and validation sets, as expressed by the formula:

[0052] ;

[0053] In the formula, Indicates the channel index. Indicates row index, Indicates column index, Represents the original feature map. , These represent row and column indices, respectively. Represents the input image tensor. Represents the convolution kernel. This represents the number of channels in the input image tensor. This indicates the number of channels in the output feature map. Indicates the first The bias vector of each channel.

[0054] The obtained original feature map is standardized, as expressed by the formula:

[0055] ;

[0056] In the formula, This represents the result after standardizing the original feature map. Indicates the first Scaling parameters for each channel, Indicates the first Translation parameters for each channel Indicates the first The mean of all values ​​across each channel. Indicates the first The standard deviation of all values ​​across each channel It represents a fixed, small constant.

[0057] The standardized result of the original feature map is input into the h-swish activation function for processing, expressed by the formula:

[0058] ;

[0059] ;

[0060] In the formula, This represents the preliminary feature map. This indicates taking the minimum value. This indicates taking the maximum value. This represents a nonlinear activation function.

[0061] S301. Perform a 1×1 convolution on the initial feature map to increase its dimensionality, expressed by the formula:

[0062] ;

[0063] In the formula, Indicates the first The feature map after dimensionality increase from 1×1 convolution in stage 1. This represents the number of channels in the output feature map after dimensionality upscaling. Indicates the first Stage-based dimensional advancement corresponds to the first Bias vectors for each channel Indicates the first Phase connection The first channel and the first Multidimensional convolution kernels with multiple channels. Indicates the first Stage input feature map.

[0064] The feature map output from the 1×1 convolution, after dimensionality upscaling, is then subjected to depthwise separable convolution, expressed by the formula:

[0065] ;

[0066] ;

[0067] In the formula, Indicates the first The output feature map after the stage undergoes depthwise separable convolution. Indicates the size of the depthwise convolution kernel. The two-dimensional convolution kernel is represented. Indicates the fill size. Indicates the first Feature map after dimensionality increase from 1×1 convolution in stage 1;

[0068] It should be noted that when the step size is 2, the first... The height and width of the feature map output by the stage depthwise convolution are expressed by the following formula:

[0069] ;

[0070] ;

[0071] In the formula, Indicates the first The stage depth can be used to determine the height of the output feature map after separable convolution. Indicates the first The height of the output feature map from the stage depthwise convolution. Indicates the first The width of the output feature map of the stage depthwise convolution. Indicates the first The width of the output feature map of the stage depth convolution.

[0072] For the first The first stage of the depthwise convolution output feature map The first channel, the... Okay, number The column values ​​are processed using the SE channel attention mechanism, which specifically includes the following steps:

[0073] Global averaging, expressed by the formula:

[0074] ;

[0075] In the formula, Indicates the first Phase, number Spatial mean of each channel;

[0076] Two fully connected layers, expressed by the formula:

[0077] ;

[0078] In the formula, Indicates the first Phase 1 The weight vector of each channel, Indicates the first Stage-reduced fully connected matrix Indicates the first Stage-up dimension fully connected matrix, This represents the sigmoid activation function. This indicates that the maximum value is used for calculation. Represents the activation function of a linear unit. Indicates the first The bias vector of the first fully connected layer in the first stage. Indicates the first The bias vector of the second fully connected layer in stage 2.

[0079] For the first Phase 1 The channel weight vectors are reweighted using the following formula:

[0080] ;

[0081] In the formula, Indicates the first Phase 1 The first channel Okay, number Feature map of column values ​​after reweighting Indicates the first Phase 1 The first channel, the... Okay, number Characteristic graph of column values;

[0082] The reweighted feature map is subjected to 1×1 convolution for dimensionality reduction, expressed by the formula:

[0083] ;

[0084] In the formula, Indicates the first Phase 1 The first channel Okay, number Stage characteristic diagram of column values; Indicates the first Stage-wise dimensionality reduction bias vector Indicates the first Phase connection The first channel and the first dimensionality-reduced convolution kernels for each channel;

[0085] Residual connections are used to preserve gradient backpropagation paths, as expressed by the formula:

[0086] ;

[0087] In the formula, Indicates the first Phase 1 The first channel Okay, number Stage characteristic diagram of column values.

