Method for identifying rural three-life space land type based on deep learning

Abstract

The invention discloses a method for identifying rural three-life space land types based on deep learning. S1, downloading data from the Google Earth and converting the data; s2, importing the converted raster image into ArcGISPro, and exporting vector data of which the format is an shp format; s3, judging the processing format of the metadata according to the land use type, and exporting the training data after selecting the metadata processing format; s4, performing model training by using the training data exported in the S3; s5, identifying the land to obtain a land type raster image; s6, smoothing the raster image exported in the S5 by using a mode filtering tool, vectorizing the processed raster by using a raster surface conversion tool, and exporting regular data; and S7, obtaining vector data of rural ecology, production and living space. According to the invention, accurate identification of ecological, production and living spaces is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of land use type identification, and in particular to a method for identifying rural three-life space land use types based on deep learning. Background Art

[0002] The three life spaces are ecological, production and living spaces. The division of rural ecological, production and living spaces is based on the identification of land use types, and plays a fundamental role in current situation analysis and research in rural planning.

[0003] There are currently a variety of methods for identifying land use types, including remote sensing multispectral image recognition, supervised learning algorithm training model recognition, and drone high-resolution image recognition. These methods generally have problems such as low recognition accuracy and difficulty in data processing. In addition, the recognition scale is large, and the processed vector data is difficult to apply in village-level planning and research.

[0004] The publication number is CN116012334A, which is a domain knowledge constrained unsupervised change detection method for construction land. Its technical solution is to process Sentinel-2 remote sensing images and then use supervised classification methods to identify land use types. The defects are that the identification scale is large and the accuracy is insufficient, and the spatial resolution is 10 meters.

[0005] A method for functional classification of urban buildings in high-resolution remote sensing images with publication number CN107247938A uses high-resolution images and a CNN method to extract buildings. However, the method has the disadvantage of identifying fewer land use types and is limited to urban areas.

[0006] In summary, the defects of the existing technology are that the recognition scale is large, the accuracy is insufficient, the types of land use types recognized are small, and the recognition areas are mostly located in cities. Summary of the invention

[0007] In order to overcome the above technical problems, the purpose of the present invention is to provide a method for identifying the land use types of rural three-life spaces (ecological, production, and living space) based on deep learning. The method is based on the deep learning module of ArcGIS Pro, and through sampling, training deep learning models, and using models to identify rural space types, it realizes accurate identification of ecological, production, and living spaces.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A method for identifying rural three-life space land use types based on deep learning, comprising the following steps;

[0010] S1: Download data from Google Earth and convert it to obtain raster images;

[0011] S2: Import the converted raster image into ArcGIS Pro, and use the surface drawing tool to collect data for residential buildings, factories, greenhouses, farmland, roads (inside and outside the group), woodlands, grasslands and shrubs, rivers and lakes. Export the vector data in shp format;

[0012] S3: Determine the metadata processing format based on the land use type. The metadata processing format used is Classified Tiles. Select the metadata processing format, import the vector data of S2, and export the training data.

[0013] S4: Use the training data exported from S3 to train the model and export the training model;

[0014] S5: In the parameter page, input the grid to be identified, select the grid to be identified, select the training model corresponding to the previous step in the model definition, identify the land use, and obtain the land use type raster image;

[0015] S6: The raster image exported from S5 is smoothed using the majority filter tool, the processed raster is vectorized using the raster to polygon tool, the vectorized data is processed with the fusion tool to process the polygon features, and the regularized data is exported;

[0016] S7: Use the merge tool to merge the residential houses and roads (within the cluster) of S6 regular data into living space. Merge the farmland, factory buildings, greenhouses, and roads (outside the cluster) of S6 regular data into production space. Merge the woodlands, grasslands, shrubs, rivers, and lakes of S6 regular data into ecological space, export vector data, and obtain vector data of rural ecology, production, and living space.

[0017] In S1, the downloaded data is an 8-bit raster image with three bands and contains geographic spatial data. The raster image contains all land use types of ecological, production and living spaces and is in Tiff format. The export tool is used to convert the geographic coordinate system (the original coordinate system of the data) into a projection coordinate system. The converted projection coordinate system is the CGCS2000 GK Zone 18 projection coordinate system. In this coordinate system, the pixel sizes of the x-axis and y-axis are 0.56 meters and 0.55 meters, respectively.

[0018] The S2 is specifically:

[0019] The surface drawing tool of the deep learning module of ArcGIS Pro was used to collect data on residential houses, factories, greenhouses, farmlands, roads (inside and outside the cluster), woodlands, grasslands and shrubs, rivers and lakes; the collected data was exported as vector data in shp format.

