Deep Learning for Forest Resource Mapping from Sub-Meter Resolution Imagery: Technical Insights and Methodologies

Deep learning has transformed numerous fields so far, propelled by improved algorithms and increased data accessibility. Remote sensing, in particular, offers significant opportunities for vision applications with direct socio-ecological implications. Forests represent a major environmental component, offering essential functions including climate regulation, biodiversity preservation, and interaction with living creatures. Departing from studies that reveal forest patterns using coarse resolution imagery or structural sensors like LiDAR, this thesis explores sub-meter resolution imagery - human-interpretable and highly detailed visual data covering the Earth. Forest attributes of different types are investigated, varying from tree characteristics to forest height and biomass. Technically, this thesis starts with a vanilla setup of indomain semantic segmentation, delves into the regression of structural attributes, and ends with cross-domain adaptation. This brings insights into the learning capacity of vision models in the context of natural scene image understanding. The first part of the thesis introduces a deep learning framework to count, locate, and estimate the height of individual trees from aerial images at the national scale. An attention UNet is utilized to delineate individual tree crowns and count trees with point supervision. Tree height is estimated from optical imagery by learning a mapping from visual cues to canopy heights projected from LiDAR point clouds. Results are compared against field data to assess the practical values of the framework as supplementary data to support digitized national forest management. This study highlights the capacity of deep learning in characterizing visually interpretable forest and tree structures. Yet, challenges persist in learning more complex patterns from the optical data. This motivates the second study on forest biomass at stand level, a structural measure of forests typically collected on the ground. At larger scales, predominating methods apply statistical or machine learning models on multispectral imagery, often complemented by height data and calibrated with field data. To our knowledge, we are the first to demonstrate that stand-level biomass can be directly learned from sub-meter resolution RGB imagery, rich in forest and tree details, using convolutional neural networks and field data. This provides avenues for efficient and accurate quantification of forest biomass, a critical indicator of forest resources that supports nature preservation and carbon-neutral commitments. The first two studies employ deep learning systems to quantify forest attributes given a specific dataset, which follows the principal indomain assumption of machine learning. Yet, degraded performance is often observed when applied to out-of-distribution data, a common scenario in practice. The third study aims to address the domain shift issue, exploring whether deep learning models trained on one dataset can be quickly adapted to new datasets with marginal efforts. We release a new dataset consisting of sub-meter resolution optical imagery collected in five countries and assess the cross-domain adaptability of various image-level regression tasks, including tree cover, total tree count, and average canopy height. By enforcing ordered embedding space during training, models are effectively prepared for later adaptation in source-free low-shot setups. Overall, this thesis introduces a collection of deep learning systems tailored for forest resource mapping with depth into technical and applied insights, contributing to sustainable management efforts for a greener future.