Kolmogorov-Arnold Networks for Spatially Independent Multispectral Land Classification

Katherine L. Bauer; Teemu Harkonen; Simo Sarkka; Arturo Sanchez-Azofeifa

cs.CVAug 13, 2026

Kolmogorov-Arnold Networks for Spatially Independent Multispectral Land Classification

Authors: Katherine L. Bauer, Teemu Harkonen, Simo Sarkka, Arturo Sanchez-Azofeifa

Organizations: Department of Earth and Atmospheric Sciences, University of Alberta, Edmonton, AB, Canada · Department of Electrical Engineering and Automation, Aalto University, Helsinki, Finland

Abstract

Land classification from satellite imagery is important for land management, environmental monitoring, and urban planning. Machine learning methods such as random forests and multilayer perceptrons have shown strong performance on multispectral data, while the Kolmogorov-Arnold network has emerged as an alternative architecture with compact model structures. This study evaluates the Kolmogorov-Arnold network for land classification using Landsat 8 imagery and compares it with random forest and multilayer perceptron models. The models were trained and tested on data from Edmonton, Alberta and evaluated on an independent dataset from Calgary, Alberta across five land classes: agriculture, urban, water, forest, and bare ground. For the Calgary dataset, the Kolmogorov-Arnold network matched the accuracy of the random forest and outperformed the multilayer perceptron, while requiring substantially fewer trainable parameters and providing greater interpretability.

Figures & tables

Fig. 1: Landsat 8 OLI Level-2 imagery of Edmonton (acquired July 21, 2022) and Calgary (acquired July 24, 2022), displayed in UTM Zone 12N as a true colour composition [ 21 ] .

Fig. 2: Average reflectance and standard deviation of each Landsat 8 band for each respective class.

Fig. 3: Examples of Landsat 8 OLI (left) data points with the Google Satellite overlay (right) for each LULC class.

Panel of Figure 4 — Fig. 4: Loss curves for the normalized KAN (top) and indices KAN (bottom) models show a smooth decline with minimal separation between training and validation loss, indicating no evidence of overfitting.

Panel of Figure 5 — Fig. 5: Confusion matrices for each model on the stratified Edmonton test set, where diagonal elements represent correct predictions and off-diagonal elements indicate misclassifications.

Normalized Bands
Model	City	OA (%)	$\text{CI}_{95}$ (%)	F1
RF	Edmonton	84.4	[80.8, 88.0]	0.78
MLP	Edmonton	98.7	[97.6, 99.8]	0.98
KAN	Edmonton	81.0	[77.1, 84.9]	0.73
RF	Calgary	97.3	[96.0, 98.6]	0.97
MLP	Calgary	94.7	[92.9, 96.5]	0.95

TABLE I: Classification performance across models, data types, and cities. The highest values of each data type and city are bolded. 95% confidence intervals are computed using the stratified proportion estimator with n=385 (Edmonton) and n=600 (Calgary).

Normalized Bands
Model	City	Agriculture	Urban	Water	Forest	Bare
RF	Edmonton	0.87	0.42	0.90	0.95	0.28
MLP	Edmonton	1.00	0.93	0.92	0.99	0.94
KAN	Edmonton	0.15	0.86	0.47	0.76	0.92
RF	Calgary	0.97	0.98	0.93	0.89	0.98
MLP	Calgary	0.99	0.98	0.81	0.76	0.98

TABLE II: Per-class intersection over union across models, data types, and cities. Bolded values show the highest IoU per data type and city.

Fig. 6: Classified Edmonton and Calgary using the different KAN models: Normalized KAN (Edmonton) (a), Indices KAN (Edmonton) (b), Normalized KAN (Calgary) (c), Indices KAN (Calgary) (d).

Fig. 7: Heat map illustrating the generalizability of each model across Edmonton and Calgary datasets.

Fig. 8: A parameter analysis to evaluate each model’s architectural efficiency during training and validation on the Edmonton dataset using the spectral indices feature set.

Panel of Figure 11 — Fig. 9: Feature attribution comparison of each model for the normalized dataset (left column) and the indices dataset (right column).

Panel of Figure 12 — Fig. 10: Architectures of the RF, MLP, and KAN models used in this study. The KAN visualization highlights learned spline activations and their relative importance as indicated by edge darkness, the MLP displays all node connections, and the RF shows a subset of 10 trees for clarity.

Explore similar work

Improved Knowledge Distillation for Land-Use Image Classification
Jun 12, 2026Arundhuti Sur, Abhiroop Chatterjee, Susmita Ghosh +1Cross-Architecture Knowledge Distillation Knowledge Distillation
Jadavpur University, India · Rochester Institute of Technology, USA
Vision Transformers and Convolutional Neural Networks for Land Use Scene Classification
May 20, 2026Arun D. KulkarniRemote Sensing Image Understanding Vision Transformer
Computer Science Department, University of Texas at Tyler, Tyler, TX 75799, USA
When Less Is More: A Controlled Benchmark of Lightweight CNNs for Satellite Land-Cover Segmentation on DeepGlobe
Jul 25, 2026Atiq Ur Rehman, Joseph Michael DonovanRemote Sensing Image Segmentation Land Cover Classification
Department of Economics & Decision Sciences, University of South Dakota, Vermillion, SD, USA · University of South Dakota, Vermillion, USA

Jun 12, 2026J/K move · Enter open · S save

Improved Knowledge Distillation for Land-Use Image Classification

Arundhuti Sur, Abhiroop Chatterjee, Susmita Ghosh, Emmett Ientilucci

Jadavpur University, India · Rochester Institute of Technology, USA

Cross-Architecture Knowledge Distillation Knowledge Distillation Neural Network Compression Image Classification

Open paper AskPDF arXiv

In the present article, an improved Knowledge Distillation (KD) framework has been proposed for efficient compression of deep convolutional neural networks for land-use image classification task. Motivated by the need to achieve competitive classification accuracy while reducing computational complexity, a teacher-student learning paradigm is adopted in which a VGG16 network transfers knowledge to a lightweight MobileNetV2 model. The proposed framework integrates hard supervision from ground truth labels with a soft supervision strategy that combines Kullback-Leibler divergence and Cosine Similarity losses. Experiments conducted on three land-use datasets show that the proposed KD-based method yields improved performance, and achieves an accuracy of 99.04%, outperforming both baseline student training and single-loss distillation approaches, while retaining substantial model compression.

Abstract

Explore similar work

Improved Knowledge Distillation for Land-Use Image Classification

Vision Transformers and Convolutional Neural Networks for Land Use Scene Classification

When Less Is More: A Controlled Benchmark of Lightweight CNNs for Satellite Land-Cover Segmentation on DeepGlobe