# Fossett Laboratory for Virtual Planetary Exploration

> The Fossett Laboratory at Washington University in St. Louis is shared research infrastructure. It collects environmental and planetary data with drones, laser scanners, GNSS receivers, satellite imagery, and other methods, processes that data through automated pipelines into finished products, and provides compute and model pipelines so researchers spend their time on scientific questions rather than workflows. The lab is building toward field science that runs continuously: observation that arrives on its own, analysis that keeps pace, and a scientist's attention drawn to what is scientifically significant. This is grounded in services across the research lifecycle (Collect → Process → Analyze → Share): drone surveys, terrestrial laser scanning, GNSS surveying, remote sensing, automated processing, GIS and statistical analysis, deep-learning analysis and geospatial foundation models, extended reality (XR) visualization, and an open STAC catalog of research-ready, machine-accessible datasets. The lab also provides hands-on training and works with academic, public, nonprofit, and industry partners. Research spans geology, ecology, archaeology, environmental science, and planetary surfaces.

## Pages

- [Home](https://fossettlab.org/): Overview of the lab — what it collects and delivers, the compute and model pipelines available for work beyond geospatial science, the four lifecycle services, and how to start a project.
- [Vision](https://fossettlab.org/vision): Where the lab is headed — moving field science from occasional observations toward continuous monitoring; the six commitments that shape it (analysis that doesn't stop, a queryable record, AI as a collaborator, reproducibility by default, rapid iteration, discovery is our measure).
- [What We Do](https://fossettlab.org/services): Overview of the research lifecycle — the Field Data → Cloud Storage → Automated Processing → AI Analysis → Visualization → Research Products → Publication pipeline — with links to the four stage pages and to Access & Rates. Legacy anchors #collection, #analysis, #visualization, #access still land on matching sections.
- [Collect](https://fossettlab.org/collect): Field data collection — drone surveys (RGB, multispectral, lidar, thermal), terrestrial laser scanning, GNSS surveying, robotic total station, ground-penetrating radar, and coordinated commercial-satellite tasking (WashU clients only). Includes the field equipment fleet (drones, sensors, survey and positioning, connectivity) and the collaborator upload portal.
- [Process](https://fossettlab.org/process): Photogrammetric reconstruction (SfM/MVS), orthomosaic generation, multispectral and thermal indices, and point cloud processing, run through automated, version-controlled pipelines. Deliverables cloud-native by default (COG, COPC, GeoPackage); published outputs surface in the data catalog. Collaborators can process drone imagery through the lab's hosted WebODM.
- [Analyze](https://fossettlab.org/analyze): GIS workflows, statistical analysis, and containerized deep-learning models — imagery analysis, remote-sensing foundation models, point-cloud segmentation, and models from fields beyond geospatial science. Names the compute the work runs on. Data labeling via the lab's hosted Label Studio.
- [Share](https://fossettlab.org/share): Publication maps and figures, browser-based 3D viewers, Gaussian splats, XR experiences (Meta Quest 3, HoloLens 2), interactive web maps, and the open data catalog with STAC API.
- [Access & Rates](https://fossettlab.org/access): The lab operates as shared research infrastructure, offered as a recharge facility. Free initial scoping, internal (WashU recharge) vs external rates, and written estimates under a User Agreement provided at project scoping. WashU users can access the internal rate schedule (WashU login required).
- [Data](https://fossettlab.org/data): Browsable catalog of geospatial datasets (imagery, point clouds, radar, 3D models) organized by field site. Includes interactive map, dataset previews, programmatic access via STAC, a quickstart guide with Python and R examples at #stac-quickstart, and a per-dataset "Cite" expander on each row exposing a stable STAC URL plus plain-text and BibTeX citations.
- [Models](https://fossettlab.org/models): Searchable catalog of the scientific models and methods the lab can run, with a page per model. Filterable by research area, data type, method (foundation model, deep learning, classical) and license. Covers remote sensing, forestry, biology, geophysics, hydrology, planetary science, materials and weather. Models not listed can usually be added on request.
- [People](https://fossettlab.org/people): Lab director, staff, and affiliated faculty.
- [About](https://fossettlab.org/about): Lab mission, history, and institutional support.
- [Media](https://fossettlab.org/press): Press coverage and media mentions.
- [Contact](https://fossettlab.org/contact): Contact form for inquiries about equipment, services, and collaboration. Includes FAQ.

