LANDMINE RECOGNITION USING A KOLMOGOROV-ARNOLD NEURAL NETWORK BASED ON MAGNETIC SENSING DATA

Background . Landmine contamination remains a critical issue for more than 60 countries, including Ukraine, where the recovery of agriculture and infrastructure is hampered by hidden explosive devices. The authors propose a passive land-mine recognition approach that combines measurements of magnetic anomalies obtained with an FLC-100 sensor and a Kolmogorov-Arnold Network (KAN), delivering high accuracy while minimising the risk of detonation. Materials and Methods. A baseline data set of 338 real measurements (sensor voltage V, sensor height H, soil type S) was balanced by generating 50 synthetic records for every “soil–mine” pair using a parameterised normal distribution. After normalising V and applying one-hot encoding to S and the mine classes M, a three-dimensional feature space was formed. Two KAN architectures were evaluated: KAN (3-16-16-4) and KAN (3-64-64-4), both employing cubic B-splines to achieve high-precision mine recognition (> 95 % accuracy). Training was conducted in PyKAN (PyTorch backend) for 65 epochs with a fixed spline grid (k = 3, m = 10) using the Adam optimiser. Results and Discussion. The KAN (3-16-16-4) model achieved an accuracy of 93.56 % without overfitting; the main confusion occurred between the “anti-personnel” and “booby-trap” classes. Increasing the number of neurons in each hidden layer to 64 raised the accuracy to 95.59 % and eliminated the erroneous assignment of “anti-personnel” mines to the “booby-trap” class. Both networks perfectly distinguished the “no-mine” and “anti-tank” cases, confirming the robustness of spline activations to sensor noise. Conclusion. The computer experiment shows that a Kolmogorov–Arnold neural network with cubic B-spline weights provides robust recognition of different mine types (“no mine”, “anti-tank mine”, “anti-personnel mine”, “booby-trap”) using magnetic-field sensor data (10⁻¹⁰ Tesla) with accuracy exceeding 95 %. Interpretable spline weights allow the contribution of each feature to be analysed, ensuring high sensitivity to small anomalies and demonstrating the scalability of KAN. Keywords : Kolmogorov–Arnold network, fluxgate magnetic sensor, passive mine detection