CN-121982710-A - Semi-automatic labeling method and system for remote sensing image

CN121982710ACN 121982710 ACN121982710 ACN 121982710ACN-121982710-A

Abstract

The application relates to a semi-automatic remote sensing image labeling method and a system thereof, which realize remote sensing image labeling with high efficiency and high quality. The method comprises the steps of preprocessing a to-be-processed remote sensing image to obtain a remote sensing image slice sample set, screening the remote sensing image slice sample set by adopting an active learning strategy based on a pre-training CNN model to obtain a remote sensing image slice, extracting depth features of the remote sensing image slice by adopting the pre-training CNN model and carrying out multi-scale feature fusion to obtain a fusion feature map, reconstructing the fusion feature map to obtain a pixel-feature matrix, predicting the pixel-feature matrix by adopting a pre-training random forest model to obtain class probability of pixels in the remote sensing image slice, labeling the remote sensing image slice according to the class probability of the pixels in the remote sensing image slice to obtain a pixel-level probability map, carrying out boundary refinement processing on the pixel-level probability map by adopting a conditional random field mechanism to obtain a pre-labeling map, and carrying out manual correction on the pre-labeling map to obtain an accurate labeling map.

Inventors

LIN XIAOBO
Feng Gongxue
AI BIN
WANG ZHEN
Qu Poyuan
WANG JIALIN
LUO XIAOMEI
LI LUYAN
SHI XIAOCHUN
Xu Gengran

Assignees

广东省国土资源测绘院
中山大学

Dates

Publication Date: 20260505
Application Date: 20251209

Claims (10)

1. A semi-automatic labeling method for remote sensing images is characterized by comprising the following steps: S1, preprocessing a remote sensing image to be processed, and constructing a remote sensing image slice sample set; S2, screening the remote sensing image slice sample set by adopting an active learning strategy based on a pre-training CNN model to obtain a remote sensing image slice; S3, extracting depth features of the remote sensing image slices by adopting a pre-training CNN model and carrying out multi-scale feature fusion to obtain a fusion feature map; S4, reconstructing the fusion feature map to obtain a pixel-feature matrix, predicting the pixel-feature matrix by adopting a pre-training random forest model to obtain the class probability of pixels in the remote sensing image slice, and labeling the remote sensing image slice according to the class probability of the pixels in the remote sensing image slice to obtain a pixel-level probability map; s5, carrying out boundary refinement treatment on the pixel-level probability map by adopting a conditional random field mechanism to obtain a pre-labeling map; and S6, manually correcting the pre-labeling graph to obtain an accurate labeling graph.
2. The method for semi-automatic labeling of remote sensing images according to claim 1, wherein step S1 comprises: Performing gridding treatment on the remote sensing image to be treated by adopting a sliding window cutting algorithm to obtain a remote sensing image slice set; zero filling is respectively carried out on the edge area of each remote sensing image slice in the remote sensing image slice set, so that a unified-size remote sensing image slice set is obtained; performing region selection on the remote sensing image to be processed by adopting a preset interactive map interface to obtain a target image region; and screening the unified-size remote sensing image slice set according to the target image area to obtain a remote sensing image slice sample set.
3. The method for semi-automatic labeling of remote sensing images according to claim 1, further comprising, prior to step S2: manually labeling the remote sensing image to be processed in a preset small-range area to obtain an initial labeling diagram; and inputting the training set formed by the initial annotation graph into a CNN model for training to obtain a pre-training CNN model.
4. The method for semi-automatic labeling of remote sensing images according to claim 1, wherein step S2 comprises: respectively predicting each remote sensing image slice in the remote sensing image slice sample set by adopting a pre-training CNN model to obtain the class probability of pixels in each remote sensing image slice; Calculating the normalized information entropy of each remote sensing image slice according to the class probability of the pixels in each remote sensing image slice, and obtaining the average information entropy value of the pixels in each remote sensing image slice; and screening the remote sensing image slice sample set according to the average information entropy value to obtain a remote sensing image slice.
5. The method for semi-automatic labeling of remote sensing images according to claim 4, wherein the average information entropy value The computational expression is as follows: Wherein, the And Representing the width and height of the remote sensing image slice respectively, The number of categories is indicated and, Representing pixels Probability of belonging to category c.
6. The method for semi-automatic labeling of remote sensing images according to claim 1, wherein step S2 comprises: Carrying out version transformation on each remote sensing image slice in the remote sensing image slice sample set according to a transformation combination formed by a preset geometric transformation and radiation transformation to obtain a transformation version of each remote sensing image slice; Respectively predicting each remote sensing image slice and a transformation version thereof in the remote sensing image slice sample set by adopting a pre-training CNN model to obtain the class probability of pixels in each remote sensing image slice and the class probability of pixels in the transformation version thereof; calculating to obtain Jensen-Shannon divergence according to the class probability of the pixels in each remote sensing image slice and the class probability of the pixels in the transformed version of each remote sensing image slice; And screening the remote sensing image slice sample set according to the Jensen-Shannon divergence to obtain a remote sensing image slice.
7. The method for semi-automatic labeling of remote sensing images according to claim 6, wherein the Jensen-Shannon divergence The computational expression is as follows: Wherein, the Indicating that each remote sensing image slice has The number of transformed versions is chosen such that, The average class probability distribution vector of the pixels representing each remote sensing image slice is generated by counting the average value of class probabilities of the pixels in each remote sensing image slice, Represent the first The pixel mean class probability distribution vectors of the transformed versions, Indicating Kullback-Leibler divergence, Representation of And (3) with Is arranged in the middle of the distribution, Representation of And To the point of Is a symmetric average of KL divergence of (c).
8. The method for semi-automatic labeling of remote sensing images according to claim 3, further comprising, prior to step S4: sampling the corresponding positions of the fusion feature images according to the geometric coordinate information of the initial annotation image to obtain training samples; Dividing the training sample into a training set and a testing set according to a preset proportion by adopting a layered sampling strategy; And extracting a characteristic-label pair of the initial annotation map, and inputting the training set, the testing set and the characteristic-label pair into a random forest model based on a Bootstrap aggregation and random characteristic selection mechanism for training to obtain the pre-training random forest model.
9. The method according to claim 3, further comprising adding the accurate annotation map to a training set formed by the initial annotation map after step S6.
10. The semi-automatic remote sensing image labeling system is characterized by comprising a preprocessing module, a screening module, a feature fusion module, a prediction module, a refinement processing module and a manual correction module; The preprocessing module is used for preprocessing the remote sensing image to be processed and constructing a remote sensing image slice sample set; the screening module is used for screening the remote sensing image slice sample set by adopting an active learning strategy based on a pre-training CNN model to obtain a remote sensing image slice; The feature fusion module is used for extracting depth features of the remote sensing image slices by adopting a pre-training CNN model and carrying out multi-scale feature fusion to obtain a fusion feature map; the prediction module is used for reconstructing the fusion feature map to obtain a pixel-feature matrix, predicting the pixel-feature matrix by adopting a pre-training random forest model to obtain the class probability of the pixels in the remote sensing image slice, and labeling the remote sensing image slice according to the class probability of the pixels in the remote sensing image slice to obtain a pixel-level probability map; The refinement processing module is used for performing boundary refinement processing on the pixel-level probability map by adopting a conditional random field mechanism to obtain a pre-labeling map; And the manual correction module is used for manually correcting the pre-annotation graph to obtain an accurate annotation graph.