[0088] S302, based on the first Phase 1 The first channel Okay, number The phase characteristic map of column values, taking The stage characteristic diagram of the stage, denoted as Perform a 1×1 projective convolution on the feature map, expressed by the formula:

[0089] ;

[0090] In the formula, The channels in the high-dimensional feature map are... , No. Okay, number The value of the column, Indicates connection channel and channels The projected convolution weights, This represents the feature map channel of stage 11. , No. Okay, number The value of the column, Indicates channel The bias vector, Indicates transpose;

[0091] S303, see also Figure 3 First, the input high-dimensional feature map is decomposed into two regions, high frequency and low frequency, by a non-subsampled pyramidal filter bank (NSP), as expressed by the formula:

[0092] ;

[0093] ;

[0094] ;

[0095] In the formula, Indicates the first Low-frequency subbands of layer decomposition Indicates a low-pass decomposition kernel. Indicates the first Hole spacing of the layers Indicates the first The hole spacing of the layer is Hollow convolution, Indicates the first High-frequency subbands of layer decomposition This indicates a Qualcomm decomposition kernel. Indicates the first Single-channel slice of the layer, Indicates the first Single-channel slice of the layer;

[0096] The high-frequency subband is then decomposed into subbands in multiple directions using a non-subsampled directional filter bank (NSDFB), achieving multi-directional decomposition, as expressed by the formula:

[0097] ;

[0098] In the formula, Indicates the first Layer direction is High-frequency directional subband, Indicates the first Layer direction is Directional filter kernel, , Indicates the first Layer direction is ;

[0099] Then, all the high-frequency and low-frequency subbands are fused separately to obtain high-frequency and low-frequency images, expressed by the formula:

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] In the formula, Indicates the first A low-frequency feature map, Indicates the first A high-frequency feature map, Indicates traversal and accumulation All layer directions are High-frequency directional subband, Indicates channel stacking. Indicates the first Low-frequency feature maps are stacked into channels. Indicates the first Channel stacking of high-frequency feature maps. Represents low-frequency feature maps. Represents high-frequency feature maps;

[0105] S304. The obtained high-frequency and low-frequency feature maps are fused, and the fused image is reduced in dimensionality by 1×1 convolution, expressed by the formula:

[0106] ;

[0107] ;

[0108] In the formula, This represents the feature map resulting from the fusion of high-frequency and low-frequency feature maps. Represents NSCT feature maps, This indicates the fused feature map in the channel. No. Line number The value of the column, Indicates the input channel is Output channel is The first projective convolution weight vector, Indicates the output channel is The first bias vector; Indicates the channel Accumulate;

[0109] The NSCT feature map is fused with the feature maps from the first four branches. The fused image is then subjected to 1×1 convolutional dimensionality reduction, expressed by the formula:

[0110] ;

[0111] ;

[0112] In the formula, This represents the feature map obtained from the first four branches of ASPP. Represents multi-scale contextual feature maps. Indicates the input channel is Output channel is The second projected convolution weight vector, Indicates the output channel is The second bias vector. Indicates the input channel is No. Line number NSCT feature map of the column, Indicates the channel Accumulate.

