Urban and rural space recognition method and system based on full-resolution neural network model

Through the urban and rural spatial recognition method based on the full resolution neural network model, the multi-resolution fusion unit and dice loss function are used to solve the problem of insufficient urban and rural spatial recognition accuracy and data resolution in the existing technology, and high-precision urban and rural areas are achieved Identification, especially in the identification of cities, urban-rural transition zones and rural areas, the verification results show that the accuracy reaches more than 94%.

CN119091157AInactive Publication Date: 2024-12-06INST OF AGRI RESOURCES & REGIONAL PLANNING CHINESE ACADEMY OF AGRI SCI
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411163133.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-12-06
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing urban and rural spatial recognition technology has shortcomings in accuracy and data resolution, especially the lack of effective methods for identifying urban and rural transition zones, and the lack of high-quality training samples and high-resolution data, which limits the wide range of applications.

Method used

Using a full-resolution neural network model (FR-Net), a multi-resolution fusion unit (MRFU) is constructed to integrate data features with different resolutions through preprocessing, data augmentation and cropping index data sets, and high-resolution features are retained. And use dice loss function and adaptive moment estimation optimization model training.

Benefits of technology

It realizes high-precision urban and rural spatial recognition, overcomes the problems of lack of high-quality training sets and high data resolution requirements, and can quickly and effectively identify urban, urban and rural transition zones and rural areas, with verification accuracy of more than 94%, false alarm rate and The omission rate is within an acceptable range.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091157A_ABST
    Figure CN119091157A_ABST
Patent Text Reader

Abstract

The invention provides an urban and rural space recognition method based on a full-resolution neural network model, and the method comprises the steps: S1, carrying out the preprocessing of an index data set which comprises flows with different resolutions; s2, mask data are prepared and preprocessed; s3, performing data enhancement and cutting on the index data set and the mask data; s4, constructing an FR-Net model which receives all types of streams with different resolutions, integrates the streams with different resolutions, and retains high-resolution features; s5, dividing the index data set obtained in the step S3 into a training set and a verification set, and performing model training; and S6, performing prediction by using the trained model. The invention further corresponds to an urban and rural space recognition system based on the full-resolution neural network model. According to the method, streams with different resolutions can be integrated, and the characteristic of high resolution is reserved. The problems that an existing high-quality training set lacks and the requirement for data resolution is high are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of region recognition, and more specifically, to a method and system for urban and rural space recognition based on a full-resolution neural network model. Background Art

[0002] It is estimated that by 2030, the global urban land area will triple from 650,000 km2 in 2000 to 2 Increased to 1.86 million km 2 . Although rapid urbanization has brought significant benefits in terms of economies of scale and technological progress, it has also led to various environmental consequences that limit sustainable development. In particular, the urbanization process has promoted global climate change, ecosystem degradation, and biodiversity loss, with significant impacts on the environment and human well-being. In addition, academic research and regional planning require accurate information on the spatial distribution of urban and rural areas. Therefore, it is of great significance to establish a standard and transferable method system to quickly and accurately spatially identify urban and rural areas.

[0003] Current research on urban-rural spatial identification mainly focuses on urban-rural division at the administrative scale and urban-rural spatial identification at the grid scale. Commonly used methods can be divided into three categories: 1) indicator-based unsupervised classification method; 2) remote sensing-based mutation detection method; 3) sample-based supervised classification method.

