Solar Panel Detection Demo

In this video, Federico walks through his entire solar-panel detection pipeline — from collecting and hand-annotating satellite imagery, splitting the data into train/val/test sets with location-based metadata, training segmentation models in PyTorch, and finally deploying a serverless web app on AWS so anyone can run predictions. PROJECT REPOS • Capstone code & notebooks – https://github.com/FederCO23/UCSD_MLBootcamp_Capstone • QGIS Serval raster-editing plugin – https://github.com/lutraconsulting/serval 🚀 TIMELINE 00:00 Welcome & project goals 00:20 Dataset creation — sourcing 272 GeoTIFF tiles, AutoMask script, hand-labelling with Serval 11:29 Model selection & training — UNet, FPN, UNet++ & PSPNet (EfficientNet encoders) 14:29 Baselines — random, threshold and logistic-regression classifiers 18:46 Data-centric tricks — bicubic up-scaling, Real-ESRGAN super-resolution & heavy augmentation 22:42 Cloud deployment on AWS Step Functions 30:31 Live demo — web app inference on Brazilian PV farms & rooftops 35:17 Lessons learned, cost-savers (tile caching, S2 grid) & future work KEY TAKE-AWAYS • Data first – meticulous tiling and NIR-based pre-masking cut labelling time from weeks to days. • Fine-tuning beats training from scratch – ImageNet-pretrained encoders slashed compute while hitting production-grade accuracy. • Lean serverless stack – Lambda handles I/O; GPU Batch does heavy lifting, keeping costs predictable. • Modular design – the four-stage pipeline (fetch → enhance → predict → report) lets you swap in better super-res models or caching without rewrites. Enjoy the episode? Like, Subscribe, and share your questions below!