## Models

Every model the lab can run, one page each at https://fossettlab.org/models. Listed once the lab can run it. Each entry gives what the model does, its research area, method type, the data it takes, and the license that binds a user — some carry non-commercial weights, share-alike terms, or a supplier's own agreement rather than an open-source license. Models not listed can usually be added on request.

- [3DFin](https://fossettlab.org/models/three-dfin): Measures stem diameter, height, and taper from terrestrial scans without machine learning (forestry vegetation; classical; point-cloud). License: GPL-3.0.
- [ADAF](https://fossettlab.org/models/adaf): Detects archaeological features such as barrows, ringforts and enclosures in lidar terrain models (geomorphology; deep learning; imagery). License: Apache-2.0 code, CC-BY-SA-4.0 weights.
- [AFwizard](https://fossettlab.org/models/afwizard): Filters ground points from lidar with a different filter applied to each terrain segment (geomorphology; classical; point-cloud). License: MIT.
- [AIFS](https://fossettlab.org/models/aifs): Produces global medium-range weather forecasts at about 31 km resolution, six-hourly (weather climate; deep learning; gridded-fields). License: Apache-2.0.
- [AlphaFold 3](https://fossettlab.org/models/alphafold3): Predicts the structure of proteins and their complexes with ligands and nucleic acids (biology; deep learning; sequence). License: CC-BY-NC-SA-4.0.
- [AMS3D crown segmentation](https://fossettlab.org/models/ams3d-crownseg): Delineates tree crowns in drone lidar using adaptive mean-shift clustering (forestry vegetation; classical; point-cloud). License: GPL-3.0-or-later.
- [autoXRD](https://fossettlab.org/models/xrd-classifier): Identifies mineral phases in powder X-ray diffraction patterns (materials mineralogy; deep learning; spectra). License: MIT.
- [BiRefNet](https://fossettlab.org/models/birefnet): Produces one high-resolution foreground mask per image, holding edges on thin structures (cross domain; deep learning; imagery). License: MIT.
- [Boltz-2](https://fossettlab.org/models/boltz): Predicts structures of molecular complexes and estimates binding affinity (biology; deep learning; sequence). License: MIT.
- [Chai-1](https://fossettlab.org/models/chai-1): Predicts structures of proteins and their complexes (biology; deep learning; sequence). License: Apache-2.0.
- [Clay](https://fossettlab.org/models/clay): Creates embeddings from satellite imagery that can be used for classification and other downstream tasks with relatively few labeled examples (remote sensing; foundation model; imagery). License: Apache-2.0.
- [CLEAN](https://fossettlab.org/models/clean): Predicts enzyme function (EC numbers) from protein sequence (biology; deep learning; sequence). License: Research use only.
- [Concerto](https://fossettlab.org/models/concerto): Creates point-cloud embeddings learned jointly from 3D scans and 2D imagery (cross domain; foundation model; point-cloud, imagery). License: Apache-2.0 code, CC-BY-NC-4.0 weights.
- [CorrDiff](https://fossettlab.org/models/corrdiff): Downscales ERA5 reanalysis to kilometre-scale fields over Europe (weather climate; deep learning; gridded-fields). License: Apache-2.0.
- [Crater Detection](https://fossettlab.org/models/crater-detection): Detects lunar craters and matches them against a catalog to fix absolute position (planetary; deep learning; imagery). License: MIT.
- [CROMA](https://fossettlab.org/models/croma): Creates embeddings from paired radar and optical satellite imagery (remote sensing; foundation model; imagery). License: MIT.
- [CrossEarth](https://fossettlab.org/models/crossearth): Segments satellite imagery in regions and conditions it was not trained on (remote sensing; foundation model; imagery). License: MIT.
- [DBloops](https://fossettlab.org/models/dbloops): Measures grain size and maps boulders from 3D point clouds of rock surfaces (geomorphology; classical; point-cloud). License: GPL-3.0.
- [DeepForest](https://fossettlab.org/models/deepforest): Detects individual tree crowns in aerial RGB imagery (forestry vegetation; deep learning; imagery). License: MIT.
- [Depth Anything 3](https://fossettlab.org/models/depth-anything-3): Recovers depth, multi-view geometry and camera pose from ordinary photographs (cross domain; foundation model; imagery). License: Apache-2.0.
- [DINOv3](https://fossettlab.org/models/dinov3): Extracts dense visual features from ordinary photographs, used as a frozen backbone for downstream tasks (cross domain; foundation model; imagery). License: LicenseRef-DINOv3 (not open source).
