Analyze
Data we collect, and data brought to us, analyzed with GIS pipelines, statistical methods, and deep-learning tools where they fit the task.
Data Analysis
The endpoint is a scientific result, not a data product — GIS and statistics, deep-learning segmentation and classification, and geospatial foundation models applied or fine-tuned to the question at hand. Every analysis ships as version-controlled code, with its inputs, parameters, and outputs written down so the run can be repeated.
GIS workflows
ArcGIS Pro and QGIS for spatial overlays, watershed analysis, change detection, terrain analysis, vector editing, and map production.
Statistical analysis
Time-series analysis of repeat surveys, uncertainty quantification, spatial statistics. Reproducible Python and R pipelines, version-controlled, delivered with raw data, code, and methods documentation.
Deep learning
Containerized deep-learning models, applied where they suit the task and benchmarked against classical baselines. Build recipes are version-controlled, so a run this year and a run next year produce the same numbers.
Imagery analysis
Object detection and segmentation in 2D imagery from any source — aerial, drone, satellite, handheld, and camera-trap stills. SAM 2 (with samgeo for georeferenced raster I/O) handles general-purpose promptable segmentation. DeepForest specializes in tree-crown detection from aerial RGB, and thermal-video models handle animal detection from drone surveys.
Remote-sensing foundation models
Pre-trained backbones for Sentinel-1/2, Landsat, NAIP, MODIS, and UAV RGB imagery — used for zero-shot scene classification, embedding-based similarity search, and transfer learning to project-specific tasks with small label budgets. Models: Clay, Prithvi-EO (IBM/NASA), Satlas (Allen AI), RemoteCLIP.
Point-cloud segmentation
Individual-tree instance segmentation from UAV and airborne lidar (SegmentAnyTree) and from ground-based lidar / TLS / MLS (TreeLearn), with classical methods (3DFin) as a baseline for stem detection and DBH. Forest structure is the most developed application; the same approach carries to other point-cloud segmentation problems.
Models beyond geospatial data
The same pipelines run models from other fields: seismic phase picking (SeisBench: PhaseNet, EQTransformer), LSTM rainfall-runoff prediction (NeuralHydrology), powder-XRD multi-phase identification (autoXRD), and genome-scale sequence modeling (Evo 2). Bring a model and data, or a problem you think a model could address.
What it runs on
Our standard workflows run on H100-class GPUs for model training and large inference jobs, with pipeline and processing workloads on AWS (EC2, Fargate, and Lambda). For larger jobs we use burstable GPUs on neoclouds. Everything runs in version-controlled containers, so a result can be reproduced on any of them.
Data labeling
Training labels for imagery and point-cloud projects are managed in the lab's hosted Label Studio instance; labeling workflows are covered in the lab documentation.
Have an analysis question?
Bring the research question — we'll scope the GIS, statistical, or deep-learning approach that fits it. Initial scoping conversations are free — see Access & Rates for how projects are priced.
Contact the Lab