Land cover classification of SAR images based on the fuzzy-TWDTW model

Accurate land cover classification is fundamental for environmental monitoring and resource management. Sentinel-1A Synthetic Aperture Radar (SAR) time-series imagery provides rich temporal and seasonal features, effectively enhancing classification accuracy. However, identifying and processing “fuzzy pixels” at type boundaries remain challenges due to high classification uncertainty. To address the above issues, this paper proposes a time weighted dynamic time warping method based on fuzzy membership (Fuzzy-TWDTW). This method first constructs a normalized Time Weighted Dynamic Time Warping (TWDTW) distance matrix to calculate the fuzzy membership. Subsequently, a confusion index calculation criterion is designed to quantify the classification uncertainty of each pixel by measuring the difference in membership degrees between the two most likely classes, and divide pixels into classified pixels and fuzzy pixels by setting a confusion index threshold. Finally, an ensemble-learning voting mechanism is adopted to determine the final type of fuzzy pixels. The results show that the Fuzzy-TWDTW method proposed in this paper has an overall accuracy of 97.71% and a Kappa coefficient of 0.96. Compared with the TWDTW method, the overall accuracy is improved by approximately 1% and the Kappa coefficient is increased by about 1.6, verifying the effectiveness of this method in complex time-series classification tasks.