- [DINOv3-SAT](https://fossettlab.org/models/dinov3-sat): Extracts dense visual features from aerial and satellite RGB imagery, used as a frozen backbone for downstream tasks (remote sensing; foundation model; imagery). License: LicenseRef-DINOv3 (not open source).
- [DNABERT-S](https://fossettlab.org/models/dnabert-s): Creates embeddings from DNA sequences for comparing genomes, identifying functional regions, and other genomic analyses (biology; foundation model; sequence). License: Apache-2.0.
- [DOFA](https://fossettlab.org/models/dofa): Handles imagery with any number of spectral bands using a single model, and returns embeddings (remote sensing; foundation model; imagery). License: CC-BY-4.0.
- [DOFA-CLIP](https://fossettlab.org/models/dofa-clip): Scores multispectral or RGB imagery against text descriptions, without task-specific training (remote sensing; foundation model; imagery). License: CC-BY-NC-4.0.
- [ESM C](https://fossettlab.org/models/esm-c): Creates protein-sequence embeddings for tasks such as function and property prediction (biology; foundation model; sequence). License: MIT.
- [Evo 2](https://fossettlab.org/models/evo2): Models genomic sequences and can be used to score sequence variation and predict the effects of genetic variants (biology; foundation model; sequence). License: Apache-2.0.
- [ForestFormer3D](https://fossettlab.org/models/forestformer3d): Segments forest point clouds into individual trees and semantic classes in a single pass (forestry vegetation; deep learning; point-cloud). License: CC-BY-NC-4.0.
- [FourCastNet 3](https://fossettlab.org/models/fourcastnet3): Produces global weather forecasts at 0.25-degree resolution, generating an ensemble rather than a single trajectory (weather climate; deep learning; gridded-fields). License: Apache-2.0.
- [FSCT](https://fossettlab.org/models/fsct): Segments stems in terrestrial scans and fits cylinders to measure diameter (forestry vegetation; deep learning; point-cloud). License: GPL-3.0.
- [GeoCLIP](https://fossettlab.org/models/geoclip): Estimates where in the world a photograph was taken (cross domain; foundation model; imagery). License: MIT.
- [GeoLG-3DFaultNet](https://fossettlab.org/models/geolg-3dfaultnet): Labels fault surfaces voxel by voxel within a 3D seismic reflection volume (geophysics; deep learning; volume). License: MIT.
- [Grounding DINO](https://fossettlab.org/models/grounding-dino): Detects objects named in a free-text prompt, with no fixed class list (cross domain; foundation model; imagery). License: Apache-2.0.
- [hylite](https://fossettlab.org/models/hylite): Maps minerals in hyperspectral scans of drillcore and outcrop (materials mineralogy; classical; imagery, spectra). License: MIT.
- [InSAR phase unwrapping](https://fossettlab.org/models/insar-unwrap): Converts wrapped InSAR interferogram phase into continuous line-of-sight ground displacement (geophysics; deep learning; imagery). License: MIT code, CC-BY-4.0 weights.
- [KPConv](https://fossettlab.org/models/kpconv): Segments 3D point clouds using convolution defined directly on points rather than on a grid (cross domain; deep learning; point-cloud). License: MIT.
- [MOMO](https://fossettlab.org/models/momo): Analyzes Mars orbital imagery for craters, boulders, and surface landforms (planetary; foundation model; imagery). License: MIT AND CC-BY-4.0.
- [Myria3D](https://fossettlab.org/models/myria3d): Segments airborne lidar into ground, vegetation, building, water, bridge and permanent structure (remote sensing; deep learning; point-cloud). License: BSD-3-Clause.
- [N2N4M](https://fossettlab.org/models/n2n4m): Removes noise from Mars CRISM shortwave-infrared hyperspectral data (planetary; deep learning; imagery, spectra). License: MIT.
- [NeuralHydrology](https://fossettlab.org/models/neuralhydrology): Predicts streamflow from rainfall using LSTM models (hydrology; deep learning; time-series). License: BSD-3-Clause.
- [NTv3](https://fossettlab.org/models/ntv3): Produces genomic representations and predicts thousands of functional tracks across long stretches of sequence (biology; foundation model; sequence). License: Non-commercial, gated.
- [OctFormer](https://fossettlab.org/models/octformer): Segments large 3D point clouds using an octree to keep attention tractable (cross domain; deep learning; point-cloud). License: MIT.
- [Point Transformer V3](https://fossettlab.org/models/point-transformer-v3): Assigns a semantic class to every point in a 3D scan (cross domain; deep learning; point-cloud). License: MIT.
