A novel-based Kolmogorov-Arnold Networks for high accuracy mapping of water and vegetation: a study of the global main rivers

Accurate mapping of water and vegetation is crucial for environmental monitoring, resource management, and land use planning. This study presents the application of a novel deep learning architecture, Kolmogorov-Arnold Networks (KANs), featuring learnable activation functions, for a pixel-based classification of water and vegetation using Sentinel-2 imagery. A dataset from 2023 observations, consisting of nine major global rivers-Amazon, Karun, Ganges, Mississippi, Nile, Rhine, Seine, Shatt-al-Arab, and Yangtze-was used to evaluate the performance of KANs against a baseline multi-layer perceptron (MLP) model. The results indicate that KANs achieved a general kappa coefficient of 0.9949, comparable to the MLP's 0.995, while offering a more compact architecture. The adaptive B-spline activation functions in KANs demonstrated particular effectiveness in classifying mixed and boundary pixels, areas where conventional algorithms often struggle. A sensitivity analysis further revealed KANs' robustness, achieving 99.52% accuracy with only Red and Near-Infrared bands' information, which highlights their potential for effective classification with limited spectral information. While KANs did not consistently outperform MLPs across all regions, their comparable accuracy, enhanced interpretability, and efficient training make them a promising alternative for remote sensing applications.