Geospatial analysis for wildfire detection, fire spread prediction and risk assessment in natural landscapes

Abstract

Wildfires, are one of the most devastating natural-disasters, threatening ecosystems, human life, and infrastructure. Aerial-based or ground sensor-based conventional monitoring systems are subject to low resolution, high cost, and weak response times. With the introduction of unmanned aerial vehicles (UAVs) and low-weight edge single-board computer such as the Raspberry Pi, real-time wildfire detection in the field is now possible. This research presents a lightweight deep learning robust model (W-fire MobilenetV4 Small) for classifying fire, further segmenting fire pixels, and assessing risk built upon the FLAME dataset for UAV deployment. The model is a customization of MobileNetV4 conv small architecture, and for further segmentation, the same classification architecture was kept as an encoder while decoding in FPN style with depth-wise separable fusion blocks for efficiency. Both for the classification and segmentation tasks, keeping the model lightweight while efficient was the priority. Moreover, this was achieved with the help of a custom flame block, SE Attention, ECA, Sandglass block & DANet in the MobileNetV4 small architecture. Additionally, keeping the customized architecture as encoder FPN-style decoding was done for the segmentation task. The model is merged to fight against image distortions like haze, smoke, blur, night vision, cloud, noise, and other atmospheric conditions that drone images might face. This ultimately makes the model robust for different conditions. For the Segmentation task, the dataset had a huge class imbalance as fire regions were very small To effectively address the extreme class imbalance (fire-pixel ratio ≈ 0.000026), a novel Balanced Focal-Dice Loss is incorporated. Most significantly, on top of pure detection, the system also provides severity ranking of fire and propagation direction prediction, transforming raw segmentation outputs into vital information for firefighting missions. Our Novel model, W-Fire_MobileNetV4_Small achieved 79.32% accuracy, 79.33% F1-score, 79.33% precision, and 79.32% recall, outperforming the baseline MobileNetV4 Conv Small (72.66% accuracy, 70.73% F1-score). Further, Using this model as backbone our segmentation model achieved 0.6051 IoU Score, Loss of 0.2654, F1-Score 0.7504 and Recall 0.8638.