Burned Forest Areas Mapping Using Semantic Segmentation Networks with Group and Shuffle Embeddings

Forests are one of the most important resources in a country. However, the number of forest fires is increasing worldwide, and more than 10 million acres of forest are lost every year. Therefore, it is important to have an effective automated forest management system that can monitor and predict the state of the forests in real time. Hence, remote sensing modality through satellite imaging has been used to monitor the state of the forests such as the affected areas due to the fire incidents. It will be very impractical to manually annotate the burned areas, which is the main driving factor for various development of semantic segmentation models for automated mapping process. This work introduces the group and shuffle embeddings to a symmetric encoder-decoder network to enhance forest fire mapping performance. The main task of these embeddings is to increase the variety of features that represent the burned forest areas, as well as forcing each group of filters to extract diverse sets of features. The proposed model, GaS-Net managed to deliver respectable performance with a mean intersection over union of 0.9664, whereby the embeddings are placed at the third block that produces 32 channels per group with a feature map size of 64 x 64 pixels. The model can be further enhanced by analyzing better configurations of the individual group and shuffling modules. This work provides significant insight to the search and rescue agencies to estimate the loss due to the forest fire disaster.