Combining Handheld and Airborne Wildland-Urban Interface Field Measurements for AI Methods

Detecting wildfire ignitions and delineating impacted ecologies is of utmost importance for protecting life, property, and prioritizing Burned Area Emergency Recovery (BAER) response plans. In this presentation, we describe our field data collection methodology and how we applied modern artificial intelligence and machine learning techniques in support of federally-funded research. We used a combination of controlled burns, near-surface camera data, and drone flight data to extract hyperspectral near-infrared measurements of surface characteristics. We then manually label the observations in terms of vegetation type. After the data has been properly pre-processed and cleaned, we perform spectral classification and image super resolution tasks. For the spectral classification, we train three varieties of decision-tree-based algorithms on handheld data to then apply the best trained model to unlabeled flight data. We also use the drone data to condition a single image super resolution (SISR) neural network while also investigating the impact of using a network trained on pre-existing satellite (Sentinel-2) data.