Remote Sensing Bio‑Digester Dataset
收藏资源简介:
Remote Sensing Bio‑Digester Dataset Overview To the best of our knowledge, this is the first large-scale satellite dataset of bio‑digesters, with facility‑level and part‑level annotations. It comprises high‑resolution aerial and satellite imagery of bio‑digester sites across France’s Grand Est region, drawn from multiple sources. Ground‑truth labels are geolocated segmentation masks for three classes: Whole installation (entire bio‑digester site) Anaerobic digestion tanks (internal tanks) Biomass piles (feedstock storage areas) Code available On Github Composition Training & validation sets include SPOT, Sentinel, and aerial modalities. Training is made easy using MMRotate as annotations are also available in DOTA format. Resolutions: 0.5 m, 1.5 m, and 5 m per pixel. Modalities: Aerial, SPOT, and Sentinel imagery—to study transferability and resolution impact. Labels remain geolocated; due to modality mismatches, some labels transfer poorly, so aerial annotations serve as the definitive reference. Coordinates provided in EPSG:2154. Directory Structure ```├── README.md├── res_0.5│ ├── image│ │ ├── BDORTHO│ │ │ ├── train│ │ │ └── val│ │ ├── SENTINEL│ │ │ ├── train│ │ │ └── val│ │ └── SPOT│ │ ├── train│ │ └── val│ ├── label│ │ ├── train│ │ └── val│ └── vpv│ ├── train│ └── val├── res_1.5│ ├── image│ │ ├── BDORTHO│ │ │ ├── train│ │ │ └── val│ │ ├── SENTINEL│ │ │ ├── train│ │ │ └── val│ │ └── SPOT│ │ ├── train│ │ └── val│ ├── label│ │ ├── train│ │ └── val│ ├── test│ │ ├── image│ │ ├── label│ │ └── meta│ └── vpv│ ├── train│ └── val└── res_5 ├── image│ │ ├── BDORTHO│ │ │ ├── train│ │ │ └── val│ │ ├── SENTINEL│ │ │ ├── train│ │ │ └── val│ │ └── SPOT│ │ ├── train│ │ └── val ├── label │ ├── train │ └── val └── vpv ├── train └── val ``` Testing Test Split Covers the entire French department of Marne, pre‑tiled at 1.5 m resolution in 1000×1000 px tiles (~5000 tiles). Environment & Modalities Available only in aerial and SPOT modalities at 1.5 m resolution.A precomputed tiling facilitates large‑scale evaluation without engineering overhead. Objective & Metrics - Detect all bio‑digester sites within 200 m accuracy (AP@200 m as introduced in the paper). - Augmented with newly discovered Marne‑region digesters to better estimate precision at scale: - SPOT: 29 sites - Aerial: 27 sites (including 4 under construction) Visualization Optionally, use vpv to explore the dataset:```vpv ac aw nw ./res_1.5/image/BDORTHO/train/ svg:./res_1.5/vpv/train/ \ nw ./res_1.5/image/PLEIADES/train/ svg:./res_1.5/vpv/train/ \ nw ./res_1.5/image/SPOT/train/ svg:./res_1.5/vpv/train/ \ nw ./res_1.5/image/SENTINEL/train/ svg:./res_1.5/vpv/train/```



