Method for constructing large-scale remote sensing image semantic segmentation model training sample set

Abstract

PaThe invention discloses a method for constructing a large-scale remote sensing image semantic segmentation model training sample set. The method comprises the steps: carrying out the registration of existing remote sensing image vector data and multi-stage remote sensing images, and carrying out the automatic cutting extraction of a primary sample set through a sliding window algorithm according to the pattern spot density; feature extraction is carried out on each image in the primary sample set, classification is carried out by adopting a clustering algorithm, samples with poor image quality are removed, and an intermediate sample set is obtained; and inputting the intermediate sample set into the semantic segmentation model in batches for iterative optimization training, predicting the samples after each iterative optimization is completed, and removing error samples in the intermediate sample set to obtain a target sample set. The method has the remarkable effects that the generation of a whole image and a mask which occupies a large space can be avoided, the sliding times of a sliding window are reduced, and the sample extraction speed and the data quality are improved; the purity of a correct sample in the sample set is improved, and the cost for manufacturing a large-scale sample set is greatly reduced.

Description

technical field [0001] The invention relates to the technical fields of remote sensing image feature extraction, remote sensing image semantic segmentation and sample set production, in particular to a method for constructing a large-scale remote sensing image semantic segmentation model training sample set. Background technique [0002] The successful launch of the Gaofen series satellites can provide PB-level remote sensing image data every year. At the same time, projects such as the "Third National Land Survey" and "Geographic and National Conditions Monitoring" use these remote sensing images to mark a large amount of vector data for ground features. If The effective use of these data forms a uniform format and a large number of sample sets, which can train a deep learning model with higher accuracy and stronger generalization ability, and provide important support for related projects and scientific research work. At present, there are two main methods for making remot...

AI Technical Summary

Problems solved by technology

[0005] Although the sample sets produced by these two methods have high accuracy, the first method requires a lot of manpower and material resources, and the efficiency is extremely low, making it difficult to build a million-level sample set; while in the second method, the vector It must be strictly corresponding to the remote sensing image, that is, the vector is drawn according to the image, but the strictly corresponding data is very limited, and it cannot meet the needs of large-scale sample set production. Due to factors such as changes and image quality, there are wrong samples caused by cloud coverage, shadow occlusion, etc. in the sample set obtained by automatic cutting, and the problem of inaccurate sample label accuracy caused by the incomplete match between the vector and the image

Therefore, neither of these two methods can be used for the production of large-scale sample sets. The first method will cause the production cost of the sample set to be too high, and the second method will make it difficult to guarantee the quality of the sample set.

Images