Key points are not available for this paper at this time.
Aiming to address the limitations in accuracy and reliability of land use classification models when encountering complex and variable land use features, a novel method for land use classification (4D-SOMKS) is developed, which constructs a low redundancy four-dimensional (4D) spectral feature system with biophysical meaning of distinct land surface properties and organizes pixel-level features in a topology-preserving space using the Self-Organizing Map (SOM), further groups the SOM neurons into spectrally coherent clusters through K-means, and uses a small number of labeled samples only for semantic assignment of the resulting clusters rather than for pixel-level supervised model training. Empirical research in the Bayannur and Hong Lake Basins (HLB) have revealed the driving role of key spectral indices in classification. High overall classification accuracies were achieved, reaching 98.45% and 98.55% respectively, with robust performance across evaluation metrics including precision, recall, F1-score, and the Kappa coefficient. The results show that 4D-SOMKS achieves high accuracy and robustness while significantly reducing reliance on large-scale labeled data, providing an effective avenue to improve the accuracy and reliability of land use classification under spatiotemporal dynamic changes.
Wang et al. (Tue,) studied this question.