Speaker
Description
Automated Deep Learning for Forest Resource Mapping to Drive Efficient Operational Planning
The efficient planning and execution of mechanised forest operations rely heavily on near-real-time, spatially explicit intelligence regarding forest resources. While the rapid growth of multi-source data—from national inventories to satellite imagery—presents a major opportunity to optimize these operations, it demands scalable, automated modelling workflows. To bridge this gap, we present an automated deep learning (DL) framework designed to generate actionable maps of forest resources for operational planning. The system ingests Sentinel-2 surface reflectance mosaics, Shuttle Radar Topography Mission (SRTM) terrain data, and climate variables, automatically optimising neural network architectures for two critical operational products: (i) precise estimation of key forest stand attributes—standing stock volume (Vol), basal area (BA), diameter at breast height (DBH), and mean height (H)—and (ii) high-resolution (10 m), national-scale tree species mapping.
Pairing harmonised field plots and inventory data from Romania and Finland with remote sensing predictors via Google Earth Engine, the workflow utilizes autoencoder-based bottleneck representations and Hyperband-driven hyperparameter tuning (via Keras Tuner) to compress high-dimensional inputs and automatically optimize network configurations. On independent test plots, the resulting models achieved R² values of 0.52 (Vol), 0.48 (BA), 0.72 (DBH), and 0.75 (H). Furthermore, the automated tuning strategy yielded a supervised bottleneck multilayer perceptron for species mapping that achieved macro F1 scores above 84% and overall accuracies of approximately 87%. By delivering wall-to-wall, consistently tuned maps of both forest structure and composition, this DL framework provides the foundational spatial data required for efficient forest engineering. The outputs directly support critical operational tasks, including optimized harvest planning, targeted machine and workforce allocation, and efficient road and transport logistics, while offering an extensible basis for operational forest digital twins.
| Keywords | Automation, DeepLearning, Mapping, Planning |
|---|
