Automated training sample definition for seasonal burned area mapping

Abstract Monitoring of environmental change can benefit from the increasing availability of multitemporal satellite imagery, and efficient and effective analysis tools are needed to generate relevant spatio-temporal land cover datasets. We present a data driven approach for automatic training sample selection to support supervised spatio-temporal mapping of seasonally burned areas in the semi-arid savannas of Southern Africa. Our approach leveraged the distinctive spectral-temporal trajectories associated with areas on the landscape burned at different times or areas remaining unburned over time. Using fuzzy c-means clustering, we extracted distinctive trajectories from the multitemporal mid-infrared burn index (MIRBI) data derived from Landsat data and characterized them based on empirically developed labeling rules. The selected training trajectories captured both the burn condition (burned or unburned) and if burned, the timeframe of the burn event. We assessed the approach by training a Random Forests model using over 2500 automatically selected training data and validated the model against ground truth for years 2009 and 2014. Based on over 1000 validation points in each year, we obtained overall accuracies above 90% showing reliable and consistent training data were supplied by our automatic training sample selection approach. The method provides a data driven and automatic approach which can reduce the time-consuming and expensive training task, enabling quicker generation of relevant burned area information that can support fire monitoring programs and climate change research.