CN-121170361-B - Intelligent classification method and device for complex scene of geographic partition collaborative coding

CN121170361BCN 121170361 BCN121170361 BCN 121170361BCN-121170361-B

Abstract

The application provides an intelligent classification method and device for complex scenes with collaborative codes of geographic partitions, which belong to the technical field of image recognition using classification, and comprise the steps of obtaining comprehensive geographic partitions of a research area, and carrying out partition coding on each comprehensive geographic partition to obtain partition coding vectors; the method comprises the steps of constructing an image coding model based on a deep learning model, taking a high-resolution remote sensing image as input data, using the image coding model to conduct scene classification on the remote sensing image, generating an image coding vector based on a last full-connection layer of the image coding model, connecting partition coding vectors with the image coding vector to generate a feature expression vector with collaborative partitions, constructing a scene supervision classification model based on the feature expression vector, conducting iterative training on the scene supervision classification model until the model converges to obtain a trained scene supervision classification model, and classifying scene units to be classified by utilizing the trained scene supervision classification model. The method can improve the accuracy of scene classification.

Inventors

WANG ZHIHUA
YANG XIAOMEI
LIU YUEMING
ZHANG QINGYANG
Gao Ku

Assignees

中国科学院地理科学与资源研究所

Dates

Publication Date: 20260505
Application Date: 20250807

Claims (10)

1. The intelligent classification method for the complex scene of the geographic partition collaborative coding is characterized by comprising the following steps of: The method comprises the steps of obtaining comprehensive geographic partitions of a research area, carrying out partition coding on each comprehensive geographic partition to obtain partition coding vectors, wherein each comprehensive geographic partition comprises a plurality of scene units, and the scene units share one partition coding under the same comprehensive geographic partition; constructing an image coding model based on a deep learning model, using a high-resolution remote sensing image as input data, using the image coding model to carry out scene classification on the remote sensing image, generating an image coding vector based on the last full-connection layer of the image coding model, and connecting the partition coding vector with the image coding vector to generate a feature expression vector with cooperative partition; Constructing a scene supervision classification model based on the feature expression vector, and carrying out iterative training on the scene supervision classification model until the model converges to obtain a trained scene supervision classification model; and classifying the scene units to be classified by using the trained scene supervision classification model to obtain scene classification results.
2. The method according to claim 1, further comprising the step of scene unit division, in particular as follows: Collecting road network, river water system and/or homeland resource boundary line data; and based on the road network, river water system and/or homeland resource boundary line data, spatial superposition is adopted to carry out spatial division on the research area, so as to obtain scene units.
3. The method according to claim 1, wherein the scene supervision classification model is iteratively trained until the model converges to obtain a trained scene supervision classification model, specifically: based on a pre-manufactured scene sample library, performing iterative training on the scene supervision classification model by adopting two-stage training; The two-stage training is that in the first stage, based on a prefabricated scene sample library, the partition coding vector is kept unchanged, and only the image coding model is subjected to iterative training to obtain an optimized image coding vector; and in the second stage, the partition coding vector and the image coding vector are changed at the same time, and the scene supervision classification model is subjected to iterative training until the model converges, so that a trained scene supervision classification model is obtained.
4. The method according to claim 1, wherein the scene supervision classification model is iteratively trained until the model converges to obtain a trained scene supervision classification model, specifically: based on a pre-manufactured scene sample library, performing iterative training on the scene supervision classification model by adopting a collaborative training method; the collaborative training is to perform collaborative iterative training on the image coding model and the scene supervision and classification model based on a pre-manufactured scene sample library until the model converges to obtain a trained scene supervision and classification model.
5. The method of claim 1, wherein the constructing a scene supervision classification model based on the feature expression vector comprises: And constructing a function mapping relation between the partition cooperative feature expression vector and the scene type by taking the partition cooperative feature expression vector as an independent variable and the scene type as a dependent variable so as to obtain the scene supervision classification model.
6. The method of claim 1, wherein the scene supervision classification model is trained based on a scene sample library obtained by: and carrying out sample preparation according to a training sample construction strategy which is randomly extracted in the geographic subareas, proportionally extracted among the geographic subareas and proportionally extracted among different scene types and simultaneously controls the minimum sample quantity of each geographic subarea so as to obtain the scene sample library.
7. The method of any one of claims 1-6, wherein the encodable range of the partition encoding vector is greater than or equal to the number of all geographic partitions of the investigation region.
8. The utility model provides a geographical subregion is complex scene intelligence sorter of coding in coordination which characterized in that includes: The system comprises a partition coding unit, a partition coding unit and a partition coding unit, wherein the partition coding unit is configured to acquire comprehensive geographical partitions of a research area, and perform partition coding on each comprehensive geographical partition to acquire partition coding vectors; the collaborative coding unit is configured to construct an image coding model based on a deep learning model, take a high-resolution remote sensing image as input data, use the image coding model to carry out scene classification on the remote sensing image, generate an image coding vector based on the last full-connection layer of the image coding model, and connect the partition coding vector with the image coding vector to generate a feature expression vector of partition collaborative; the model construction and training unit is configured to construct a scene supervision classification model based on the feature expression vector, and iteratively train the scene supervision classification model until the model converges to obtain a trained scene supervision classification model; the scene classification unit is configured to classify the scene units to be classified by using the trained scene supervision classification model to obtain scene classification results.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method according to any one of claims 1-7.
10. A computer readable storage medium having stored thereon a computer program/instruction, which when executed by a processor, implements the steps of the method of any of claims 1 to 7.

