A cloud detection method with domain adaptation enhanced generalization capability

Objectives: Due to the differences in spectral range, imaging conditions and imaging time of remote sensing sensors, there are usually color and spectral differences between remote sensing images obtained by different sensors. Therefore, it is difficult to directly apply the model pretrained with open-source datasets for the cloud detection task on current target images. Methods: In this paper, a domain adaptive cloud detection algorithm is proposed, which realizes the spectral mapping between the target remote sensing images and the images of existing training data sets, and improves the generalization ability and robustness of the pretrained deep learning cloud detection model in practical applications. Specifically, the domain adaptation based cloud detection framework in this paper uses a convolutional neural network based cloud detection model as the prediction algorithm, and CycleGAN as the spectral domain transfer between images in the source domain and the target domain. Results: The cross-domain experiments among GaoFen2 (GF2), Landsat7, landsat8 have proved the effectiveness and advance of the proposed method. Conclusions can be obtained on three different datasets: 1) When the test dataset and training dataset are obtained from the same sensor, deep learning-based cloud detection algorithms obtain better results than the traditional cloud detection algorithms. 2) When cloud detection models trained on images from the source dataset are directly used to detect test images of the target dataset, the accuracy decreases obviously. 3) When the test images are pre-transformed by the introduced domain transfer algorithm, the accuracy of the cloud detection results is significantly improved and much better than the traditional methods. In addition, we discovered that when the spectral range of source training images covers the spectral range of test images, the accuracy of the cloud detection results decreases only slightly, but in the opposite case, the accuracy of the cloud detection results decreases significantly. Conclusions： Domain transfer can map images between the test dataset and the training dataset at the pixel level, so that the distributions of them are similar both in spectral and spatial space. By introducing domain adaptation, the deep learning-based cloud detection method can make full use of the pre-trained models from existing datasets, which greatly reduces the demand on preparation of target labels.