Land-Cover Classification With Time-Series Remote Sensing Images by Complete Extraction of Multiscale Timing Dependence

The evolution rules of different land covers show some similarity, which poses a huge challenge for high-accuracy land-cover classification. Therefore, the full extraction of multiscale timing-dependence features is important for mining seasonal and phenological change laws and improving the accuracy of time-series land-cover classification. However, traditional methods are often unable to fully detect the global and local change information generated during the evolution of land covers, resulting in incomplete timing-dependence features being extracted and a low classification accuracy. The Informer network can fully capture the long-term dependence of a time series, thereby improving its classification accuracy. Therefore, we propose a high-accuracy land-cover classification method with the Informer network. First, we continuously shorten the length of the series so that the ProbSparse self-attention mechanism can consider timing dependencies on multiscale, and then we can obtain the features of the local important moments. Second, we calculate the correlation between the important moments and the other moments, as well as the correlation among each moment, to fully utilize the local and global time-dependent features of the land-cover time series. Third, we add a fully connected batch normalization module in order to use all the extracted timing dependence for classification. Finally, the proposed model is compared with traditional models on two datasets: for the reorganized BreizhCrops dataset, it achieved a performance similar to long short-term memory; for the TiSeLaC dataset, it achieved an F1-score of 96.011%, which is 0.33% higher than the second-best model.