CN-121982546-A - Natural resource remote sensing image detection system and method based on visual recognition

CN121982546ACN 121982546 ACN121982546 ACN 121982546ACN-121982546-A

Abstract

The invention discloses a natural resource remote sensing image detection system and a method based on visual identification, and relates to the technical field of remote sensing identification.A professional performs circle selection labeling on natural resource elements such as water sources, woodland and cultivated land by collecting historical remote sensing images of a target area to generate a natural resource label graph with aligned pixel levels; the method comprises the steps of dividing a year into a plurality of climates based on the climatic features, and dividing the image and the label map into corresponding sample subsets of the climates according to imaging dates. And (5) independently training a plurality of expert models by utilizing each sample subset, and constructing a multi-climate expert model library. And selecting a corresponding expert model to output a main label image according to the imaging date, inputting other models to obtain a plurality of contrast label images, and counting the total difference number by comparing pixel by pixel. And if the difference number does not exceed the consistency threshold, outputting a primary and secondary natural resource identification result, otherwise, triggering manual arbitration, and feeding back the arbitration result to the corresponding sample subset after labeling.

Inventors

YUE ZHIQIANG
JI ZHE
WEI LAI
Xu ali
WANG QIANG
SUN FANGFANG
Lin Layue
GUAN HONG
ZHEN XIANG
LI JIAQI

Assignees

长春市测绘院

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. The natural resource remote sensing image detection method based on visual identification is characterized by comprising the following steps of: Step 100, collecting a historical remote sensing image of a target area and performing artificial labeling to generate a natural resource label graph, dividing K weathers based on the weathers of the target area, and dividing the historical remote sensing image and the label graph thereof into K weathers sample subsets according to imaging dates; Step 200, respectively and independently training K expert models by utilizing the K weathered sample subsets, wherein each expert model is trained to receive an image and output a label graph; Step 300, inputting a remote sensing image to be detected, selecting a corresponding expert model for interpretation according to an imaging date, outputting a main label image, inputting the remote sensing image to be detected into other expert models to obtain K-1 contrast label images, comparing the contrast label images with the main label image pixel by pixel, and counting the total difference number; And step 400, comparing the total difference number with a consistency threshold, judging that the main label graph is reliable when the total difference number is smaller than or equal to the consistency threshold, outputting a natural resource identification result of the current remote sensing image, triggering an alarm when the total difference number is larger than the consistency threshold, and manually arbitrating the natural resource element category of the remote sensing image to be detected to be used as a training sample.
2. The method for detecting remote sensing images of natural resources based on visual recognition as set forth in claim 1, wherein the step S100 comprises the steps of: Step S101, collecting historical remote sensing images of a target area, performing circle selection marking on natural resource elements in the images by a professional, and generating a natural resource tag image for each image, wherein the natural resource elements are natural geographic entities which need to be identified and marked in the remote sensing images, the natural resource tag image is raster data which completely corresponds to the space size of the original remote sensing images, each pixel value in the raster data is endowed with a category ID, and the category numbers of the natural resource elements which belong to pixels at corresponding positions in the original remote sensing images are represented; Step S102, dividing a year into K typical weathers which are mutually disjoint based on the weathers of the target area by a professional, and marking the typical weathers as a weathers set S= { S 1 ,S 2 ,...,S K }; And step 103, dividing the historical remote sensing images and the natural resource label graph thereof into corresponding subsets of the weathers according to the imaging date of each historical remote sensing image to form K weathers sample subsets D 1 ,D 2 ,...,D K .
3. The method for detecting remote sensing images of natural resources based on visual recognition as set forth in claim 2, wherein the step S200 comprises the steps of: S201, utilizing a semantic segmentation network U-Net as an expert model, wherein the semantic segmentation network receives a historical remote sensing image of H rows, W columns and X3 channels as a training sample, outputs a segmentation map of H rows, W columns and X C channels, wherein H and W are respectively the pixel height and the pixel width of the image, C is the total number of natural resource element categories, and a kth candidate sample subset D k is used as the training sample to independently train a kth expert model M k ; Step S202, inputting the training sample into a semantic segmentation network U-Net, performing convolution calculation on each layer of the semantic segmentation network U-Net, and training and outputting a predicted value of each pixel belonging to each natural resource element class number by adopting a pixel level cross entropy loss function; Step S203, after training is completed, K expert models { M 1 ,M 2 ,...,M K }, which are used for forming a multi-object expert model library, are obtained, the expert models receive an image and output a label graph of H rows and W columns, the label graph is a two-dimensional matrix, and each pixel in the two-dimensional matrix takes the natural resource element class number with the highest predicted value as the class ID of the current pixel.
4. The method for detecting remote sensing images of natural resources based on visual recognition as set forth in claim 3, wherein said step S300 comprises the steps of: Step 301, for an input remote sensing image I to be detected, determining a belonging weather period S x , x epsilon [1, k ] according to an imaging date, selecting an expert model M x , performing visual identification on the remote sensing image I to be detected, and outputting a main label image L x , wherein L x is a matrix of H rows and W columns; Step S302, sequentially inputting the remote sensing image I to be detected into the remaining K-1 expert models, and outputting a contrast label image L i by each model M i (I is not equal to x), wherein L i is a matrix of H rows and W columns; And S303, comparing each comparison label image L i with the main label image L x pixel by pixel, comparing the values of L i (h, w) and L x (h, w), wherein (h, w) is the pixel position, h epsilon [1, H ], w epsilon [1, W ], when L i (h,w)≠L x (h, w) is adopted, judging that the expert model M i and the expert model M x have a difference at the pixel position, traversing all K-1 comparison label images, and counting the total number x of the differences.