[0113] S305. Utilizing an efficient multi-scale attention mechanism, multi-scale contextual feature maps are divided along the channel dimension. Each feature, expressed by the formula:

[0114] ;

[0115] In the formula, Representing the multi-scale context feature map Individual characteristics, Indicates feature index, This indicates the number of channels in each group. Indicates that the channel is Multi-scale contextual feature maps;

[0116] For each of the identified sub-features, a parallel two-branch operation is performed; the parallel two-branch operation includes a first branch and a second branch, wherein:

[0117] A. The first branch encodes the channels using two one-dimensional global average pooling methods in the horizontal and vertical directions, expressed by the following formula:

[0118] ;

[0119] ;

[0120] In the formula, Indicates column index, Indicates row index, Indicates the first in the horizontal direction One-dimensional pooling results for each feature Indicates the first [number]th [unit] along the vertical direction One-dimensional pooling results for each feature Indicates that the channel is No. Line number Multi-scale context feature map One characteristic, Indicates that the channel is No. Line number Multi-scale context feature map One characteristic, Indicates the maximum number of row indexes. Indicates the maximum number of column indexes;

[0121] The obtained one-dimensional pooling results along the horizontal and vertical directions are convolved by 1×1 to output a linear response map, expressed by the formula:

[0122] ;

[0123] ;

[0124] In the formula, Indicates that the channel is No. Liede The linear response plot of each feature in the horizontal direction. Indicates that the channel is No. Line number Linear response plots in the vertical direction of each feature. Indicates channel The vertical bias vector. Indicates channel The horizontal bias vector, Indicates channel No. The vertical convolution weights of the rows, Indicates channel No. Horizontal convolution weights of the row;

[0125] The sigmoid function is used to activate the linear response plots in the horizontal and vertical directions. After activation, clustering and reweighting are performed, as expressed by the formula:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] In the formula, This indicates that the channel after activation by the sigmoid function is No. Liede The linear response plot of each feature in the horizontal direction. This indicates that the channel after activation by the sigmoid function is No. Liede The linear response plot of each feature in the horizontal direction. Indicates that the channel is No. Line number Liede Spatial attention map of features, The channel after reweighting is No. Line number Multi-scale context feature map Individual characteristics;

[0131] The reweighted multi-scale context feature map The sub-features are grouped and normalized, expressed by the formula:

[0132] ;

[0133] ;

[0134] ;

[0135] In the formula, This represents the group mean. Indicates the within-group standard deviation. The normalized channel is represented as No. Line number Multi-scale context feature map Individual characteristics, Represents a minimal constant;

[0136] After normalization, two-dimensional average pooling is performed, expressed by the formula:

[0137] ;

[0138] In the formula, Represents the channels of the normalized multi-scale context feature map No. Individual characteristics;

[0139] The sub-features are subjected to Softmax operation and 3×3 convolution, expressed by the following formula:

[0140] ;

[0141] ;

[0142] In the formula, This represents the normalized channel attention weight vector. Indicates that the channel is No. Line number Multi-scale context feature map Each output sub-feature map Indicates connection to channel and channels No. Line number Liede Convolution kernels for individual features, Indicates channel No. Line number Liede Individual feature maps, Indicates channel No. Bias vectors of individual features;

[0143] The normalized channel attention weight vector and the output sub-feature map are multiplied by a matrix, expressed by the formula:

[0144] ;

[0145] In the formula, express The flattened form, This represents the first spatial attention weight map. This represents a matrix transformation function.

[0146] B. The second branch encodes the channels using two-dimensional average pooling, expressed by the formula:

[0147] ;

[0148] In the formula, Represents the global statistical vector of the channel;

[0149] The result of two-dimensional average pooling is then processed by Softmax, and expressed by the formula:

[0150] ;

[0151] In the formula, Normalized channel attention weight vector;

[0152] The normalized channel attention weight vector is combined with the normalized sub-features in the 1×1 convolution branch. Combined with the generation of a second spatial attention weight map, expressed by the formula:

[0153] ;

[0154] In the formula, This represents the second spatial attention weight map. express The flattened form.

[0155] S306. Add the first and second attention weight maps together and input them into the Sigmoid activation function to obtain the final spatial attention weight map, expressed by the formula:

[0156] ;

[0157] ;

[0158] In the formula, This represents the weight map resulting from the sum of the first and second attention weight maps. The first attention weight map represents the... Line number The value of the column, The second attention weight map represents the first... Line number The value of the column, This represents the final spatial attention weight map;

[0159] The grouping results are reweighted, expressed by the formula:

[0160] ;

[0161] In the formula, This represents the final spatial attention weight map after reweighting.

