Abstract: Multilabel learning is highly conducive to the accurate classification of remote sensing (RS) images that contain rich feature information, and the deep-learning-based multilabel classification of RS images has recently gained popularity owing to its exceptional performance. However, the categorical annotation of training samples in large-scale datasets is a time-consuming and labor-intensive task whose quality cannot be guaranteed, leading to the potential occurrence of missing labels. Incomplete label information can mislead the model training process, thereby negatively impacting classification performance. To address this problem, an incomplete multilabel classification method for RS images based on local semantics (LS) is proposed in this paper. First, a semantic representation learning (SRL) module is used to extract deep features from input images, which are then used to generate pseudo-labels. Next, a local semantic learning module is deployed to capture local semantic correlations within the features. These correlations are weighted and subsequently used to enhance the pseudo-labels. Finally, a weighted combination of pseudo-labels and enhanced pseudo-labels is used as the final predicted label, and the network is optimizes using the cross-entropy loss function. The experimental results demonstrate that the proposed method outperforms existing multilabel classification methods with different missing label rates.

Key words: remote sensing images, multilabel classification, incomplete multilabel, local semantics (LS)