Methane emissions are responsible for at least one-third of the total anthropogenic climate forcing and current estimations expect a significant increase in these emissions in the next decade. Consequently, methane offers a unique opportunity to mitigate climate change while addressing energy supply problems. From the five primary methane sources, residual water treatment provided 7% of the emissions in 2010. This ratio will undoubtedly increase with global population growth. Therefore, locating sources of methane emissions is a crucial step in characterizing the current distribution of GHG better. Nevertheless, there is a lack of comprehensive global and uniform databases to bind those emissions to concrete sources and there is no automatic method to accurately locate sparse human infrastructures such as wastewater treatment plants (WWTPs). WWTP detection is an open problem posing many obstacles due to the lack of freely accessible high-resolution imagery, and the variety of real-world morphologies and sizes. In this work, we tackle this state-of-the-art complex problem and go one step forward by trying to infer capacity using one end-to-end Deep Learning architecture and multi-modal remote sensing data. This goal has a groundbreaking potential impact, as it could help estimate mapped methane emissions for improving emission inventories and future scenarios prediction. We will address the problem as a combination of two parallel inference exercises by proposing a novel network to combine multimodal data based on the hypothesis that the location and the capacity can be inferred based on characteristics such as the plant situation, size, morphology, and proximity to water bodies or population centers. We explore technical documentation and literature to develop these hypotheses and validate their soundness with data analysis. To validate the architecture and the hypotheses, we develop a model and a dataset in parallel with a series of ablation tests. The process is facilitated by an automatic pipeline, also developed in this work, to create datasets and validate models leveraging those datasets. We test the best-obtained model at scale on a mosaic composed of satellite imagery covering the region of Catalonia. The goal is to find plants not previously labeled but present in wastewater treatment plant (WWTP) databases and to compare the distribution and magnitude of the inferred capacity with the ground truth. Results show that we can achieve state-of-the-art results by locating more than half of the labeled plants with the same precision ratio and by only using orthophotos from multispectral imagery. Moreover, we demonstrate that additional data sources related to water basins and population are valuable resources that the model can exploit to infer WWTP capacity. During the process, we also demonstrate the benefit of using negative instances to train our model and the impact of using an appropriate loss function such as Dice's loss.