[0020] The S3 is specifically:

[0021] Parameter page settings: select the raster image downloaded in S1 for input raster, select Tiff format for image format, select map space for reference system, and select classified tiles for metadata processing format (Classified Tiles, this format only outputs statistical data with detailed information about the class, such as class name, class value and output statistics, and is mainly used for pixel classification);

[0022] After the above parameters are set, select Run.

[0023] The S4 is specifically:

[0024] In the parameter page, enter the training data and select the raster image used for sampling. Select 20 for the maximum epoch. The model type is determined by the site and the algorithm used is:

[0025] Among them, the BDCNEdgeDetector algorithm is selected for residential houses, factories, greenhouses, woodlands, grasslands, shrubs, rivers and lakes; the U-Net algorithm is selected for farmland and greenhouses; and the MultiTaskRoadExtractor algorithm is selected for roads (highways and village roads) to train deep learning models.

[0026] The BDCNEdgeDetector (Bidirectional Cascade Network Edge Detector) is an edge detection algorithm based on deep learning; this algorithm can accurately identify the edges of residential buildings, factory buildings, greenhouses, woodlands, grasslands, shrubs, rivers, lakes and other land use types. The following is its detailed calculation formula and processing flow:

[0027] Step 1: Input Image Processing

[0028] The training data X exported from S3 is input into the VGG16-based network to extract features layer by layer. The network is divided into multiple incremental detection blocks (ID Blocks), each of which is responsible for edge detection at a specific scale;

[0029]

[0030] Parameter definition: X is the training data exported from S3; N s (X): feature map of the sth layer; w, h are the width and height of the feature map; c is the number of channels of the feature map; Y is the complete true edge map; Y s is the real edge map corresponding to the scale sss; S is the total number of scales;

[0031] Step 2: Scale Enhancement Module (SEM)

[0032] A Scale Enhancement Module (SEM) is inserted into each ID Block to expand the receptive field of each layer through multi-branch dilated convolution to capture multi-scale features;

[0033] Dilated convolution calculation formula:

[0034]

[0035] Parameter definition: x[i,j] is the value of the input feature map at position (i,j); r is the expansion rate, which controls the sampling step; w[m,n] is the weight of the convolution kernel at position (m,n); y i,j is the value of the output feature map at position (i, j); m, n are the size indexes of the convolution kernel; K is the number of branches with different expansion rates; r k =max(1,r0·k) is the dilation rate of the kth dilated convolution; r0 is the dilation rate reference factor;

[0036] Step 3: Bidirectional Cascading

[0037] A bidirectional cascade structure is used to transfer information between shallow and deep layers, focusing on detail edges from shallow to deep and overall boundaries from deep to shallow.

[0038] Shallow to Deep (S2D): Low-level features extract detail edges;

[0039]

[0040] Deep to Shallow (D2S): high-level features capture large-scale boundaries;

[0041] Parameter definition: Y s s2d ,Y s d2s is the layer-specific supervision signal from shallow to deep layers and from deep to shallow layers; P i s2d ,

[0042]

[0043] P i d2s is the edge prediction result from shallow to deep and deep to shallow; Y is the complete real edge map; i, s are the scale indexes of the current and other layers;

[0044] Step 4: Edge Blending

[0045] The edge prediction results of each layer are fused through 1×1 convolution to generate the final edge detection result;

[0046] Multi-scale edge fusion:

[0047]

[0048] Parameter definition: P is the final fusion edge prediction map; P s s2d ,P s d2s For the shallow-to-deep and deep-to-shallow edge predictions at the sth layer;

[0049] Step 5: Loss calculation and optimization

[0050] Use the class-balanced cross entropy loss to optimize the network, and calculate the loss of edge predictions at each layer and the loss of the final fusion result separately;

[0051] Total loss function:

[0052] L=w side ·L side +w fuse ·L fuse (P,Y)

[0053] Side loss:

[0054]

[0055] Class-balanced cross entropy:

[0056]

[0057] Parameter definition: L is the total loss; w side ,w fusew _ is the weight of the side loss and fusion loss; L side is the edge detection loss for each layer; L fuse is the fusion edge detection loss; P s s2d , P s d2s Y is the edge prediction of the sth layer; s s2d , Y s d2s is the layer-specific edge supervision; Y - is a set of non-edge pixels; Y + is the edge pixel set; α, β are the weights for balancing positive and negative samples; y ′ j ,y j is the predicted probability and true label of the j-th pixel.

