<![CDATA[Forest fires pose a significant threat, requiring advanced detection systems. Conventional deep learning models, such as Convolutional Neural Networks (CNNs), are often limited by fixed activation functions that struggle to model the complex, irregular visual patterns of fire. This architectural rigidity presents a research gap for more adaptive neural architectures. This study addresses this gap by proposing and evaluating a novel method for forest fire classification using a Convolutional Kolmogorov-Arnold Network (CKAN), an architecture featuring learnable activation functions to improve detection accuracy and flexibility. Following a systematic machine learning lifecycle, this research utilized a public Kaggle dataset of 14,063 'Fire' and 'Nofire' images. Extensive data augmentation was applied to enhance model robustness. We designed a hybrid CKAN model combining a CNN feature extractor with a KAN module that uses learnable B-spline activation functions for classification. The model was trained for 30 epochs with the AdamW optimizer and Binary Cross-Entropy loss, followed by a rigorous evaluation on an unseen test set. The proposed CKAN model demonstrated exceptional performance, achieving 98.04% accuracy and an AUC-ROC of 0.9955, significantly outperforming conventional architectures. Grad-CAM analysis confirmed that the model focused on relevant visual features of fire and smoke, thereby validating its decision-making process. The findings establish the CKAN architecture as a highly effective and computationally efficient approach for forest fire classification, making it a powerful and promising solution for deployment in real-world, resource-constrained environmental monitoring systems.]]>
Copyrights © 2026
