Efficient Burn Region Detection Using Scale Estimation Networks

Abstract We introduce a novel approach for burn region detection using convolutional neural networks (CNNs), addressing the challenge of efficiently detecting burn areas across various scales. Traditional burn detection methods often suffer from high computational costs due to the need for either large-scale models or multi-scale testing. To mitigate this, we propose a Scale-aware Burn Region Detection (SABRD) framework, which explicitly handles scale variations by predicting the scale distribution of burn regions prior to detection. This prediction, obtained through a lightweight CNN, generates a scale histogram that guides the resizing of the image to focus on specific scales. As a result, burn regions across different scales are normalized, enabling more accurate detection using smaller, more efficient CNNs. Extensive evaluations on multiple burn datasets demonstrate that SABRD significantly reduces computational overhead while maintaining high detection accuracy.