Fig. 1

Proposed solar PV mapping pipeline. Given a small set of point labels and its corresponding Sentinel 2 imagery, pixels are clustered into multiple clusters (64 for our experiments). These clusters are merge into a user defined smaller set of classes (three in this example) using a linear classifier. Cluster merge results are shown in a web tool where a human user provides feedback on which pixels belong to the solar farms class or to the other background classes and the linear classifier if finetuned based on the feedback from the user. This weakly supervised segmentation process is represented at the top of this figure and is interactively performed to obtain weak semantic labels like the example shown at the top right of the figure. These labels paired with the corresponding geo-located Sentinel 2 image are used to create a semantic segmentation dataset suitable for supervised training of a solar farm semantic segmentation model. The obtained segmentation neural network can be used to perform inference for solar farms in novel scenes as shown at the bottom of the figure. False positive predictions are considered hard negatives and are used to augment the training dataset and finetune the supervised segmentation neural network improving its false positive rate. This process of performing inference in novel scenes, adding hard negative to the training set and finetuning the supervised model further can be repeated multiple times until the performance of the results is good enough for large scale inference.