<![CDATA[Land cover maps obtained from satellite imagery are used in environmental management and spatial planning. Deep learning now outperforms traditional machine learning for this task, but Kolmogorov-Arnold Networks (KAN) have rarely been tested on multispectral remote sensing data. This paper evaluates two KAN strategies for classifying nine land cover types from Sentinel-2 imagery in Jambi, Indonesia. ResNet-KAN adds a KAN-based classifier head to a standard CNN backbone, while ConvKAN builds the entire network from KAN-based convolution layers. Both are compared against seven CNN, Transformer, and machine learning baselines using 23 spectral features with Google Dynamic World labels as reference, and ablation experiments test spectral feature composition, ImageNet transfer learning, and input patch size. Swin Transformer reaches the highest overall accuracy (88.34%), but ConvKAN better separates rare land cover classes like Grass and Shrub, achieving the best F1-Macro (0.5870) with only 2.91 million parameters, 89.4% fewer than Swin-T. Adding spectral indices raises ConvKAN F1-Macro by 13.8%, but lowers ViT accuracy by 3.19% OA because self-attention can already learn band-ratio operations from raw bands. KAN models also perform better when trained from scratch, because most Sentinel-2 channels fall outside the visible spectrum that ImageNet covers. Spatially, ConvKAN produces maps as clean as Swin Transformer despite being ten times smaller. KAN can therefore match larger models in accuracy and map quality for multispectral land cover classification.]]>
Copyrights © 2026