[0004] Unsupervised classification methods based on indicators mainly include cluster analysis and principal component analysis. They divide urban and rural areas by setting thresholds. Such methods are relatively simple and fast. The setting of thresholds involves subjectivity, which may lead to low spatial resolution or inaccurate identification of urban-rural transition zones. Mutation detection methods based on remote sensing mainly include wavelet transform, etc., which delineate urban and rural boundaries by monitoring mutation points of land use to achieve urban-rural identification. Such methods often involve subjective visual mapping of mutation points of land use, etc., and their accuracy is difficult to grasp. Sample-based supervised classification methods mainly include random forests, support vector machines and other methods. Such methods identify urban and rural areas through models trained with a large number of samples. Such methods have high accuracy but require a large number of high-quality samples and have high data requirements. The three types of methods have their own advantages and disadvantages. However, through the analysis of existing technologies and literature, compared with the first two types of methods, the third type of methods often have higher accuracy and smaller errors in the application of urban-rural spatial identification because they extract and learn the features of training samples. However, at present, such methods are still relatively rarely used in urban-rural identification, mainly affected by the low spatial resolution of indicator data sets and the lack of large-scale high-quality samples. Few studies have constructed high-quality training data sets by outlining samples through field surveys, and have used deep learning methods of convolutional neural networks (CNNs) for high-precision recognition. At the same time, existing technologies and research mainly identify cities and villages under the traditional urban-rural binary structure, and have not yet identified the urban-rural transition zone between urban and rural areas. Summary of the invention

[0005] In view of the problems in the background technology, the present invention proposes a method for urban and rural space identification based on a full-resolution neural network model, including: S1, preprocessing an indicator data set, wherein the indicator data set includes streams of different resolutions; S2, preparing mask data and preprocessing; S3, performing data enhancement and cropping on the indicator data set and the mask data; S4, constructing a FR-Net model, which receives all types of streams of different resolutions, integrates streams of different resolutions, and retains high-resolution features; S5, dividing the indicator data set obtained in step S3 into a training set and a validation set, and performing model training; S6, using the trained model for prediction. The present invention can integrate streams of different resolutions and retain high-resolution features.

[0006] The present invention also proposes an urban and rural space identification system based on a full-resolution neural network model, including a computer executable program, and when the executable program is executed, the method described in the above claims is completed.

[0007] The beneficial effects of the present invention are as follows: the present invention proposes a full-resolution network model (FR-Net) based on residual units, which is used to extract features of low-resolution data sets and maintain high precision. The present invention overcomes the problems of lack of existing high-quality training sets and high data resolution requirements. The present invention constructs a multi-resolution fusion unit (MRFU) that can integrate streams of different resolutions and retain high-resolution features. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to make the present invention more easily understood, the present invention will be described in more detail by referring to the specific embodiments shown in the accompanying drawings. These drawings only depict typical embodiments of the present invention and should not be considered as limiting the scope of protection of the present invention.

[0009] Figure 1 is a flow chart of the method of the present invention.

[0010] Figure 2 The present invention is a flow chart of another embodiment of the method.

[0011] Figure 3 This is the architecture diagram of the FR-Net model of the present invention.

[0012] Figure 4 Schematic diagram of a multi-resolution fusion unit of the model of the present invention

[0013] Figure 5 It is a comparison chart of the prediction results of the method of the present invention and the prior art. DETAILED DESCRIPTION

[0014] The following describes the implementation modes of the present invention with reference to the accompanying drawings so that those skilled in the art can better understand the present invention and implement it. However, the enumerated embodiments are not intended to limit the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments may be combined with each other, wherein the same components are represented by the same figure marks.

[0015] It should be noted that urban expansion is mainly reflected in the transformation of urban and rural landscape patterns, which is mainly reflected in the transformation of natural and semi-natural land to urban land. The essence is the transformation of land use types, but at the same time, the population is becoming more concentrated, economic development is becoming more rapid, and it is related to factors such as sea slope and slope. Therefore, land use type, population density, GDP, altitude, etc. frequently appear in the description. In addition, this method uses multiple indicators to characterize urban development in different dimensions, so the indicator data set is used in the description of the present invention to enhance the versatility of this technology.

[0016] The present invention is described below with reference to the accompanying drawings. In one embodiment, Figure 1 As shown, the method of the present invention includes steps S1-S6: S1, preprocessing the indicator data set; S2, preparing mask data and preprocessing; S3, performing data enhancement and cropping on the indicator data set and the mask data; S4, constructing a FR-Net model, which receives all types of streams of different resolutions, integrates streams of different resolutions, and retains high-resolution features; S5, dividing the indicator data set obtained in step S3 into a training set and a validation set, and performing model training; S6, using the trained model for prediction.

