Accelerating Geospatial Analysis with GPUs
summary
An end-to-end pipeline loads Sentinel‑2 tiles and ESA WorldCover labels, preprocesses clouds and overlaps into a median 94.34GB → ~200MB, 240 chunks Zarr dataset, trains a cuML random forest on a single GPU, and runs multi‑GPU Dask CUDA preprocessing on RAPIDS Xarray cuML Dask with 8 Tesla V100 GPUs, 3–5× speedup reported. Results show crisp water/land segmentation and common mislabels in dense urban shadows (lower Manhattan), with notebooks and timings included.
AI-written summary · about
chapters (16)
- 0:00Geospatial data importance
- 3:59Satellite imagery concepts
- 7:11Geospatial workflow challenges
- 10:45NVIDIA Rapids for GPU
- 13:14Data acquisition with STAC
- 16:27Pre-processing on GPUs
- 19:54Land use classification
- 22:46Model training with RAPIDS
- 25:35GPU vs CPU performance
- 27:00Key takeaways and future
- 28:23Q&A: Shadows and time
- 29:06Q&A: CUDA tiling
- 30:09Q&A: Cost and budget
- 31:35Q&A: GPU config
- 33:00Q&A: Programming overhead
- 34:45Q&A: Beyond pixel classification
related talks
01h 03mWorkshop: Intro to GPU data science with RAPIDS
39mAccelerating Data Science with RAPIDS
01h 36mRAPIDS: GPU-Accelerated Data Science for PyData Users
30mDask ❤️ Xarray: Geoscience at Massive Scale
31mOpenDataCube - Petabyte Scale Satellite Imagery Processing in Python
24mTiling & Parallel Processing of Large Images
37mVisualizing Large Datasets with JavaScript Using deck.gl
39mDensity! The not-so-secret trend driving the future of Data Analytics