SEMANTIC SEGMENTATION OF ENHANCED LANDFORM MAPS USING HIGH RESOLUTION SATELLITE IMAGES

High resolution fuel maps are useful for high resolution wildfire simulations and detection of hazards on the landscape. In general, high resolution Enhanced Lifeform Maps (ELMs) are used in conjunction with other data layers to create these fuel maps. However, ELMs are costly to make with substantial manual editing involved. In response, this study uses deep learning-based semantic segmentation models to generate 5-m resolution ELMs (14 classes) in Marin and San Mateo, California using high resolution remote sensing datasets. ELM classes were found to be severely imbalanced, leading to model overfitting. Sample weighted loss functions helped alleviate this issue to an extent. High resolution ELMs are bound to be more valuable with the growing fire risk and landscape heterogeneity, particularly near the wildland urban interface. All codes, future updates, and further details can be found at https://github.com/minhokim93/elm_mapping.