Drones and remote sensing can provide observations of forests at scale, but this raw data needs to be interpreted to further our scientific understanding and inform effective management decisions. This thesis studies two problems under the realistic constraint of limited domain-specific training data: tree detection for understanding carbon sequestration and vegetation mapping for forest fire mitigation.

For tree detection, we process the drone data using structure from motion and then register it to remote sensing imagery. Then, we compare different strategies for using a deep learning detector with these modalities and limited training data. For vegetation mapping, we show that we can localize fuel that causes forest fires using image-based semantic segmentation trained on very few examples and LiDAR-based geometric reasoning. Finally, we introduce RAPTORS, a novel algorithm that plans where to collect sparse drone observations based on existing remote sensing data. We show that training a remote sensing-based vegetation classification model on observations from RAPTORS is more effective at identifying rare classes than training on observations from a coverage-based approach. Overall, these experiments show how using machine learning, data harmonization across scales, and intelligent sampling can facilitate automated forest understanding with limited training data.