[0017] In one embodiment, in step S1, the indicator data set refers to a set of multiple indicators that the technicians believe are related to the changes in the urban-rural transition zone-rural spatial pattern. For example: land use / land cover data, population density data, altitude data, elevation / slope data, GDP, night light data, distance to urban roads, distance to urban administrative centers, etc. Figure 1 As shown, the green block represents the countryside, the yellow block represents the urban-rural transition zone, and the red block represents the city.

[0018] The operations for processing multiple indicator data include but are not limited to: "band synthesis" and "raster projection" based on ArcGIS software. The purpose of this step is to unify the spatial coordinate system and spatial resolution of the selected multiple indicator data and synthesize them into a raster dataset (.tif) to facilitate subsequent data processing.

[0019] In one embodiment, in step S2, the mask data refers to the urban and rural raster data that has been tested by the user, confirmed to be problem-free, has spatial reference, and has been classified. Use the data set publicly available on the Internet, but the accuracy cannot be guaranteed, so it is preferred to use a technician to first outline and export (.tif). Then, the spatial reference and resolution of the mask data are transformed to be consistent with the indicator data set through the "raster projection" and "resampling" operations in the ArcGIS software. The purpose of this step is to unify the spatial reference and spatial resolution of the original data (indicator data set, mask data) to facilitate subsequent data processing.

[0020] In one embodiment, in step S3, considering that the convolutional neural network (CNN) method itself requires a large number of samples, and in order to improve the computational efficiency, the present invention performs multiple (preferably 5) data enhancement operations on the preprocessed indicator data set and the corresponding mask data: 1) setting the overlap rate, defining the cropping method, and cropping the original data (indicator data set, mask data) into an image of a set size (preferably 256*256 or 128*128); 2) rotating the indicator data set and the corresponding mask data together clockwise by 5°; 3) repeating operations 1)-2) 5 times. The purpose of this step is to generate model input data from the indicator data set and the corresponding mask data as required. Figure 2 The data processing steps are shown in the figure.

[0021] In one embodiment, in step S4, the architecture diagram of the FR-Net model constructed by the present invention is as follows: Figure 2 and Figure 3 As shown, it mainly includes convolution (Conv), activation function (ReLU), batch normalization (BN), pooling, upsampling, multi-resolution fusion unit (MRFU), etc. The basic operation process of the model includes:

[0022] S41, input the training data into a convolution kernel (usually with a step size of 1 and a convolution kernel size of 3*3), and increase the number of channels to 32. By using the convolution kernel as a filter to extract corresponding features from the image, the network capacity is expanded by increasing the number of image channels to capture deeper features.

[0023] S42, the image resolution in the vertical direction is reduced to 1 / 2, 1 / 4, and 1 / 8 respectively through maximum pooling operations of different strengths. The maximum pooling operation is used to reduce the number of model parameters and the feature map so as to retain the most significant features of the input image. The purpose of reducing the resolution is to extract corresponding features at different resolutions and reduce the loss of significant features.

[0024] S43, the network is divided into 4 parallel streams in the horizontal direction, each stream performs convolution, batch normalization and non-linear activation operations of different strengths. The purpose of performing convolution, batch normalization and non-linear activation operations is to extract features at a deeper level through multi-layer convolution, so as to facilitate fusion in the multi-resolution fusion unit (MRFU).

[0025] S44, a multi-resolution fusion unit (MRFU) is used to connect and fuse multiple parallel streams of different resolutions from the previous stage. Multi-dimensional features are extracted by fusing parallel streams of multiple resolutions.

[0026] S45, upsample the three low-resolution streams and connect them together to the first full-resolution stream to obtain the final feature map. The purpose of upsampling is to eventually form a feature map with the same size and number of channels, so as to improve the accuracy of subsequent results.

