Deep Chromatic Feature Engineering For High-Accuracy Wildfire Prediction Using Enhanced Convolutional Neural Network Architectures

Abstract

Wildfires constitute a major global environmental hazard, causing extensive ecological damage, economic losses, infrastructure disruption, and threats to public safety. Although deep learning has substantially improved automated wildfire image classification, conventional spatial-domain models may be affected by illumination changes, atmospheric interference, shadows, and background clutter. This chapter presents a chromatic feature-engineering framework for wildfire classification based on the YUV color space and enhanced Convolutional Neural Network (CNN) architectures. Images from the Fire and No Fire Images (FFNI) subset of the DeepFire benchmark are spatially standardized and transformed from RGB into YUV representations. The luminance (Y) and chrominance (U and V) components are analyzed as complementary representations of image structure and color information. A baseline LeNet-5 model is compared with a structurally modified LeNet-5 designed to improve the extraction of discriminative features from chromatic representations. The reported experiments show that the modified LeNet-5 achieves 96.0% classification accuracy on the YUV features, compared with 44.0% for the baseline LeNet-5. These findings indicate that an appropriate combination of chromatic representation and architecture design can substantially improve wildfire image classification. The chapter also reviews related deep learning approaches, describes the dataset and preprocessing pipeline, formulates the YUV transformation mathematically, and discusses the experimental implications of the proposed approach.