5. The method for remote sensing image detection of natural resources based on visual recognition as set forth in claim 4, wherein the step S400 comprises the steps of: step S401, comparing the total number x of differences with a preset consistency threshold T: When x is less than or equal to T, judging that an interpretation result L x of an expert model M x is reliable, counting the number of pixels appearing in each class ID in an L x matrix, selecting the class ID with the largest number of pixels, determining natural resource elements corresponding to the class ID as main natural resources of the current remote sensing image I, sequencing the rest of the appearing class IDs from high to low according to the number of pixels, and determining the corresponding natural resource elements as secondary natural resources of the current remote sensing image I; When x > T, triggering an alarm by the system, and manually arbitrating to confirm the nature resource element category of the remote sensing image I to be detected; And step S402, after the correct natural resource elements confirmed by the manual arbitration, carrying out circle selection labeling on the natural resource elements in the current remote sensing image, generating a natural resource label graph for each image, and feeding back to a physical sample subset of the corresponding date as a training sample.
6. The system is characterized by comprising a data preparation and physical sample division module, a multi-physical expert model training module, a cross verification and difference calculation module and an arbitration and model evolution module; The data preparation and climate sample division module collects historical remote sensing images of the target area and performs manual labeling to generate a natural resource tag map; dividing the historical remote sensing image and a label graph thereof into K weather period sample subsets according to imaging dates; The multi-weatherometer expert model training module is used for independently training K expert models by utilizing the K weatherometer sample subsets, and each expert model is trained to receive an image and output a label graph; The cross verification and difference calculation module inputs a remote sensing image to be detected, selects a corresponding expert model for interpretation according to an imaging date, outputs a main label image, inputs the remote sensing image to be detected into other expert models to obtain K-1 contrast label images, compares the contrast label images with the main label image pixel by pixel, and counts the total difference number; The arbitration and model evolution module compares the total difference number with a consistency threshold, judges that the main label graph is reliable when the total difference number is smaller than or equal to the consistency threshold, and outputs a natural resource identification result of the current remote sensing image, when the total difference number is larger than the consistency threshold, an alarm is triggered, and the natural resource element category of the remote sensing image to be detected is arbitrated by manpower to be used as a training sample.
7. The natural resource remote sensing image detection system based on visual identification is characterized in that the data preparation and physical sample division module comprises a natural resource labeling unit and a sample organization unit; the natural resource labeling unit performs circle selection and identification on natural resource elements in the historical remote sensing images, generates a natural resource label graph aligned with pixels of each image, and assigns a preset category ID to each pixel point in the natural resource label graph; The sample organization unit divides one year into K typical climates according to the climatic characteristics of the target area, and classifies each historical image and the corresponding natural resource label graph into corresponding climatic period sample subsets according to the imaging date of each historical image.
8. The natural resource remote sensing image detection system based on visual identification is characterized in that the multi-weather expert model training module comprises a model framework, a training unit and an expert model library construction unit; The model framework and the training unit select a semantic segmentation network U-Net as a basic framework of the model, and the semantic segmentation network U-Net receives a remote sensing image with a standard size and outputs a segmentation map corresponding to the space size; The expert model library construction unit uses each candidate period sample subset to independently train the corresponding expert model, and the training process optimizes the pixel-level cross entropy loss function to enable each model to learn to classify the image pixels into different natural resource categories in the exclusive candidate period, and K expert models jointly form a multi-candidate expert model library after training is completed.
9. The natural resource remote sensing image detection system based on visual recognition is characterized in that the cross verification and difference calculation module comprises a multi-model parallel interpretation unit and a difference statistics unit; The multi-model parallel interpretation unit interprets the input image to be detected, invokes the corresponding expert model according to the imaging date to generate a main label image, and then inputs the same image into all other expert models to obtain a plurality of contrast label images from different object viewing angles; and the difference statistics unit compares each comparison label graph with the main label graph pixel by pixel, counts the comparison label graph as one difference when the class IDs of the same pixel position are inconsistent, and counts the total difference number after traversing all comparison models and all pixels.
10. The natural resource remote sensing image detection system based on visual identification is characterized in that the arbitration and model evolution module comprises a result output unit and a closed loop feedback unit; the result output unit compares the total difference number with a preset consistency threshold value, when the divergence degree is low, the interpretation result of the main label graph is judged to be reliable, the system automatically analyzes the main label graph and outputs the identification result classified by main and secondary natural resources; the closed loop feedback unit marks the image to be detected through manual arbitration for the case with high divergence degree, generates a new natural resource label graph according to the correct marking result, and feeds back the new natural resource label graph to the corresponding sub-sample set of the weathered period according to the imaging date of the image.

