Spectral-Spatial Optimization of Deep Networks: A Frequency-Domain VGG Framework For Wildfire Detection

Abstract

Wildfires pose an escalating global threat, inflicting severe ecological devastation and substantial economic liability annually. While conventional spatial-domain deep learning frameworks have advanced automated fire detection, their efficacy remains heavily constrained by environmental noise, variable illumination, and atmospheric occlusion. This paper addresses these vulnerabilities by introducing a novel frequency-domain deep learning paradigm for predictive wildfire classification. Utilizing the Fire and No Fire Images (FFNI) from the DeepFire benchmark dataset, it deploys the 2D Fast Fourier Transform (FFT) to map spatial visual data into spectral matrices, thereby isolating robust structural textures and high-frequency pixel transition dynamics. The design rigorously evaluate a series of customized Convolutional Neural Network (CNN) topologies, specifically benchmarking standard VGG16 architectures against structurally optimized variants tailored with localized receptive fields, adaptive filter counts, and modified kernel sizes. Quantitative empirical evaluations demonstrate that the proposed FFT-driven modified VGG16 framework achieves a classification accuracy of 94.0% and an optimized Area Under the Receiver Operating Characteristic (AUC-ROC), substantially outperforming the baseline spatial-domain VGG16 configuration which achieved 62.0%. These findings establish that spectral feature extraction via FFT effectively mitigates ambient spatial noise, offering a computationally efficient, highly resilient trajectory for real-time computer vision deployments in disaster mitigation frameworks.