A coupled multi-task feature boosting method for remote sensing scene classification

The scene classification plays an essential role in processing very high resolution (VHR) images for understanding. The scene classification in remote sensing faces two difficulties: the mismatching features caused by the model overfitting problem and the semantic information losing problem. The multi-task method helps solve the problems by using the share weights of multiply tasks. We propose a feature boosting method with a multi-task framework that combines the scene classification task and the semantic segmentation task to overcome the difficulties. Different from the traditional multi-task learning method, the two tasks are coupled together via a weakly supervised learning method so that it does not require the labelled semantic segmentation samples. First, we proposed a weakly supervised segmentation method to create the interconnection of the segmentation task and the classification task. And we achieve a coarse segmentation result which is highly correlated to the classification by the weakly supervised method. Second, according to the surface distribution of remote sensing, we propose a sparse surface constraint to obtain fine segmentation results. Fine features are obtained by constraining the shared weights of the weakly supervised segmentation method. Last, we classify the scenes using the fine features and conduct experiments on the public remote sensing scene classification datasets. Experimental results demonstrate that the proposed coupled multi-task model outperforms the state-of-the-art methods on remote sensing scene classification.