CCD-TBLCA: A Land Cover Continuous Change Detection Method Fusing Transformer and Bi-Directional LSTM via Cross-Attention

Continuous land cover changes in remote sensing observations can provide crucial data support for evaluating urban spatial planning and sustainable development, but their acquisition remains challenging because it must overcome missing observations caused by clouds and cloud shadows while accurately characterizing the spatio-temporal dynamics of land cover semantics. Therefore, we proposed a Continuous Change Detection method fusing Transformer and Bi-directional LSTM (BiLSTM) via Cross-Attention (CCD-TBLCA) in land cover using Sentinel-2 imagery. Specifically, it employed a dual-branch encoding design for spectral time-series data, enabling deep information fusion of long-term dependencies and local dynamics. Firstly, we used the optical temporal information from raw irregular time-series, achieving monthly mapping from spectral anomalies to land cover change events based on CCD-TBLCA. Subsequently, we proposed a breakpoint detection criterion that enforced semantic consistency before and after changes, thereby forming a more comprehensive framework for evaluating breakpoint detection alongside spatial existence and temporal accuracy. Across seven urban regions, CCD-TBLCA outperformed existing methods in both temporal classification and breakpoint detection, with semantic-consistent breakpoint detection accuracy generally improving by over 15%. In spatial mapping, CCD-TBLCA achieved F1 accuracy above 0.9 for most land cover types, surpassing existing products. CCD-TBLCA’s robustness to sequence updates and irregular time-series had also been validated. The results show that CCD-TBLCA provided in this study can accurately describe continuous spatio-temporal land cover changes and help to understand how human activities reshape land cover patterns.