SPLC: An Attention-Based End-to-End Superpixels and Land Cover Classification Jointly Learning Architecture for HR Remote Sensing Image

High-resolution (HR) remote sensing images typically have abundant textual details, various land cover (LC) categories bridged by very complex contour boundaries, which often bring challenges for generating superpixels (SPs) and predicting LC classification. In this work, an end-to-end architecture is presented to jointly learn SPs and LC classification (SPLC), whereby LC classification is boosted with accurate boundaries that are resulted from our learned SPs, in addition, better SPs can be generated by investigating more semantic-constrained features that are learned with LC classification task. More specifically, based on a convolutional neural network (CNN)-based backbone, two individual heads for generating SPs and predicting LC category are appended with a weighted loss adopted for our jointly learning, and an attention mechanism for considering the global context of SPs is further employed to improve LC classification. The experimental results show that our method outperforms other LC classification methods on both the ISPRS Vaihingen 2-D dataset, the overall accuracy (OA) reaches 88.97%. Furthermore, when compared to existing SPs methods, our approach can produce better SPs with more precise boundary localization and higher accuracy.