MTCSNet: Mean Teachers Cross-Supervision Network for Semi-Supervised Cloud Detection

Cloud detection methods based on deep learning depend large and reliable training datasets to achieve high accuracy. There will be a significant impact their performance, however when the data are insufficient or label quality is low. Thus, alleviate this problem, semi-supervised cloud method, named mean teacher cross-supervision network (MTCSNet) proposed. This method enforces both consistency accuracy two student branches, which perturbed with different initializations, for same input image. For each of respective used generate high-quality pseudo labels, constructed using an exponential moving average (EMA). A one-hot label, produced by one branch, supervises other branch standard cross-entropy loss, vice versa. To incorporate additional prior information into model, presented uses near-infrared bands instead red as model inputs injects strong augmentations unlabeled images fed model. induces learn richer representations ensure constraints predictions image across batches. attain more refined equilibrium between supervised loss in process, proposed learns optimal weights homoscedastic uncertainty, thus effectively exploiting advantages tasks elevating overall performance. Experiments SPARCS GF1-WHU public show that outperforms several state-of-the-art algorithms only limited number labeled samples available.