- [PointsToWood](https://fossettlab.org/models/pointstowood): Separates leaf from wood in high-resolution terrestrial lidar (forestry vegetation; deep learning; point-cloud). License: AGPL-3.0.
- [Prithvi-EO](https://fossettlab.org/models/prithvi-eo): Earth-observation foundation model developed by NASA and IBM, with models adapted for tasks including burn-scar, flood, and crop mapping (remote sensing; foundation model; imagery). License: Apache-2.0.
- [Raman mineral classifier](https://fossettlab.org/models/raman-classifier): Identifies minerals from Raman spectra by matching against the RRUFF database (materials mineralogy; classical; spectra). License: MIT.
- [RemoteCLIP](https://fossettlab.org/models/remoteclip): Links satellite imagery with natural-language descriptions, allowing images to be searched or classified without task-specific training (remote sensing; foundation model; imagery). License: Apache-2.0.
- [RF-DETR](https://fossettlab.org/models/rf-detr): Detects objects in real time, and fine-tunes on a small set of local annotations (cross domain; deep learning; imagery). License: Apache-2.0.
- [SAM 2](https://fossettlab.org/models/sam2): Segments objects in images and video without task-specific training (cross domain; foundation model; imagery). License: Apache-2.0 AND BSD-3-Clause.
- [SaProt](https://fossettlab.org/models/saprot): Creates protein embeddings that use structure as well as sequence (biology; foundation model; sequence, structure). License: MIT AND GPL-3.0.
- [Satlas](https://fossettlab.org/models/satlas): Pretrained backbones for satellite and aerial imagery, used as a starting point for task-specific models (remote sensing; foundation model; imagery). License: Apache-2.0.
- [SegmentAnyTree](https://fossettlab.org/models/segment-any-tree): Separates a forest lidar point cloud into individual trees (forestry vegetation; deep learning; point-cloud). License: MIT.
- [SeisBench](https://fossettlab.org/models/seisbench): Picks earthquake phase arrivals from seismic waveforms (geophysics; deep learning; waveform). License: GPL-3.0.
- [SeisT](https://fossettlab.org/models/seist): Extracts several earthquake measurements from a single waveform: phase arrivals, polarity, magnitude, back-azimuth, and distance (geophysics; deep learning; waveform). License: MIT.
- [SigLIP 2](https://fossettlab.org/models/siglip2): Scores images against text, for zero-shot triage and text search across image collections (cross domain; foundation model; imagery). License: Apache-2.0.
- [Sonata](https://fossettlab.org/models/sonata): Creates embeddings from 3D point clouds without requiring labelled training data (cross domain; foundation model; point-cloud). License: Apache-2.0 code, CC-BY-NC-4.0 weights.
- [StormCast](https://fossettlab.org/models/stormcast): Nowcasts convective storms over the continental United States on a 3 km grid (weather climate; deep learning; gridded-fields). License: Apache-2.0.
- [SuperPoint Transformer](https://fossettlab.org/models/superpoint-transformer): Performs semantic and panoptic segmentation of very large scans by grouping points before classifying them (cross domain; deep learning; point-cloud). License: MIT.
- [TabFM](https://fossettlab.org/models/tabfm): Classifies or predicts values from tabular data without training a model for the dataset (cross domain; foundation model; tabular). License: Apache-2.0 code, TabFM Non-Commercial License v1.0 weights.
- [TerraMind](https://fossettlab.org/models/terramind): Multimodal Earth-observation foundation model that combines radar, optical, elevation, and land-cover data, and can generate one modality from another (remote sensing; foundation model; imagery). License: Apache-2.0.
- [TimesFM](https://fossettlab.org/models/timesfm): Forecasts regularly sampled time series without training a new model for each dataset (cross domain; foundation model; time-series). License: Apache-2.0.
- [TreeLearn](https://fossettlab.org/models/treelearn): Segments individual trees from ground-based and mobile lidar scans (forestry vegetation; deep learning; point-cloud). License: MIT.
- [treeX](https://fossettlab.org/models/treex): Separates individual trees from laser scans without any trained model (forestry vegetation; classical; point-cloud). License: MIT.
- [TRELLIS.2](https://fossettlab.org/models/trellis2): Generates a textured 3D mesh with materials from a single image (cross domain; foundation model; imagery). License: MIT.
- [Utonia](https://fossettlab.org/models/utonia): Creates embeddings from remote-sensing point clouds without requiring labelled training data (cross domain; foundation model; point-cloud). License: Apache-2.0 code, CC-BY-NC-4.0 weights.

