Generation of Synthetic Wildfire Smoke Images with Generative Adversarial Networks

Abstract

The scarcity of annotated images significantly hin ders the development of robust deep learning models for early wildfire smoke detection. Traditional augmentation methods, such as rotation or mirroring, are often insufficient. This is particularly true for detecting subtle smoke formations at long distances, where smoke is often barely perceptible even to human observers. Existing smoke datasets predominantly feature devel oped smoke plumes or closer views, making them unsuitable for training models for this critical early-phase detection. To address this, we propose generating synthetic images using Generative Adversarial Networks. Unlike typical GAN applications that aim for high-fidelity object representation, our objective is different. We synthesize realistic, fuzzy images of subtle, distant smoke — blurred and blended with the background — yet retaining characteristic features essential for classifier training. We propose a GAN architecture based on a modified Super-Resolution GAN, specifically adapted without B residual blocks, in order to produce realistic images of smoke at long distances. Experimental evaluation demonstrates that augmenting datasets with GAN generated smoke images significantly improves the performance of classifiers in detecting early-stage wildfire smoke, affirming the utility of GANs for data enhancement even when generating low quality, realistic imagery. This method mitigates data scarcity, offering a viable solution for training effective early wildfire detection systems.

Keywords