The BAMBI Dataset: Canopy and Deadwood Annotations for Nadir UAV-Recordings of Forest Wildlife (non-commercial) (Part 2)

This dataset is part of the BAMBI (Biodiversity Airborne Monitoring Based on Intelligent UAV sampling; https://www.bambi.eco/) project. It gives, per frame, tree cover (which pixels are canopy) and standing deadwood. Licence: CC BY-NC 4.0, unlike the rest of the BAMBI dataset. Tree cover comes from Restor's TCD model and deadwood from deadtrees.earth. Both are built on NVIDIA's SegFormer, released under the NVIDIA Source Code License, which permits use "non-commercially, meaning for research or evaluation purposes only". The deadwood case is not obvious from its own repository: deadtrees.earth release under MIT and so does segmentation_models_pytorch, but the MiT encoder inside that MIT package carries an explicit NVIDIA carve-out. The remaining environment classes -- snow, water, road, grass, rock, bare ground, roof, vehicle -- are free of that restriction and are published separately under CC-BY-4.0, so that they are not needlessly encumbered. Tree cover is inferred at 10 cm/px and deadwood at 5 cm/px, the resolutions those models were trained for; the frames are resampled accordingly. Tree cover is a per-pixel canopy mask, not individual crowns: at 10 cm inference its effective resolution is about 40 cm on the ground. This is an annotation layer, not a standalone dataset. It carries no imagery. The recordings it annotates are the processed BAMBI Dataset (10.5281/zenodo.21874826 and its sibling parts), which has to be downloaded alongside it. download_from_zenodo.py --version environment fetches both in one step and places them in the same directory. Experimental. These annotations are machine-generated and have not been reviewed by hand. BAMBI carries no ground truth for these classes, so no accuracy figures are given and none should be inferred. Treat them as a starting point to be checked, not as ground truth. The per-frame masks are the raw model output and are not smoothed. On frames that are one material edge to edge -- a flat snowfield, a fog whiteout -- detection is unstable and the masks flicker between adjacent frames. Each flight therefore ships a smoothed coverage series and three flags: undetermined for frames whose luminance dynamic range is too low to classify at all, unstable for a class that disagrees with its temporal neighbourhood, and unreliable_classes for a class that flips presence too often across the flight. Nothing is filled in and no mask pixel is invented, so a reader who ignores the flags will see the flicker. Masks are stored as COCO run-length encoding over the full 1024x1024 frame. Frames are 16:9 content letterboxed into that square; the black bands are excluded from both inference and the reported coverage. You can find scripts for working with the dataset at https://github.com/bambi-eco/Dataset.