SciPy 2026

Motivation

Monitoring land use and land cover (LULC) change is essential for understanding deforestation, urban growth, and the effects of climate change. Satellite missions like ESA's Sentinel-2 provide openly available multispectral imagery at up to 10-meter resolution with global coverage and a revisit frequency of roughly 5 days. However, a single tile contains millions of pixels across multiple spectral bands, and producing accurate LULC maps requires preprocessing raw imagery (cloud removal, temporal compositing, index computation), training a classifier, and running inference across large regions. On a CPU, each of these stages can take tens of minutes per scene, making regional-scale analysis slow and difficult to iterate on.

Because each pixel in a satellite image can be processed independently, these raster operations are naturally parallelizable and well suited for GPU computation. This talk shows how existing Python tools in the geospatial ecosystem can be combined with GPU-accelerated libraries to make geospatial workflows significantly faster, often with minimal code changes.

Methodology

We present an end-to-end LULC classification pipeline built on publicly available datasets. Sentinel-2 Level-2A imagery provides the input features and ESA WorldCover provides per-pixel land cover labels, both accessed through the SpatioTemporal Asset Catalogs (STAC) specification. The pipeline covers several stages common to remote sensing workflows such as querying and loading cloud-hosted imagery into Xarray using Dask for chunked computation, masking clouds using the Sentinel-2 Scene Classification Layer, mosaicing overlapping tiles, computing an annual median composite, and deriving spectral indices such as NDVI and NDWI as additional features. The resulting data cube and matched labels are then used to train a Random Forest classifier on millions of labelled pixels, and the trained model is applied to previously unseen satellite tiles to generate LULC maps.

We compare wall times for each stage (preprocessing, training, inference) against a CPU baseline (using scikit-learn) to quantify the practical benefits and ease of using GPUs when working with data in the geospatial domain.

Results

Across the full pipeline, GPU-accelerated stages run 3x to 5x faster than their CPU equivalents, with the largest gains in model training and full-scene inference. The trained model performs especially well at distinguishing major land cover classes like built area and water bodies. We also present visual comparisons of model predictions against reference maps over held-out regions for qualitative analysis.

Conclusion

Attendees will come away with a practical understanding of how to efficiently leverage GPUs when working with geospatial data in Python. We will also discuss the design choices made, challenges we encountered, and potential improvements to provide a complete understanding to attendees which they can leverage in their own work. The full notebook for this exercise with detailed explanations for attendees to follow is available at https://docs.rapids.ai/deployment/stable/examples/lulc-classification-gpu/notebook/

Talk Outline (25 min + 5 min Q&A)

Introduction (5 min): What LULC classification is and why it matters for environmental monitoring, how satellite imagery is structured (tiles, bands, resolution, coordinate reference systems), and the datasets used in this talk (Sentinel-2 and ESA WorldCover).
Data access and preprocessing (10 min): Querying cloud-hosted imagery via STAC, loading into Xarray/Zarr, cloud masking, temporal compositing, and computing spectral indices (NDVI, NDWI).
Model training and inference (5 min): Building a feature cube from preprocessed imagery, training a Random Forest classifier on millions of labelled pixels, and generating LULC maps as inference over unseen tiles using the trained model.
Performance comparison and potential improvements(5 min): CPU vs GPU wall-time comparisons across preprocessing, training and inference. Discussion of problems like class imbalance approaches on how to solve these issues. A brief discussion about best practices for chunking and memory management.
Q&A (5 min)