Precision-Recall curves for different detection methods over all the Grand Est region. The probabilistic part-based method (orange) significantly boosts performance.

Humans naturally identify fine-grained objects like biodigesters by first spotting potential candidates, then confirming them by checking for characteristic parts—such as tanks and piles. Following this intuition, we define the detection confidence based on the presence of these parts within the detected bio-digester.

With only a limited annotation starting point and vast unlabeled territory, bio-digester detection models suffer from high false positive rates, since digesters show visual similarity to industrial zones, oil storage areas, or farms. Hard negative mining tackles this by identifying high-confidence false positive detections via human inspection and adding these hard examples back into the training set.

Detection results in Marne, Grand Est, and Bretagne sub-regions. Results are split between in-database detections (existing ground truth) and overall detections, including new discoveries. Precision and recall are computed using confirmed matches; overall counts include detections outside the reference database.

Our system was able to detect new sites outside of the reference training set and generalize to new regions, such as Bretagne, showing transferability and discovery capability.

Power Production (kW) against Tank Area (m²) in Bretagne region (r² = 0.332).

Estimating power production over a large area provides valuable insight into potential methane leaks and atmospheric impact. Leveraging the generated bio-digester inventory of Bretagne, we estimate its aggregated production in kilowatts (kW), for which ground truth is available. We build a linear regression model that predicts power production based on the overall detected tank area at each site. While the model exhibits a large predictive error, it provides reasonable aggregated estimates over many observations.