Description

Semi-automatic labeling method and system for remote sensing image Technical Field The application relates to the technical field of artificial intelligence and remote sensing image processing, in particular to a semi-automatic labeling method and system for remote sensing images. Background The remote sensing image annotation is basic work of remote sensing image processing and analysis, and has important significance in the fields of land utilization monitoring, resource investigation, environmental protection and the like. With the rapid development of the high-resolution satellite remote sensing technology, the acquired remote sensing image data volume grows exponentially, and urgent demands are made on the efficient and accurate image labeling technology. Traditional remote sensing image labeling mainly relies on professional personnel to perform manual visual interpretation. And a professional manually sketches the target area and carries out category labeling according to the information such as spectral features, texture features, geometric shapes and the like in the image and combining prior knowledge. However, this conventional method has significant limitations. Firstly, the manual labeling efficiency is extremely low, a large amount of time and labor cost are required for processing large-scale remote sensing images, and the requirement of rapid processing in practical application is difficult to meet. Secondly, the labeling quality is highly dependent on the professional level and experience of operators, the subjectivity is strong, and the consistency of results among different labeling operators is difficult to ensure. In addition, in the face of complicated ground object types and fuzzy boundary areas, omission and misjudgment are easy to occur in manual labeling, and the accuracy of subsequent analysis is affected. In order to improve the labeling efficiency, the prior art uses a deep learning technology to construct an automatic semantic segmentation model for labeling. The method generally adopts a convolutional neural network to carry out pixel-level classification on the remote sensing image, and realizes automatic identification and segmentation of a target area. However, training of the deep learning model requires a large amount of high-quality annotation data as a supervisory signal, and obtaining these annotation data itself faces the aforementioned artificial annotation dilemma, which creates a cyclic dependency problem. Meanwhile, due to the complexity and diversity of remote sensing image scenes, the accuracy standard required by practical application is difficult to achieve by simply relying on an automatic model, and particularly, the performance is poor when the targets with fuzzy boundaries and similar categories are processed. The patent name is 201910491117.7, which is found by searching the prior art document and is a semantic labeling method of a remote sensing image based on reinforcement learning, and the method comprises the steps of data acquisition, data preprocessing, image data cutting, sample set labeling, reinforcement learning network model construction, training parameter setting, training set selection, training of a visual semantic embedded network by using the training set, training of a value network by using the training set, training of a policy network by using the training set, combined training of policy and value network by using the verification set, further optimizing network parameters by using the verification set, and model effect verification. However, this patent has problems such as lack of high quality sample data, lack of model processing, and combination of manual processing. Disclosure of Invention Based on the above, it is necessary to provide a method and a system for semi-automatically labeling remote sensing images to achieve remote sensing image labeling with high efficiency and high quality. In order to solve the technical problems, the invention provides the following technical scheme: In one aspect, the invention provides a semi-automatic labeling method for remote sensing images, which comprises the following steps: S1, preprocessing a remote sensing image to be processed, and constructing a remote sensing image slice sample set; S2, screening the remote sensing image slice sample set by adopting an active learning strategy based on a pre-training CNN model to obtain a remote sensing image slice; S3, extracting depth features of the remote sensing image slices by adopting a pre-training CNN model and carrying out multi-scale feature fusion to obtain a fusion feature map; S4, reconstructing the fusion feature map to obtain a pixel-feature matrix, predicting the pixel-feature matrix by adopting a pre-training random forest model to obtain the class probability of pixels in the remote sensing image slice, and labeling the remote sensing image slice according to the class probability of the pixels in the remote sensing image slice to