[0058] The U-Net is an edge detection algorithm based on deep learning. The algorithm can accurately identify the edges of farmland and greenhouse land types. The following is its detailed calculation formula and processing flow:

[0059] Step 1: Shrink the Path

[0060] The contraction path is responsible for extracting contextual features of the image (training data derived from S3); it uses two consecutive 3x3 convolutions (without padding), each followed by a ReLU activation function; it then applies 2x2 max pooling (with a stride of 2) for downsampling, doubling the number of channels after each pooling;

[0061] convolution:

[0062]

[0063] Parameter definition: W (l) is the convolution kernel of the lth layer; * is the convolution operation; x (l-1) is the input feature map of the l-1th layer; b (l) is the bias term;

[0064] Max Pooling:

[0065]

[0066] Parameter definition: z i,j is the feature map after downsampling;

[0067] Step 2: Expand the path

[0068] The expansion path is responsible for gradually restoring the spatial resolution; the feature map is enlarged to the original resolution using 2x2 deconvolution (upsampling); the feature map of the corresponding layer of the contraction path is combined with the upsampled feature map through a splicing operation; after splicing, it is processed through two consecutive 3x3 convolutions and ReLU activations;

[0069] Deconvolution:

[0070]

[0071] Parameter definition: y i,j up is the feature map after upsampling; W up [m,n] is the deconvolution kernel;

[0072] Stitching:

[0073]

[0074] convolution:

[0075]

[0076] Parameter definition: y i,j contracty is the feature map from the contraction path;

[0077] Step 3: Final classification layer

[0078] In the last layer, 1x1 convolution is used to map the number of channels to the number of target categories; the category probability of each pixel is calculated by Softmax;

[0079] 1x1 Convolution:

[0080]

[0081] Parameter definition: p k (x) is the predicted probability that pixel x belongs to category k;

[0082] Softmax:

[0083]

[0084] Parameter definition: a k (x) is the activation value of category k; K is the total number of categories;

[0085] Step 4: Loss Function

[0086] Use pixel-wise cross entropy as the loss function, combined with weights to balance the training weights of small objects and background;

[0087] Cross Entropy Loss:

[0088]

[0089] Parameter definition: Ω is the pixel set of the image; l(x) is the true label of pixel x; p l(x) (x) is the predicted probability of pixel x; w (x) is the pixel weight;

[0090] Pixel weight:

[0091]

[0092] Parameter definition: w c (x) is the weight for balancing the class frequency; w0 is a fixed weight parameter; d1(x) and d2(x) are the distances from the pixel to the nearest and second nearest target boundaries, respectively; σ is a smoothing parameter.

[0093] The MultiTaskRoadExtractor framework regards the road extraction task as a multi-task learning problem. Through this algorithm, the edges of road land types can be accurately identified. The specific steps include: semantic segmentation task, boundary detection task, and centerline detection task;

[0094] This multi-task learning structure is expressed as:

[0095] L MTRE =λ1Lsegmentation +λ2L boundary +λ3L centerline

[0096] Parameter definition: L MTREL is the total loss function, L segmentation , L boundary , L centerline They represent the losses of segmentation, boundary and centerline detection respectively, λ1, λ2, λ3 are the weight coefficients of the tasks;

[0097] Step 1: Segmentation Loss

[0098] Semantic segmentation usually uses a pixel-level cross entropy loss function:

[0099]

[0100] Parameter definition: y(x) is the true label of pixel x, p(x) is the predicted probability, and Ω represents the pixel set;

[0101] Step 2: Boundary Detection Loss

[0102] Boundary detection is achieved through cross entropy loss based on edge maps or specific boundary loss to enhance the clarity of the boundary. The usual boundary detection loss function is:

[0103]

[0104] Parameter definition: y b (x) and p b (x) are the true boundary label and predicted boundary probability of pixel x respectively;

[0105] Step 3: Centerline Detection Loss

[0106] Centerline detection is used to ensure road connectivity, also calculated by cross entropy or other loss functions:

[0107]

[0108] Parameter definition: y c (x) and p c (x) are the true centerline labels and predicted probabilities, respectively.

[0109] The S6 raster data processing steps are:

[0110] Majority filter: Use the majority filter tool for smoothing, set the parameter page, select the raster data exported in S5 for input setting, select the number of adjacent pixels: four, and the replacement threshold is the majority;

[0111] After the above parameters are set, select Run;

[0112] Raster to Surface: Input raster is the raster data obtained in the previous step. After the above parameters are set, select Run to export surface vector data.

[0113] Fusion: For input features, select the vector data obtained in the previous step. After the above parameters are set, select Run to export the surface vector data.

[0114] Beneficial effects of the present invention:

[0115] Starting from how to accurately identify the types of ecological, production, and living space land from the rural scale, the present invention constructs a technical process for identifying the types of rural ecological, production, and living space land based on deep learning. Compared with the traditional ecological, production, and living space identification methods, this process can quickly and accurately identify the types of rural land, and then obtain the vector data of ecological, production, and living space after processing. The method is simple, intuitive, and easy to operate. It realizes the innovation of land type identification methods and can provide new ideas for the data collection and processing methods of ecology, production, and living space in current rural planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 It is a schematic diagram of the process of the present invention.

