Multiscale Semantic Segmentation Networks for Forest Fire Mapping

Forest fires significantly threaten global ecosystem sustainability, making early detection crucial for reducing damage to flora and fauna. This work uses a deep semantic segmentation network to segment active forest fire areas based on satellite imagery. The chosen base network is SegNet, which will be integrated with a multiscale module in order to enhance the network's capability to detect fire spots of various scales. Two multiscale methodologies, spatial pyramid pooling and atrous spatial pyramid pooling (ASPP), are explored, along with a feedforward path option. As a result, the best performance of an intersection over union (IoU) of 0.6743 was obtained by an ASPP-embedded network with three parallel pathways. Similarly, the best variant of the SPP embedded network is also produced by three parallel pathways with an IoU performance of 0.6738. By having an accurate mapping of the forest fire spots, the firefighters can manage their operations better with optimal resource allocation to put out the fires as early as possible. The proposed method can be further improved by implementing a better base semantic segmentation model to encode the fire features accurately and efficiently.