Detecting solar farms with deep learning
- 1. Detecting Solar Farms with Deep Learning Jason Brown, Data Scientist FOSS4GNA 2019, April 16 astraea.earth/careers
- 2. Project Motivation • Fast growth in renewable energy industry including photovoltaic (PV) • More investment needed for 2°C Paris target: $130B by 2030 • Solar farm location, capacity, and date are not reliably recorded • EO data archives contain verifiable, spatially-correct, timestamped images of PV • Can we extract desired info from imagery with deep learning? 2 Source: BloombergNEF
- 3. Take Home Message 3 August 2018 Yes we can! Our deep learning solution can extract verifiable, spatially-correct, timestamped info about large PV arrays from archives of Earth- observation data.
- 4. Remote Sensing Imagery Sentinel 2 – Free, publicly available • Archive begins 2015 – Global coverage – 5 day revisit – Multispectral imagery • 13 channels in visible, near infrared (IR) & short wave IR – Spatial resolution: 10m & 20m – Top of atmosphere (L1C) 4 Sentinel-2 (10m)
- 5. Computer Vision Tasks Image credit ataspinar.com 5
- 7. Creating Training Examples • Lots of effort to create images and masks from annotated imagery • Many small decisions can impact model performance • 5,000 positive examples, 15,000 or more negative Positive: Negative: 7
- 8. Deep Learning Ensemble 8 • Classification – Inception v3 architecture – Tune for high recall, with minimum precision • Segmentation – U-net architecture – Tune for F1 score (balance of precision and recall) positive predictions 𝐹1 = 2 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 2 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝐹𝑎𝑙𝑠𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
- 9. Model Performance Array ≤ 0.1 km2 Array > 0.1 m2 Recall 0.799 0.892 F1 score 0.822 0.908 9 0.1 km2 equivalent to 316 m square or 100 pixels Model Performance Array ≤ 0.1 km2 Array > 0.1 m2 Recall 0.799 0.892 F1 score 0.822 0.908 9 0.1 km2 equivalent to 316 m square or 100 pixels
- 10. Model Inference • We conducted model inference on conterminous USA & China • Approx 4GB per 120 km2 • We processed approx 13TB – Access and prepare imagery – Infer on model pipeline – Process and write outputs • Fast enough to do as scenes are published 10
- 11. Demo 11
- 12. Roll Credits 12 Collaborators Dr. Kimberly Scott, VP of Data Science Courtney Whalen, Data Scientist Eric Culbertson, Data Scientist Earth on AWS QGIS SentinelHub Folium Dash Keras Scikit-learn GeoPandas rasterio Open Source & Open Data Used astraea.earth/careers