Kolmogorov-Arnold Networks beat MLPs in land classification with fewer parameters
New study shows KANs match random forests on satellite imagery with 5-class accuracy
A new arXiv paper from Katherine L. Bauer, Teemu Harkonen, Simo Sarkka, and Arturo Sanchez-Azofeifa tests Kolmogorov-Arnold Networks (KANs) as an alternative to traditional machine learning models for multispectral land classification. Using Landsat 8 satellite imagery from Edmonton, Alberta for training, the researchers compared KANs against random forests and multilayer perceptrons (MLPs), then evaluated all models on an independent dataset from Calgary, Alberta to test spatial generalization. The task covered five land classes: agriculture, urban, water, forest, and bare ground.
Results show the KAN matched random forest accuracy on the Calgary dataset and outperformed the MLP, while requiring substantially fewer trainable parameters and providing greater interpretability. This suggests KANs offer a compact, transparent alternative for satellite-based land monitoring, potentially reducing compute costs and making model decisions easier to audit in applications like environmental tracking and urban planning.
- KAN matched random forest accuracy and beat MLPs on the independent Calgary test set
- Trained on Edmonton data, evaluated on Calgary data to test spatial independence across 5 land classes
- Achieved results with substantially fewer trainable parameters and greater interpretability
Why It Matters
More efficient, interpretable AI for environmental monitoring and urban planning from satellite data.