[0027] Furthermore, the core of the model of the present invention is a multi-resolution fusion unit (MRFU), such as Figure 4 As shown. The multi-resolution fusion unit receives all types of different resolution streams, such as high-resolution, low-resolution, and residual inputs (horizontal streams). The multi-resolution fusion unit controls the feature flow and is able to integrate streams of different resolutions while retaining high-resolution features. The multi-resolution fusion unit is essentially a residual unit, which is mainly used to improve the training characteristics and performance of the network. The output of the multi-resolution fusion unit has two paths: one involves a 3×3 convolution with a stride of 1, batch normalization, and a non-linear (ReLU) activation layer; the other requires a 3×3 convolution operation with a stride of 2 to increase the number of units, and also includes batch normalization and ReLU activation. MRFU enables the FR-Net model to inherit the advantages of easy training and high performance of the residual network.

[0028] The output formula of the residual unit of the FR-Net model is:

[0029]

[0030] Where, X m is the output of the residual unit, X n is the input, f(X i ; W i+1 ) is W i+1 . is the residual unit with parameters.

[0031] According to the chain rule, the derivative of the loss function L is used to update the parameter W n :

[0032]

[0033] In addition, the loss function L, as a scalar, quantifies the difference between the predicted output of the model and the underlying true value by generating a numerical value, thereby representing the loss of the model. Therefore, by minimizing the loss, the parameters of the model can be optimized to enhance its accuracy and generalization ability. Considering the quantitative imbalance between the three types of cities, urban-rural transition zones, and rural areas in the application scenario, the present invention adopts dice loss (DSC) as the loss function L to overcome this problem.

[0034]

[0035] Where A∩B represents the intersection of predicted city pixels (A) and ground truth city pixels (B), and their respective city pixel numbers are |A| and |B|. The present invention uses adaptive moment estimation (Adam) as an optimizer.

[0036] In one embodiment, in step S5, for the data set obtained in step S3, the trimmed data set is divided into a training set and a validation set at a ratio of 0.85:0.15, and the training set is input into the constructed initial FR-Net model for training. The batch size (batch_size) in the training is set to 8, and the number of training epochs (epochs) is set to 150. The parameters in the model will be continuously updated as the training data continues to increase. After each model training, the accuracy (Accuracy) and loss (Loss) of the model are verified using the validation data set. If the accuracy is improved and the loss converges, the model parameters are saved for the next training. Otherwise, they are not saved. The purpose of this step is to continuously update the model parameters through continuous training and verification, and save the optimal model.

[0037] In one embodiment, in step S6, the target year indicator dataset is processed with reference to the operation on the training dataset in step S1, and the data type, spatial resolution, and spatial reference of each band are processed. After processing, the optimal FR-Net model in step S5 is input to predict the urban and rural identification results of the target year.

[0038] The present invention also proposes an urban and rural space identification system based on a full-resolution neural network model, including a computer executable program, and when the executable program is executed, the method described in the above claims is completed.

[0039] The present invention uses the identification of urban-rural transition zone-rural space in the Beijing-Tianjin-Hebei region in 2000, 2005, 2010, 2015, and 2020 as an example to verify the results. The FR-Net model uses the overall accuracy (OA) and Micro-F1 score for verification, and the verification accuracy is above 94%, and the omission rate and false alarm rate are both within an acceptable range (Table 1). At the same time, through comparative analysis with existing research and public data sets, it is found that the matching degree of the results is above 85% ( Figure 5 ). This shows that the developed model can achieve high-precision urban and rural spatial recognition. In particular, MRFU can quickly and effectively maintain high-resolution streams and achieve high-performance feature extraction.

[0040] Table 1 Model evaluation

[0041]

[0042] Based on field research in the Beijing-Tianjin-Hebei region, the present invention provides a model (.h5 format) trained with a self-drawn mask data set including three types of cities, urban-rural transition areas, and villages, which can be used as an example for urban and rural identification in other years in the Beijing-Tianjin-Hebei region. The indicator data set includes: altitude (m), slope (°), GDP (10,000 yuan), population density (people), land use, night lights, distance from the center of the prefecture-level city (km), distance from the city road (km), the spatial reference is the Albers projection coordinate system, and the spatial resolution is 100m.