[0162] S307. The final reweighted spatial attention weight map is upsampled by a factor of 4 using the decoder, as expressed by the formula:

[0163] ;

[0164] In the formula, Indicates channel No. Line number The feature map is upsampled by 4 times. This represents the bilinear interpolation algorithm. This indicates the row index of the feature map after upsampling. This indicates the column index of the feature map after upsampling. Indicates the channel index;

[0165] The feature map, after being upsampled by 4 times, is concatenated with the low-level feature map output by MobileNetV3, as expressed by the formula:

[0166] ;

[0167] ;

[0168] In the formula, This represents the low-level feature map output by MobileNetV3. Represents the fused feature map;

[0169] After fusing the feature maps, a 3×3 convolution and a 4x upsampling are performed, expressed by the formula:

[0170] ;

[0171] ;

[0172] In the formula, Represents the channel after convolution refinement No. Line number Column fusion feature map, This indicates the row index of the feature map after upsampling. This indicates the column index of the feature map after upsampling. This represents the final feature map after further upsampling.

[0173] The final feature map is then subjected to class mapping to obtain the final pixel-by-pixel class score map; expressed by the formula:

[0174] ;

[0175] In the formula, This represents the final pixel-by-pixel category score map. Indicates connection channel and channels The convolution weight matrix, Represents the final pixel-by-pixel category score map aisle The bias vector;

[0176] The model training is complete, resulting in the improved DeeplabV3+ model.

[0177] S400. Obtain the orthophoto of the power grid planning area where the land cover classification is to be performed. Input the orthophoto into the trained improved DeeplabV3+ model to obtain the pixel-by-pixel land cover classification prediction result map.

[0178] Example 2:

[0179] This embodiment also provides a deep learning-based land cover classification system for power distribution network planning areas, including the following modules:

[0180] Data acquisition and preprocessing module: Use drones to collect orthophotos within the power distribution network planning area, and use image annotation tools to label the types of ground features in the orthophotos;

[0181] The ground cover classification prediction module is internally deployed with a trained improved DeeplabV3+ model. Orthophotos are input into the trained improved DeeplabV3+ model to obtain pixel-by-pixel ground cover classification prediction results. The improved DeeplabV3+ model is based on the DeeplabV3+ model, replacing the Xception backbone network with MobileNetV3, replacing the global pooling operation in the Hollow Spatial Pyramid Pooling (ASPP) module with Non-Subsampled Profilotransform (NSCT), and introducing an efficient multi-scale attention mechanism (EMA) at the end of the encoder.

[0182] Example 3:

[0183] This embodiment proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in any embodiment of the present invention.

[0184] Example 4:

