Classification of Satellite Imagery
summary
Project uses Sentinel-2 10‑m RGB tiles plus OpenStreetMap/GeoPedia labels to classify agricultural vs non‑agricultural land, focusing on tulip fields. Workflow trains a ResNet-50 cloud filter (pretrained on ~40k Planet images plus 5k hand‑labeled Sentinel chips) to drop cloudy tiles, then runs U‑Net segmentation with heavy augmentation; final segmentation reaches about IoU 0.73 after ~24–28 epochs (training runs ~8–9 hours on a single GPU). Inference runs as a Apache Flink streaming pipeline (Beam Python SDK) that filters then segments live tiles.
AI-written summary · about
chapters (10)
- 0:00Introduction to talk
- 1:37Data acquisition demo
- 3:52Sentinel 2 satellite data
- 5:40Cloud filtering with ResNet
- 9:20Data augmentation for segmentation
- 10:25U-Net for tulip field segmentation
- 15:07Streaming pipeline with Apache Beam
- 20:32Live demo of pipeline
- 24:28Future work and use cases
- 31:22Q&A session
related talks
33mWhat and where? - Machine learning for geospatial image analysis
39mSatellite Imagery to Machine-Readable Intelligence
36mDeep Learning - Segmentation, Identification and Classification
21mLand Boundary Delineation using U-Nets
28mGeoAI: Evolution of AI-Based Tree/Vegetation Detection Models Using Imagery
21mUnderstanding World food economy with satellite
27mThe View from Above Getting Started with Satellite Images
01h 36mFrom Maps to Models: An End-to-End Journey in Geospatial Machine Learning with Python