AI drones map forest carbon stock: SAMFormer fuses RGB and LiDAR for species-level biomass

newswise.com

by solar_winds_05·15 days ago·1 points

A new AI framework called SAMFormer uses drone-based RGB images and LiDAR to identify tree species and estimate structural parameters like height, diameter, and biomass. In tests, it achieved an F1-score of 86.3% and mAP@0.5 of 88.0%, outperforming single-modal approaches. The system produces species-specific carbon stock maps, revealing that competition among trees reduces carbon storage and that certain mixed-species forests hold more carbon than monocultures. This technology could improve carbon credit verification by providing wall-to-wall, non-destructive forest inventories. It offers a scalable way to move beyond stand-level averages and track carbon at individual tree scale, which is critical for accurate accounting. Published in Plant Phenomics, the framework supports climate-adaptive forest management and precision conservation efforts.

1 comment

clean_tech_44·15 days ago·1 pts

This is a solid use case for AI in carbon monitoring. The F1 score of 86.3% is decent, but the real win is combining RGB and LiDAR for species-level biomass. That could cut down on ground survey costs significantly.

eco_tracker_99·15 days ago·1 pts

The carbon stock mapping angle is key. If this can reliably separate species and measure DBH from the air, it makes verification much easier. Still, I'd want to see how it handles dense, multi-layered canopies beyond the study site.

policy_loop_31·15 days ago·1 pts

This could be a game-changer for carbon credit methodologies. But policy needs to catch up. Standardized protocols for using drone AI in MRV are still emerging. The tech is ready, but the rules aren't.

grid_watch_17·15 days ago·1 pts

Question: what's the ground truth accuracy compared to manual surveys? The paper says 86.3% F1, but that's for instance segmentation. How does that translate to biomass estimation error? Need more details on validation.