Photovoltaics Detection in Satellite Imagery using Deep Learning and Remote Sensing
de Weck, Olivier
MetadataShow full item record
The global push for renewables coupled with the steadily decreasing cost of each unit of solar energy produced has resulted in a dramatic rise in photovoltaic deployment both at a residential and commercial level. Currently, the global solar energy capacity doubles every 18 months, of which expanding solar energy facilities are the biggest contributor. The inherently decentralized nature of the deployment of photovoltaics has resulted in a dearth of reliable or verifiable information on solar energy generation both at a granular as well as a global scale. In parallel, there is an increase in the availability of high-resolution imagery from satellites and advancement in state-ofthe-art learning techniques. Together, this presents a unique opportunity to harvest previously scarce high-resolution satellite data and deploy state-of-the-art detection techniques for renewable energy applications. In this thesis, I propose, optimize, and validate several Deep Learning frameworks to detect and map residential as well as commercial solar installations. The best performing residential model achieved a precision of 96.9% and recall of 90.0%, comparable to the 93.1% precision and 88.5% recall achieved by the current state-of-the-art, DeepSolar. Notably, this was achieved with significantly reduced computational complexity - 89,000 trainable parameters compared to DeepSolar’s 21.8 million trainable parameters. A method is proposed for the extension of the custom trained CNN todifferent geographies at low cost, by using incremental training data. Performance of the model on a new geography is found to saturate with roughly 15 %incremental training data. Further, a study in resolution sensitivity showed that the optimal GSD for this problem lies in the range [0.3,0.7] m. For solar farms, a semantic segmentation neural network based model was trained on a dataset created by collecting satellite imagery of several major solar farms in the US and tested on images of solar farms unseen by the model. Objectively, the model achieved highly competitive performance indicators including a mean accuracy of 96.87%, and a Jaccard Index (intersection over union of classified pixels) score of 95.5%. Subjectively, it was found to detect spaces between panels producing a segmentation output better than human labeling. As a final step in the pipeline, a multi-step capacity evaluation model to estimate the number of panels and energy generation capacity of the detected solar energy facilities was proposed and generation capacities were compared against publicly available electricity generation data reported by various sources. The capacity evaluation model is the first of its kind, while deep learning applied specifically for the detection and mapping of solar farms is one of the first for the United States. Overall, this work fits into a longer-term goal of creating a granular global database of solar energy production which could serve as a single source of truth for industries and policymakers to inform decision-making. In the future, this approach could be used as a replacement for conventional sources of knowledge, or as a secondary source of intelligence for the cross-validation of reported figures.
DepartmentMassachusetts Institute of Technology. Center for Computational Science and Engineering
Massachusetts Institute of Technology