[0117] Figure 2 This is a table of land use types and algorithms matching in the present invention. DETAILED DESCRIPTION

[0118] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0119] See attached Figure 1 , the present invention provides the following technical solution: a method for identifying rural ecological, production and living space land types based on deep learning, comprising the following steps ( Figure 1 Identification process of ecological, production and living space land types):

[0120] S1: Data acquisition: download data from Google Earth;

[0121] The downloaded data is an 8-bit raster image with three bands and contains geospatial data. The raster image contains all land use types of ecological, production, and living space ( Figure 2), the format is Tiff; use the export tool to convert the geographic coordinate system into the projection coordinate system. The converted spatial coordinate system is the CGCS2000 GK Zone 18 projection coordinate. In this coordinate system, the pixel sizes of the x-axis and y-axis are 0.56 meters and 0.55 meters respectively;

[0122] S2: Sample collection: Import the raster image downloaded in S1 into ArcGIS Pro, and use the drawing surface tool to collect data on residential houses, factories, greenhouses, farmlands, roads (inside and outside the cluster), woodlands, grasslands and shrubs, rivers and lakes; the collection tool is the Add Surface tool in ArcGIS Pro, collect surface data of residential houses, factories, greenhouses, farmlands, roads (inside and outside the cluster), woodlands, grasslands and shrubs, rivers and lakes, and collect no less than 50 data. Export the collected data as vector data in shp format;

[0123] S3: Export training data for deep learning: Determine the metadata processing format based on the land use type. The metadata processing format used is Classified Tiles. Export the training data after selecting the metadata processing format.

[0124] Parameter page settings: select the raster image downloaded in S1 for input raster, select Tiff format for image format, select map space for reference system, and select classified tiles for metadata processing format (Classified Tiles, this format only outputs statistical data with detailed information about the class, such as class name, class value and output statistics, mainly used for pixel classification);

[0125] After the above parameters are set, select Run.

[0126] S4: Training deep learning model: The algorithm used is determined according to the land use type. The algorithm used for residential houses, factories, woodlands, grasslands and shrubs, rivers and lakes is the BDCNEdgeDetector algorithm; the algorithm used for farmland and greenhouses is the U-Net algorithm, and the algorithm used for roads (within and outside the group) is the MultiTaskRoadExtractor algorithm ( Figure 2 ), according to the above judgment criteria, the specific parameter settings are as follows:

[0127] When the land use type is residential, factory, woodland, grassland, shrubs, or river: set the parameter page, select the classification slice metadata exported in the previous step as the input training data, select 20 as the maximum period, and select BDCN edge detector (pixel classification) as the model type; set the environment page, and select GPU as the processor type.

[0128] When the land use type is farmland or greenhouse: On the parameter page, select the classification slice metadata exported in the previous step as the input training data, select 20 as the maximum period, and select U-Net (pixel classification) as the model type; On the environment page, select GPU as the processor type.

[0129] When the land use type is road (inside or outside the cluster): set the parameter page, select the classification slice metadata exported in the previous step as the input training data, select 20 as the maximum period, and select Multi-task Road Extractor (Pixel Classification) as the model type; set the environment page, and select GPU as the processor type.

[0130] After the above parameters are set, select Run.

[0131] The S4 is specifically:

[0132] In the parameter page, enter the training data and select the raster image used for sampling. Select 20 for the maximum epoch. The model type is determined by the site and the algorithm used is:

[0133] Among them, residential houses, factories, greenhouses, woodlands, grasslands, shrubs, rivers and lakes are selected

[0134] The BDCNEdgeDetector algorithm is used, the U-Net algorithm is used for farmland and greenhouses, and the MultiTaskRoadExtractor algorithm is used for roads (highways and village roads) to train the deep learning model.

[0135] The BDCNEdgeDetector (Bidirectional Cascade Network Edge Detector) is an edge detection algorithm based on deep learning; this algorithm can accurately identify the edges of residential buildings, factory buildings, greenhouses, woodlands, grasslands, shrubs, rivers, lakes and other land use types. The following is its detailed calculation formula and processing flow:

[0136] Step 1: Input Image Processing

[0137] The training data X exported from S3 is input into the VGG16-based network to extract features layer by layer. The network is divided into multiple incremental detection blocks (ID Blocks), each of which is responsible for edge detection at a specific scale;

[0138]