Description

Natural resource remote sensing image detection system and method based on visual recognition Technical Field The invention relates to the technical field of remote sensing identification, in particular to a natural resource remote sensing image detection system and method based on visual identification. Background With the rapid development of remote sensing technology, satellite and aerial remote sensing images have become important means for acquiring large-scale earth surface information, and are widely applied to the fields of natural resource investigation, environment monitoring, homeland planning and the like. The traditional natural resource identification method mainly relies on manual visual interpretation, and has the defects of high accuracy, low efficiency and difficulty in meeting the monitoring requirements of a large range and high frequency. In recent years, computer vision technology based on deep learning, in particular to semantic segmentation models (such as U-Net, segNet and the like), has great potential in remote sensing image interpretation. The models can realize end-to-end pixel level classification, and automatically identify the ground object types in the images, such as natural resource elements of water bodies, woodlands, cultivated lands and the like. Most of them train and infer based on a single model, neglecting the remarkable changes of natural resources in different seasons and different climates. For example, the vegetation coverage and the water body range of the same region in spring and autumn may have large differences, if the same model is used for identification, erroneous judgment or omission is easily caused, and accuracy and robustness of an identification result are affected. In addition, the existing system lacks a consistency checking mechanism for the model output result, and the reliability of the identification result cannot be automatically evaluated in an inference stage. When the models are inconsistent in judgment on some complex scenes or seasonal variation sensitive areas, error results are directly output, and the accuracy of subsequent decisions is affected. Therefore, there is a need for a system and a method for intelligent detection of remote sensing images of natural resources, which can integrate multi-phase and multi-weather information and have self-verification and continuous evolution capabilities, so as to improve the recognition accuracy and the practicability of the system. Disclosure of Invention The invention aims to provide a natural resource remote sensing image detection system and method based on visual identification, so as to solve the problems in the prior art. In order to achieve the purpose, the invention provides the following technical scheme that the natural resource remote sensing image detection method based on visual identification comprises the following steps: Step 100, collecting a historical remote sensing image of a target area and performing artificial labeling to generate a natural resource tag image, dividing K weathers based on the weathers of the target area, dividing the historical remote sensing image and the tag image thereof into K weathers sample subsets according to imaging dates, and providing a sample set for season adaptation for subsequent model training, so that the model can learn typical features of natural resources in different periods, and the identified season adaptability is improved; the step S100 includes the steps of: Step S101, collecting historical remote sensing images of a target area, performing circle selection and marking on natural resource elements in the images by a professional, and generating a natural resource tag image for each image, wherein the natural resource elements are natural geographic entities which need to be identified and marked in the remote sensing images and at least comprise three major categories of water sources, forest lands and cultivated lands, the natural resource tag image is grid data completely corresponding to the space size of the original remote sensing images, each pixel value in the grid data is endowed with a category ID and represents the category number of the natural resource element to which the pixel at the corresponding position in the original remote sensing images belongs, for example, 0-background, 1-water sources, 2-forest lands and 3-cultivated lands; Step S102, dividing a year into K typical weathers which are mutually disjoint based on the weathers of the target area by a professional, and marking the typical weathers as a weathers set S= { S 1,S2,...,SK }; Step S103, dividing the historical remote sensing images and natural resource label images thereof into corresponding subsets of the weathers according to the imaging date of each historical remote sensing image to form K weathers sample subsets, namely D 1,D2,...,DK, wherein the constructed weathers sample subsets with clear structures and clear seasonal features lay a high-q