Multi-source remote sensing image classification method based on robust deep semantic segmentation network

Abstract

The invention discloses a multi-source remote sensing image classification method based on a robust deep semantic segmentation network (Robust Loss Function of Remote Sensing Imagery, RSRLF). The method is formed by recombining a fault tolerance loss function and an adaptive category weight. Remote sensing semantic segmentation is carried out based on the open source land cover classification dataset so that sample annotation work can be greatly reduced, but the open source land cover classification data set contains a certain annotation error sample. The method is advantaged in that the fault tolerance loss function can inhibit the model from learning the noise label, and avoids the over-fitting of the noise label by the model; and a category equalization constraint module can solve problems of large scale difference of globally distributed samples of ground object categories and inconsistent confusion degree among the categories; combination of the two can effectively solve a problem that noise labels and category samples are unbalanced when remote sensing semantic segmentation is carried out by using an open source land cover classification data set, and ground object classification precision based on an open source land cover classification data set is improved.

Description

technical field [0001] The invention belongs to the intersection field of remote sensing interpretation and artificial intelligence, and relates to a multi-source remote sensing image classification method based on a robust deep semantic segmentation network, specifically including fault-tolerant deep learning and self-adaptive class balance weight collaborative learning remote sensing semantic segmentation robust Great way to learn. Background technique [0002] Changes in land use and land cover directly affect climate and biodiversity worldwide. Rapid and effective acquisition of real and reliable land cover data can provide important support for global climate change research, land cover change detection, and ecological model establishment. With the continuous improvement of the spatial resolution of remote sensing images, the resolution of large-scale land cover classification mapping is also continuously improved. How to quickly and accurately obtain land cover class...

AI Technical Summary

Problems solved by technology

[0005] The present invention mainly solves the problem of low accuracy of classification results existing in the process of classification of remote sensing images based on open source land cover classification data in the prior art, and provides a multi-source remote sensing image classification method based on robust deep semantic segmentation network (Robust Loss Function of Remote Sensing Imagery, RSRLF), which consists of a fault-tolerant loss function and adaptive category equalization weights, can effectively improve the classification accuracy of remote sensing images based on open source land cover classification datasets

Noisy labels can adversely affect model learning

Images