[0139] Parameter definition: X is the training data exported from S3; N s (X): feature map of the sth layer; w, h are the width and height of the feature map; c is the number of channels of the feature map; Y is the complete true edge map; Y s is the real edge map corresponding to the scale sss; S is the total number of scales;

[0140] Step 2: Scale Enhancement Module (SEM)

[0141] A Scale Enhancement Module (SEM) is inserted into each ID Block to expand the receptive field of each layer through multi-branch dilated convolution to capture multi-scale features;

[0142] Dilated convolution calculation formula:

[0143]

[0144] Parameter definition: x[i,j] is the value of the input feature map at position (i,j); r is the expansion rate, which controls the sampling step; w[m,n] is the weight of the convolution kernel at position (m,n); y i,j is the value of the output feature map at position (i, j); m, n are the size indexes of the convolution kernel; K is the number of branches with different expansion rates; r k =max(1,r0·k) is the dilation rate of the kth dilated convolution; r0 is the dilation rate reference factor;

[0145] Step 3: Bidirectional Cascading

[0146] A bidirectional cascade structure is used to transfer information between shallow and deep layers, focusing on detail edges from shallow to deep and overall boundaries from deep to shallow.

[0147] Shallow to Deep (S2D): Low-level features extract detail edges;

[0148]

[0149] Deep to Shallow (D2S): high-level features capture large-scale boundaries;

[0150]

[0151] Parameter definition: Y s s2d ,Y s d2s is the layer-specific supervision signal from shallow to deep layers and from deep to shallow layers; P i s2d ,P i d2s is the edge prediction result from shallow to deep and deep to shallow; Y is the complete real edge map; i, s are the scale indexes of the current and other layers;

[0152] Step 4: Edge Blending

[0153] The edge prediction results of each layer are fused through 1×1 convolution to generate the final edge detection result;

[0154] Multi-scale edge fusion:

[0155]

[0156] Parameter definition: P is the final fusion edge prediction map; P s s2d ,P s d2s For the shallow-to-deep and deep-to-shallow edge predictions at the sth layer;

[0157] Step 5: Loss calculation and optimization

[0158] Use the class-balanced cross entropy loss to optimize the network, and calculate the loss of edge predictions at each layer and the loss of the final fusion result separately;

[0159] Total loss function:

[0160] L=w side ·L side +w fuse ·L fuse (P,Y)

[0161] Side loss:

[0162]

[0163] Class-balanced cross entropy:

[0164]

[0165] Parameter definition: L is the total loss; w side ,w fusew _ is the weight of the side loss and fusion loss; L side is the edge detection loss for each layer; L fuse is the fusion edge detection loss; P s s2d , P s d2s Y is the edge prediction of the sth layer; s s2d , Y s d2s is the layer-specific edge supervision; Y - is a set of non-edge pixels; Y + is the edge pixel set; α, β are the weights for balancing positive and negative samples; y ′ j ,y j is the predicted probability and true label of the j-th pixel.

[0166] The U-Net is an edge detection algorithm based on deep learning. The algorithm can accurately identify the edges of farmland and greenhouse land types. The following is its detailed calculation formula and processing flow:

[0167] Step 1: Shrink the Path

[0168] The contraction path is responsible for extracting contextual features of the image (training data derived from S3); it uses two consecutive 3x3 convolutions (without padding), each followed by a ReLU activation function; it then applies 2x2 max pooling (with a stride of 2) for downsampling, doubling the number of channels after each pooling;

[0169] convolution:

[0170]

[0171] Parameter definition: W (l) is the convolution kernel of the lth layer; * is the convolution operation; x (l-1) is the input feature map of the l-1th layer; b (l) is the bias term;

[0172] Max Pooling:

[0173]

[0174] Parameter definition: z i,j is the feature map after downsampling;

[0175] Step 2: Expand the path

[0176] The expansion path is responsible for gradually restoring the spatial resolution; the feature map is enlarged to the original resolution using 2x2 deconvolution (upsampling); the feature map of the corresponding layer of the contraction path is combined with the upsampled feature map through a splicing operation; after splicing, it is processed through two consecutive 3x3 convolutions and ReLU activations;

[0177] Deconvolution:

[0178]

[0179] Parameter definition: y i,j up is the feature map after upsampling; W up [m,n] is the deconvolution kernel;

[0180] Stitching:

[0181]

[0182] convolution:

[0183]

[0184] Parameter definition: y i,j contracty is the feature map from the contraction path;

[0185] Step 3: Final classification layer

[0186] In the last layer, 1x1 convolution is used to map the number of channels to the number of target categories; the category probability of each pixel is calculated by Softmax;

[0187] 1x1 Convolution:

[0188] Parameter definition: p k (x) is the predicted probability that pixel x belongs to category k;

[0189]

[0190] Softmax:

