Cross-Temporal Self-Supervised Learning With Superpixel Mask for Multitemporal Land Cover Classification

Multitemporal remote sensing images offer multiple views of Earth observation, enabling the integration of temporal information to effectively mitigate issues caused by clouds, shadows, seasonal changes, and atmospheric reflection. This integration significantly reduces the uncertainty and noise of images, facilitating a more comprehensive and accurate extraction of essential ground object features. However, most land cover classification models based on deep learning focus on interpreting single-temporal remote sensing images, and multitemporal remote sensing datasets are relatively scarce, leading to challenges in model training. To address these issues, this article proposes a cross-temporal self-supervised learning method with superpixel mask (CTS2L-sm) for multitemporal land cover classification. Specifically, we design a superpixel-based masking strategy to replace the fixed-size patch masking strategy used in masked image modeling (MIM), which prevents the complete occlusion of small-scale remote sensing scenes and more effectively preserves structural information. To further enhance the integration of multitemporal information, we introduce a cross-temporal attention module (CTAM) in both forward and backward to extract cross-temporal features. The resulting multitemporal denoised homogeneous features, when fused with each temporal feature, significantly improve classification accuracy. CTS2L-sm combines these denoised homogeneous features extracted by CTAM with single-temporal features and simultaneously reconstructs the original remote sensing images across all temporal, extending the MIM strategy from single-temporal to multitemporal self-supervised training. Experimental results on the WUSU and FUSU datasets demonstrate the effectiveness and superiority of the proposed CTS2L-sm compared to several state-of-the-art methods.