[0185] This embodiment proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0186] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0187] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0188] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0189] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0190] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for classifying land cover in power distribution network planning areas based on deep learning, characterized in that, Includes the following steps: UAVs were used to collect orthophotos within the power grid planning area, and image annotation tools were used to annotate the land cover types in the orthophotos to form a dataset; Using the DeeplabV3+ model as a prototype, the Xception backbone network was replaced with MobileNetV3, the global pooling operation in the hollow spatial pyramid pooling ASPP module was replaced with non-subsampled contourlet transform (NSCT), and an efficient multi-scale attention mechanism (EMA) was introduced at the end of the encoder to obtain an improved DeeplabV3+ model. The improved DeeplabV3+ model is trained based on the dataset to obtain the trained improved DeeplabV3+ model. Obtain the orthophoto of the distribution network planning area where the land cover classification is to be performed, input the orthophoto into the trained improved DeeplabV3+ model, and obtain the pixel-by-pixel land cover classification prediction result map. The improved DeeplabV3+ model includes an encoder and a decoder, specifically: The encoder includes MobileNetV3 and ASPP modules and an efficient multi-scale attention mechanism EMA. The ASPP module includes 5 parallel branches and 1 projective convolution operation unit. The first 4 branches of the 5 parallel branches are the same as the first 4 branches in the ASPP module of the DeeplabV3+ model. The 5th branch includes a non-subsampled contourlet transform (NSCT) unit and a 1×1 convolution operation channel stacking unit. The decoder includes a 1×1 convolution operation unit, a 4x upsampling unit, a channel stacking unit, and a 3×3 convolution operation unit; The MobileNetV3 includes 3×3 convolutional layers, multiple inverted residual bottleneck blocks, and 1×1 convolutional layers, wherein: The 3×3 convolutional layer is used to perform a 3×3 convolution operation with a stride of 2 on the pixel values ​​of each channel, row, and column in the input image of MobileNetV3, and output the original feature map; and the original feature map is normalized and processed by the h-swish activation function to output the preliminary feature map. The multi-layer inverse residual bottleneck block is used to perform a 1×1 convolution operation on the preliminary feature map to increase its dimensionality and output the increased dimensionality preliminary feature map; a depthwise separable convolution is performed on the increased dimensionality preliminary feature map, and a channel reweighting operation is performed through the SE channel attention mechanism; the increased dimensionality preliminary feature map after the channel reweighting operation is reduced by 1×1 convolution operation to obtain the stage feature map and the low-level feature map. The 1×1 convolutional layer is used to perform projective convolution on the final stage feature map of the stage feature map to obtain a high-dimensional feature map; The NSCT unit performs a pyramid decomposition on the high-dimensional feature map output from MobileNetV3 using a non-downsampled pyramidal filter bank to obtain high-frequency and low-frequency subbands. It then decomposes the high-frequency subbands into multiple directional subbands using a non-downsampled directional filter bank. These multi-directional high-frequency and low-frequency subbands are fused to obtain corresponding high-frequency and low-frequency feature maps. The obtained high-frequency and low-frequency feature maps are then fused, and a 1×1 convolution operation is performed on the fused image to reduce its dimensionality, outputting the NSCT feature map. Finally, the NSCT feature map is fused with the feature maps output from the first four branches and subjected to a 1×1 convolution operation to reduce its dimensionality, resulting in a multi-scale context feature map.

2. The method for classifying land cover in power distribution network planning areas based on deep learning according to claim 1, characterized in that, The efficient multi-scale attention mechanism (EMA) divides the multi-scale context feature map along the channel dimension. Each sub-feature is divided into several sub-features, and a parallel two-branch operation is performed on each sub-feature. The parallel two-branch operation includes a first branch operation and a second branch operation, wherein: The first branch operation specifically involves encoding the channels of the sub-features using two one-dimensional global average pooling operations in the horizontal and vertical directions; performing a 1×1 convolution operation on the resulting one-dimensional pooling results along the horizontal and vertical directions to output linear response maps in the horizontal and vertical directions; activating the linear response maps in the horizontal and vertical directions using the sigmoid function; performing a clustering operation after activation to obtain a spatial attention map; and multiplying the spatial attention map with the sub-features channel by channel to reweight them, outputting the reweighted multi-scale context feature map. Each sub-feature is normalized and subjected to two-dimensional average pooling. Then, the sub-features after normalization and two-dimensional average pooling are subjected to Softmax and 3×3 convolution to obtain the normalized channel attention weight vector and the multi-scale context feature map. The output sub-feature map is obtained by matrix multiplication of the normalized channel attention weight vector and the output sub-feature map; The second branch encodes the channels of the sub-features using two-dimensional average pooling. The result of the two-dimensional average pooling is then processed by Softmax to obtain a normalized channel attention weight vector. This normalized channel attention weight vector is combined with the normalized sub-features from the first branch to generate a second spatial attention weight map. The first and second spatial attention weight maps are then added together and input into the Sigmoid activation function to obtain the final spatial attention weight map. The spatial attention weight map is then multiplied pixel by pixel with the sub-features to perform spatial reweighting on the current sub-features, resulting in a weighted feature map of the current sub-feature after EMA processing.