[0191] Parameter definition: a k (x) is the activation value of category k; K is the total number of categories;

[0192] Step 4: Loss Function

[0193] Use pixel-wise cross entropy as the loss function, combined with weights to balance the training weights of small objects and background;

[0194] Cross entropy loss: Parameter definition: Ω is the pixel set of the image; l(x) is the true label of pixel x;

[0195]

[0196] p l(x) (x) is the predicted probability of pixel x; w (x) is the pixel weight;

[0197] Pixel weight:

[0198]

[0199] Parameter definition: w c (x) is the weight for balancing the class frequency; w0 is a fixed weight parameter; d1(x) and d2(x) are the distances from the pixel to the nearest and second nearest target boundaries, respectively; σ is a smoothing parameter.

[0200] The MultiTaskRoadExtractor framework regards the road extraction task as a multi-task learning problem. Through this algorithm, the edges of road land types can be accurately identified. The specific steps include: semantic segmentation task, boundary detection task, and centerline detection task;

[0201] This multi-task learning structure is expressed as:

[0202] L MTRE =λ1L segmentation +λ2L boundary +λ3Lcenterline

[0203] Parameter definition: L MTREL is the total loss function, L segmentation , L boundary , L centerline They represent the losses of segmentation, boundary and centerline detection respectively, λ1, λ2, λ3 are the weight coefficients of the tasks;

[0204] Step 1: Segmentation Loss

[0205] Semantic segmentation usually uses a pixel-level cross entropy loss function:

[0206]

[0207] Parameter definition: y(x) is the true label of pixel x, p(x) is the predicted probability, and Ω represents the pixel set;

[0208] Step 2: Boundary Detection Loss

[0209] Boundary detection is achieved through cross entropy loss based on edge maps or specific boundary loss to enhance the clarity of the boundary. The usual boundary detection loss function is:

[0210]

[0211] Parameter definition: y b (x) and p b (x) are the true boundary label and predicted boundary probability of pixel x respectively;

[0212] Step 3: Centerline Detection Loss

[0213] Centerline detection is used to ensure road connectivity, also calculated by cross entropy or other loss functions:

[0214]

[0216] Parameter definition: y c (x) and p c (x) are the true centerline labels and predicted probabilities, respectively.

[0217] S5: Use deep learning to classify pixels. The specific parameters are as follows: Parameter page settings, select the raster image downloaded in S1 for input raster selection, and select the exported model corresponding to the previous step for model definition; Environment page settings, processor type GPU.

[0218] After the above parameter settings are completed, select Run to generate land use type raster data.

[0219] S6 raster data processing steps:

[0220] Majority filter: Use the majority filter tool for smoothing. In the parameter page, select the raster data exported from S5 in the input setting, select four for the number of adjacent pixels, and set the replacement threshold to the majority.

[0221] After the above parameters are set, select Run;

[0222] Raster to Surface: Input raster is the raster data obtained in the previous step. After the above parameters are set, select Run to export surface vector data.

[0223] Fusion: For input features, select the vector data obtained in the previous step. After the above parameters are set, select Run to export the surface vector data.

[0224] S7 was merged into three living spaces according to the land use type;

[0225] According to the land use type, the types of ecological, production and living space land use are judged, and the rules for merging and classification are as follows:

[0226] a) Living space: residential buildings, roads (within the cluster);

[0227] b) Production space: farmland, factory buildings, greenhouses, roads (outside the cluster);

[0228] c) Ecological space: woodlands, grasslands, shrubs, rivers and lakes;

[0229] Each vector data is merged separately using the merge tool. The merged vector data are the ecological, production and living land types.

[0230] In summary, the present invention starts from how to accurately identify the types of ecological, production, and living space land from the rural scale, and constructs a technical process based on deep learning to identify the types of rural ecological, production, and living space land. Compared with traditional ecological, production, and living space identification methods, this process can quickly and accurately identify rural land types, and then obtain vector data of ecological, production, and living space after processing. This method is simple and intuitive, and easy to operate, with the characteristics of "clear logic" and "convenient process". To a certain extent, it has achieved method innovation in the field of land type identification, and can provide new ideas for the current ecological, production, and living space data collection and processing methods in rural planning.

