Enhancing the Binary Classification of Wildfire Smoke Through Vision-Language Models

Forest fires pose a significant threat to ecosystems, biodiversity, and human communities. As fires can get out of control very quickly, early detection of forest fires is imperative for saving the lives of not only humans but also wildlife. The detection of smoke, which is detectable earlier than fire, thus becomes a good target for prediction, as it allows officials to react and contain wildfires quickly. Historically, there has been much research on the smoke classification task on smoke datasets. However, many of these datasets are unrealistic and cannot be applied to real-world scenarios for a variety of reasons, including unrealistic camera angles that cannot be obtained in real life. Therefore, to create a more applicable algorithm, we focus on the detection of smoke in forest images in the form of a video series utilizing the FIgLib dataset. Traditionally, deep learning models such as CNNs and LSTMs have been used in image sequence classification tasks. In this paper, we utilize a novel classification approach through multimodal vision language models (VLMs) such as LLaVA and demonstrate that its performance is more generalizable, interpretable, and is also less data-intensive than traditional methods and on par with state-of-the-art methodologies.