| 内容提要: |
In recent years, several advanced convolutional neural net-work (CNN) models have been proposed that can successfully detect clouds in multispectral satellite images. These networks include classical super-pixel classification, typical end-to-end CNN-based frameworks, and multi-scale/level feature fusion networks. However, such networks rely heavily on large number of samples annotated at pixel-level for parameter tuning, and sample collection is tedious, time-consuming, and expensive. To reduce the labeling cost, we propose an unsupervised domain adaptation (UDA) approach to enable the model trained on labeled source satellite images to generalize to unlabeled target satellite images. Specifically, we propose a fine-grained feature alignment (FGFA) domain adaptation strategy that encourages a cloud detection network to extract domain-invariant representations, which improves the accuracy of cloud detection in unlabeled target satellite images. The proposed FGFA strategy consists of two steps: i) class-relevant feature selection based on an attention-guided mechanism and ii) class-relevant feature alignment based on a proposed grouped feature alignment approach. Experimental results on the “Landsat-8 → ZY-3” and “GF-1→ ZY-3” domain adaptation tasks demonstrate the effectiveness of our method and its superiority to existing state-of-the-art UDA approaches. |