Claims

1. A method for identifying rural three-life space land use types based on deep learning, characterized in that: The steps include: S1: Download data from Google Earth and convert it to obtain raster images; S2: Import the converted raster image into ArcGIS Pro and export it as vector data in shp format; S3: Determine the metadata processing format according to the land use type. The metadata processing format used is classified slicing. After selecting the metadata processing format, export the training data. S4: Use the training data exported from S3 to train the model; S5: In the parameter page, input the grid to be identified, select the grid to be identified, select the training model corresponding to the previous step in the model definition, identify the land use, and obtain the land use type raster image; S6: The raster image exported from S5 is smoothed using the majority filter tool, the processed raster is vectorized using the raster to polygon tool, the vectorized data is processed with the fusion tool to process the polygon features, and the regularized data is exported; S7: Use the merge tool to merge the residential houses and roads in the S6 regularized data into living space. Merge the farmland, factory buildings, greenhouses, and roads in the S6 regularized data into production space. Merge the woodlands, grasslands, shrubs, rivers, and lakes in the S6 regularized data into ecological space. Export vector data to obtain vector data of rural ecology, production, and living space.

2. According to claim 1, a method for identifying rural three-life space land use types based on deep learning is characterized by: In S1, the downloaded data is an 8-bit raster image with three bands and containing geographic spatial data. The raster image contains all land use types of ecological, production and living spaces and is in Tiff format. The geographic coordinate system is converted into a projection coordinate system using an export tool. The converted projection coordinate system is the CGCS2000 GK Zone 18 projection coordinate system. In this coordinate system, the pixel sizes of the x-axis and y-axis are 0.56 meters and 0.55 meters, respectively.

3. According to the method of identifying rural three-life space land use types based on deep learning in claim 1, it is characterized by: The S2 is specifically: The surface drawing tool of the deep learning module of ArcGIS Pro was used to collect data on residential houses, factories, greenhouses, farmlands, roads (inside and outside the cluster), woodlands, grasslands and shrubs, rivers and lakes; the collected data was exported as vector data in shp format.

4. According to the method of identifying rural three-life space land use types based on deep learning in claim 1, it is characterized by: The S3 is specifically: Parameter page settings: select the raster image downloaded in S1 for input raster, select Tiff format for image format, select map space for reference system, and select classified slices for metadata processing format; After the above parameters are set, select Run.

5. According to the method of identifying rural three-life space land use types based on deep learning in claim 1, it is characterized by: The S4 is specifically: In the parameter page, enter the training data and select the raster image used for sampling. Select 20 for the maximum epoch. The model type is determined by the land use and the algorithm used is: Among them, the BDCNEdgeDetector algorithm is selected for residential houses, factories, greenhouses, woodlands, grasslands, shrubs, rivers and lakes; the U-Net algorithm is selected for farmland and greenhouses; and the MultiTaskRoadExtractor algorithm is selected for roads (highways and village roads) to train deep learning models.

6. The method for identifying rural three-life space land use types based on deep learning according to claim 5 is characterized in that: The BDCNEdgeDetector identifies the edges of land use types such as residential buildings, factories, greenhouses, woodlands, grasslands and shrubs, rivers and lakes, specifically: Step 1: Input Image Processing The training data X exported from S3 is input into the VGG16-based network to extract features layer by layer. The network is divided into multiple incremental detection blocks, each of which is responsible for edge detection at a specific scale. Parameter definition: X is the training data exported from S3; N s (X): feature map of the sth layer; w, h are the width and height of the feature map; c is the number of channels of the feature map; Y is the complete true edge map; Y s is the real edge map corresponding to the scale sss; S is the total number of scales; Step 2: Scale Enhancement Module Insert a scale enhancement module into each ID Block and expand the receptive field of each layer through multi-branch dilated convolution to capture multi-scale features; Dilated convolution calculation formula: Parameter definition: x[i,j] is the value of the input feature map at position (i,j); r is the expansion rate, which controls the sampling step; w[m,n] is the weight of the convolution kernel at position (m,n); y i,j is the value of the output feature map at position (i, j); m, n are the size indexes of the convolution kernel; K is the number of branches with different expansion rates; r k =max(1,r0·k) is the dilation rate of the kth dilated convolution; r0 is the dilation rate reference factor; Step 3: Bidirectional Cascading A bidirectional cascade structure is used to transfer information between shallow and deep layers, focusing on detail edges from shallow to deep and overall boundaries from deep to shallow. From shallow to deep layers: low-level features extract detailed edges; Deep to shallow: high-level features capture large-scale boundaries; Parameter definition: Y s s2d ,Y s d2s is the layer-specific supervision signal from shallow to deep layers and from deep to shallow layers; P i s2d , P i d2s is the edge prediction result from shallow to deep and deep to shallow; Y is the complete real edge map; i, s are the scale indexes of the current and other layers; Step 4: Edge Blending The edge prediction results of each layer are fused through 1×1 convolution to generate the final edge detection result; Multi-scale edge fusion: Parameter definition: P is the final fusion edge prediction map; P s s2d ,P s d2s For the shallow-to-deep and deep-to-shallow edge predictions at the sth layer; Step 5: Loss calculation and optimization Use the class-balanced cross entropy loss to optimize the network, and calculate the loss of edge predictions at each layer and the loss of the final fusion result separately; Total loss function: L=w side ·L side +w fuse ·L fuse (P,Y) Side loss: Class-balanced cross entropy: Parameter definition: L is the total loss; w side ,w fusew _ is the weight of the side loss and fusion loss; L side is the edge detection loss for each layer; L fuse is the fusion edge detection loss; P s s2d , P s d2s is the edge prediction of the sth layer; Y s s2d , Y s d2s is the layer-specific edge supervision; Y - is a set of non-edge pixels; Y + is the edge pixel set; α, β are the weights for balancing positive and negative samples; y ′ j ,y j is the predicted probability and true label of the j-th pixel.

