DilAtSE-Net: An Encoder Decoder Network for Burnt Area Delineation

Laalenthika Konthalapalli, Sarvesh Kumar, Advika Srivastava, Ananya Das, Ananya Shriram, Ishita Choudhary

Keywords: Environmental Science, Wildfires, Deep Learning

Abstract: Burned area delineation following forest wildfires plays a critical part in quantification for disaster management, post-disaster assessment, and restoration planning. In recent years, advances in deep learning and computer vision, coupled with the growing availability of Earth observation and remote sensing datasets, have significantly contributed to progress in addressing this problem. However, burned area delineation from satellite imagery faces several interrelated challenges that dampen the effectiveness of current approaches. Severe class imbalance causes models to exhibit majority-class bias and poor recall. Additionally, region-based loss functions optimize overall overlap rather than edge accuracy, often over-looking precise boundary detection. This study introduces an encoder-decoder based architecture, inspired by the conventional U-Net, that incorporates dilated convolution blocks for increased field of view, squeeze-and-excitation layers for dynamic channel weighting, and a self-attention bottleneck, trained with a combination of losses. The proposed DilAtSE-Net model outperforms methods previously benchmarked by achieving higher mean Intersection over Union (mIoU) and comparable Dice scores on the Wildfire-CEMS dataset with 12.8M parameters compared to existing benchmarks ranging from 31M to 64.1M parameters.

Submission Number: 25