Semi-Supervised Cloud Detection for Remote Sensing Imagery via Self-Training

Abstract

Most existing deep learning-based cloud detection methods rely on high-quality pixel-level annotations, but acquiring them is prohibitively expensive. Some researchers have attempted to supplement annotated data with unannotated data by referring to the self-training paradigm. This kind of methods often involves an iterative "training-predicting" process, i.e. training on annotated data, then making predictions on unannotated data and using them as pseudo annotations for retraining. However, the lack of sufficient initial annotated images results in that there is a lot of noise in the pseudo-annotations, which is not conducive to the training of subsequent models. In this paper, we propose a novel self-training cloud detection framework with an Otsu thresholding-based multi-stage pseudo annotation filtering strategy, which can automatically filter pseudo annotations in the iterative process without extra supervision and train a robust model. In addition, a new loss function is specially designed for this method, which further improves the model performance by making better use of filtered pseudo annotations. Experimental results on GF-1 wide-field satellite images demonstrate that the proposed method obtains an overall accuracy of 96.16% and an mIoU of 88.84%, both outperforming other semi-supervised methods.