7. The method for identifying rural three-life space land use types based on deep learning according to claim 5 is characterized in that: The U-Net identifies the edges of farmland and greenhouse land types, specifically: Step 1: Shrink the Path The contraction path is responsible for extracting contextual features of the image; it uses two consecutive 3x3 convolutions, each followed by a ReLU activation function; then it applies 2x2 max pooling for downsampling, doubling the number of channels after each pooling; convolution: Parameter definition: W (l) is the convolution kernel of the lth layer; * is the convolution operation; x (l-1) is the input feature map of the l-1th layer; b (l) is the bias term; Max Pooling: Parameter definition: z i,j is the feature map after downsampling; Step 2: Expand the path The expansion path is responsible for gradually restoring the spatial resolution; Use 2x2 deconvolution to expand the feature map to the original resolution; combine the feature map of the corresponding layer of the contraction path with the upsampled feature map through a splicing operation; after splicing, process it through two consecutive 3x3 convolutions and ReLU activations; Deconvolution: Parameter definition: y i,j up is the feature map after upsampling; W up [m,n] is the deconvolution kernel; Stitching: convolution: Parameter definition: y i,j contracty is the feature map from the contraction path; Step 3: Final classification layer In the last layer, 1x1 convolution is used to map the number of channels to the number of target categories; the category probability of each pixel is calculated by Softmax; 1x1 Convolution: Parameter definition: p k (x) is the predicted probability that pixel x belongs to category k; Softmax: Parameter definition: a k (x) is the activation value of category k; K is the total number of categories; Step 4: Loss Function Use pixel-wise cross entropy as the loss function, combined with weights to balance the training weights of small objects and background; Cross Entropy Loss: Parameter definition: Ω is the pixel set of the image; l(x) is the true label of pixel x; p l(x) (x) is the predicted probability of pixel x; w (x) is the pixel weight; Pixel weight: Parameter definition: w c (x) is the weight for balancing the class frequency; w0 is a fixed weight parameter; d1(x), d2(x) are the distances from the pixel to the nearest and second nearest target boundaries, respectively; σ is a smoothing parameter.

8. The method for identifying rural three-life space land use types based on deep learning according to claim 5 is characterized in that: The MultiTaskRoadExtractor framework identifies the edges of road land types, specifically including: semantic segmentation tasks, boundary detection tasks, and centerline detection tasks; This multi-task learning structure is expressed as: L MTRE =λ1L segmentation +λ2L boundary +λ3L centerline Parameter definition: L MTREL is the total loss function, L segmentation , L boundary , L centerline They represent the losses of segmentation, boundary and centerline detection respectively, λ1, λ2, λ3 are the weight coefficients of the tasks; Step 1: Semantic Segmentation Loss Semantic segmentation usually uses a pixel-level cross entropy loss function: Parameter definition: y(x) is the true label of pixel x, p(x) is the predicted probability, and Ω represents the pixel set; Step 2: Boundary Detection Loss Boundary detection is achieved through cross entropy loss based on edge maps or specific boundary loss to enhance the clarity of the boundary. The usual boundary detection loss function is: Parameter definition: y b (x) and p b (x) are the true boundary label and predicted boundary probability of pixel x respectively; Step 3: Centerline Detection Loss Centerline detection is used to ensure road connectivity, also calculated by cross entropy or other loss functions: Parameter definition: y c (x) and p c (x) are the true centerline labels and predicted probabilities, respectively.

9. The method for identifying rural three-life space land use types based on deep learning according to claim 1 is characterized in that: The S6 raster data processing steps are: Majority filter: Use the majority filter tool for smoothing, set the parameter page, select the raster data exported in S5 for input setting, select the number of adjacent pixels: four, and the replacement threshold is the majority; After the above parameters are set, select Run; Raster to Surface: Input raster is the raster data obtained in the previous step. After the above parameters are set, select Run to export surface vector data. Fusion: For input features, select the vector data obtained in the previous step. After the above parameters are set, select Run to export the surface vector data.

Images