[0043] The embodiments described above are only preferred specific implementations of the present invention. This specification uses the phrases "in one embodiment", "in another embodiment", "in yet another embodiment" or "in other embodiments", which may refer to one or more of the same or different embodiments in this announcement. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for urban and rural space recognition based on a full-resolution neural network model, characterized in that: include: S1, preprocessing an indicator data set, wherein the indicator data set includes streams of different resolutions; S2, prepare mask data and perform preprocessing; S3, data augmentation and cropping of the indicator dataset and mask data; S4, build the FR-Net model, which receives all types of streams with different resolutions, integrates streams with different resolutions, and retains high-resolution features; S5, dividing the indicator data set obtained in step S3 into a training set and a validation set to perform model training; S6, use the trained model to make predictions.

2. The method for identifying urban and rural space according to claim 1, characterized in that: The data enhancement method in step S3 includes: S31, set the overlap rate, define the cropping method, and crop the index data set and the mask data into an image of a set size; S31, rotate the original data clockwise by a set angle; S31, repeat the operations of steps S31-S32 a set number of times.

3. The method for identifying urban and rural space according to claim 1, characterized in that: Step S4 includes: S41, input the training data into the convolution kernel, extract the corresponding features from the image by using the convolution kernel as a filter, and expand the network capacity by increasing the number of image channels; S42, reduce the number of model parameters and reduce the feature map through the maximum pooling operation to retain the most significant features of the input image; S43, divide the network into 4 parallel streams in the horizontal direction, each stream performs convolution, batch normalization and non-linear activation operations; S44, connect and fuse multiple parallel streams of different resolutions from the previous stage; S45, upsample the three low-resolution streams and connect them together to the first full-resolution stream to obtain the final feature map.

4. The method for identifying urban and rural space according to claim 3, characterized in that: In step S41, the training data is input into a 3×3 convolution kernel with a stride of 1, and the number of channels is increased to 32; in step S42, the image resolution in the vertical direction is reduced to 1 / 2, 1 / 4, and 1 / 8 respectively through maximum pooling.

5. The method for identifying urban and rural space according to claim 4, characterized in that: The step S44 is completed by the multi-resolution fusion unit. The multi-resolution fusion unit is a residual unit whose output includes two paths: one involves a 3×3 convolution with a stride of 1, batch normalization, and a non-linear activation layer; the other involves a 3×3 convolution operation with a stride of 2, batch normalization, and an activation function.

6. The method for identifying urban and rural space according to claim 4, characterized in that: The output formula of the residual unit of the FR-Net model is: Where, X m is the output of the residual unit, X n is the input, f(X i ; W i+1 ) is W i+1 . is the residual unit with parameters.

7. The method for identifying urban and rural space according to claim 6, characterized in that: Also includes: Use the derivative of the loss function L to update the parameter W n :

8. The method for identifying urban and rural space according to claim 7, characterized in that: Also includes: The loss function L is the dice loss function DSC: Where A∩B represents the intersection of the predicted city pixel A and the ground truth city pixel B, and their respective city pixel numbers are |A| and |B|.

9. The method for identifying urban and rural space according to claim 8, characterized in that: Step S2 includes: The spatial coordinate systems and spatial resolutions of multiple indicator data are unified and synthesized into a raster dataset.

10. A system for identifying urban and rural space based on a full-resolution neural network model, characterized in that: The invention comprises a computer executable program, and when the executable program is executed, the method as claimed in any one of claims 1 to 9 is performed.

Citation Information

Patent Citations

  • Middle and small city marginal area extraction method based on multi-source data

    CN116310853A

  • Urban and rural space structure identification method based on night light two-dimensional characteristics

    CN117876871A