Category Correlation and Adaptive Knowledge Distillation for Compact Cloud Detection in Remote Sensing Images

Cloud detection relying on deep convolutional neural networks (DCNNs) obtains remarkable accuracy gains at the expense of high computation and storage costs, which are difficult to deploy to resource-constrained devices, such as intelligent satellites. Recently, knowledge distillation (KD) has been a promising solution for a compact model. However, most existing KD methods only transfer the feature relationship of pairwise pixel, which fails to cope with thin clouds and cloud-like objects in complex scenes. Furthermore, those KD methods directly imitate the output of the complicated model regardless of the correctness. In this article, we propose a novel category correlation and adaptive KD (CAKD) framework for the lightweight cloud detection network. We design a category relational context (CRC) module to refine the structured pixel-category correlation from the teacher and student network. Then, we perform the category correlation distillation (CCD) to make the student model better address the intraclass consistency and the interclass difference, thus reducing the category confusion. Besides, a pixel-adaptive distillation (PAD) module is utilized to adaptively transfer the soft-output knowledge of a teacher model by extracting the teacher’s pixel prediction probability. Extensive experiments on Landsat 8, Landsat 7, Gaofen-2, Gaofen-1, and Google Earth dataset report the effectiveness and universality of our distillation method. The CAKD allows MobileNetV2 with 2.31M parameters and 4.63G float-point operations (FLOPs) to outperform advanced cloud detection methods without the added inference overhead.