3. The method for classifying land cover in power distribution network planning areas based on deep learning according to claim 2, characterized in that, The decoder is used to upsample the weighted feature map by 4 times; the weighted feature map after 4 times upsampling is concatenated with the low-level feature map to obtain a fused feature map; the fused feature map is subjected to a 3×3 convolution operation and 4 times upsampling to obtain the final feature map; the final feature map is subjected to class mapping, and the final pixel-by-pixel class score map is output.

4. A deep learning-based land cover classification system for power distribution network planning areas, characterized in that, Includes the following modules: Data acquisition and preprocessing module: Use drones to collect orthophotos within the power distribution network planning area, and use image annotation tools to label the types of ground features in the orthophotos; The land cover classification prediction module uses the DeeplabV3+ model as a prototype. The Xception backbone network is replaced with MobileNetV3, the global pooling operation in the Spatial Pyramid Pooling (ASPP) module is replaced with Non-Subsampled Profilotransform (NSCT), and an efficient multi-scale attention mechanism (EMA) is introduced at the encoder end to obtain an improved DeeplabV3+ model. The improved DeeplabV3+ model is trained based on the dataset to obtain a fully trained model. Orthophotos of the distribution network planning area where land cover classification is to be performed are acquired, and these orthophotos are input into the trained improved DeeplabV3+ model to obtain pixel-by-pixel land cover classification prediction results. The improved DeeplabV3+ model includes an encoder and a decoder, specifically: The encoder includes MobileNetV3 and ASPP modules and an efficient multi-scale attention mechanism EMA. The ASPP module includes 5 parallel branches and 1 projective convolution operation unit. The first 4 branches of the 5 parallel branches are the same as the first 4 branches in the ASPP module of the DeeplabV3+ model. The 5th branch includes a non-subsampled contourlet transform (NSCT) unit and a 1×1 convolution operation channel stacking unit. The decoder includes a 1×1 convolution operation unit, a 4x upsampling unit, a channel stacking unit, and a 3×3 convolution operation unit; The MobileNetV3 includes 3×3 convolutional layers, multiple inverted residual bottleneck blocks, and 1×1 convolutional layers, wherein: The 3×3 convolutional layer is used to perform a 3×3 convolution operation with a stride of 2 on the pixel values ​​of each channel, row, and column in the input image of MobileNetV3, and output the original feature map; and the original feature map is normalized and processed by the h-swish activation function to output the preliminary feature map. The multi-layer inverse residual bottleneck block is used to perform a 1×1 convolution operation on the preliminary feature map to increase its dimensionality and output the increased dimensionality preliminary feature map; a depthwise separable convolution is performed on the increased dimensionality preliminary feature map, and a channel reweighting operation is performed through the SE channel attention mechanism; the increased dimensionality preliminary feature map after the channel reweighting operation is reduced by 1×1 convolution operation to obtain the stage feature map and the low-level feature map. The 1×1 convolutional layer is used to perform projective convolution on the final stage feature map of the stage feature map to obtain a high-dimensional feature map; The NSCT unit performs a pyramid decomposition on the high-dimensional feature map output from MobileNetV3 using a non-downsampled pyramidal filter bank to obtain high-frequency and low-frequency subbands. It then decomposes the high-frequency subbands into multiple directional subbands using a non-downsampled directional filter bank. These multi-directional high-frequency and low-frequency subbands are fused to obtain corresponding high-frequency and low-frequency feature maps. The obtained high-frequency and low-frequency feature maps are then fused, and a 1×1 convolution operation is performed on the fused image to reduce its dimensionality, outputting the NSCT feature map. Finally, the NSCT feature map is fused with the feature maps output from the first four branches and subjected to a 1×1 convolution operation to reduce its dimensionality, resulting in a multi-scale context feature map.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 3.

Images