Description

Intelligent classification method and device for complex scene of geographic partition collaborative coding Technical Field The application relates to the technical field of image recognition by using classification, in particular to an intelligent classification method and device for complex scenes by geographic partition collaborative coding. Background Geographical partitioning is considered as a very effective knowledge integration approach in the intelligent classification of complex scene remote sensing, and more researches are being acquired. However, the current intelligent classification method for remote sensing of complex scenes based on geographical partitions mostly adopts a partition Shi Ce mode, namely training and classifying by one model in one area. Although the prior method can apply geographic partition to emphasize the difference between different areas of the same type of scene, the similarity of the same type of scene between the different areas is ignored at the same time. Studies have shown that if geographical partition intelligent interpretation of "partition Shi Ce" is used entirely, there is often a problem that classification accuracy does not rise and fall due to insufficient sample of some areas. Accordingly, there is a need to provide an improved solution to the above-mentioned deficiencies of the prior art. Disclosure of Invention The application aims to provide a complex scene intelligent classification method and device for geographic partition collaborative coding, which can solve or alleviate the problems in the prior art by designing a classification model framework which can not only consider the differences among areas but also consider the similarity of similar scenes. In order to achieve the above object, the present application provides the following technical solutions: In a first aspect, the present application provides a method for intelligently classifying complex scenes by geopartition cooperative coding, including: The method comprises the steps of obtaining comprehensive geographic partitions of a research area, carrying out partition coding on each comprehensive geographic partition to obtain partition coding vectors, wherein each comprehensive geographic partition comprises a plurality of scene units, and the scene units share one partition coding under the same comprehensive geographic partition; constructing an image coding model based on a deep learning model, using a high-resolution remote sensing image as input data, using the image coding model to carry out scene classification on the remote sensing image, generating an image coding vector based on the last full-connection layer of the image coding model, and connecting the partition coding vector with the image coding vector to generate a feature expression vector with cooperative partition; Constructing a scene supervision classification model based on the feature expression vector, and carrying out iterative training on the scene supervision classification model until the model converges to obtain a trained scene supervision classification model; and classifying the scene units to be classified by using the trained scene supervision classification model to obtain scene classification results. In one possible embodiment, the method further comprises the step of scene unit division, in particular by collecting road network, river water system, and/or homeland resource boundary line data; and based on the road network, river water system and/or homeland resource boundary line data, spatial superposition is adopted to carry out spatial division on the research area, so as to obtain scene units. In one possible implementation manner, the iterative training is performed on the scene supervision classification model until the model converges to obtain a trained scene supervision classification model, which specifically includes: based on a pre-manufactured scene sample library, performing iterative training on the scene supervision classification model by adopting two-stage training; The two-stage training is that in the first stage, based on a prefabricated scene sample library, the partition coding vector is kept unchanged, and only the image coding model is subjected to iterative training to obtain an optimized image coding vector; and in the second stage, the partition coding vector and the image coding vector are changed at the same time, and the scene supervision classification model is subjected to iterative training until the model converges, so that a trained scene supervision classification model is obtained. In one possible implementation manner, the iterative training is performed on the scene supervision classification model until the model converges to obtain a trained scene supervision classification model, which specifically includes: based on a pre-manufactured scene sample library, performing iterative training on the scene supervision classification model by adopting a collaborative t