## XR Scenes

- [XR scenes](https://fossettlab.org/xr): Interactive 3D scenes that run in a browser over WebXR and can be entered with a Meta Quest or another headset. Each scene loads its model only on request.
- [Gateway Arch](https://fossettlab.org/xr/gateway-arch/): The Gateway Arch and the St. Louis riverfront. Arch geometry swept in Blender from the National Park Service's published centroid equation and section dimensions; terrain from USGS with Missouri 2023-24 aerial imagery; buildings, paths, ponds and trees placed from OpenStreetMap. A geographic visualization with simplified interpretations of rooflines, facades and materials — not a survey and not a photographic reconstruction. Three camera presets, a walking mode beneath the Arch, and a downloadable glTF file.
- [Moon](https://fossettlab.org/xr/moon/): A global lunar terrain view, streamed rather than downloaded: LOLA topography carrying the LROC wide-angle morphology mosaic, with one area refined over the Apollo 11 landing site. Terrain and imagery are sampled differently and the viewer reports each separately at the view center. The imagery keeps the shadows the spacecraft photographed; the brightness and continuity controls are reversible display treatments rather than calibrated reflectance or physical relighting, and close views go dark or coarse where the source detail runs out. Feature labels mark where something is, not where the terrain is most detailed. Runs in a desktop browser and in a headset browser over WebXR; headset readability, comfort and frame rate have not been measured on a device.

## Guides

Technical guides at https://fossettlab.org/guides — annotated GNSS and lidar file formats, and the calculations behind survey planning. Each is written from a real lab file or dataset; samples are byte-exact and every number traces to the file, a datasheet, or a script that regenerates it.

- [Guides index](https://fossettlab.org/guides): All six, grouped into reading a file you already have and estimating something before you collect it.
from the site navigation; each answers one specific question.

- [Reading a RINEX observation file](https://fossettlab.org/guides/rinex-anatomy): An annotated RINEX 3.03 GNSS observation file, field by field — every header record, the epoch record, the per-satellite observation record, and how to decode observation codes such as C1C, L1C, D1C and S1C. Explains the GPS-time vs UTC 18-second offset that makes a file's write time appear to precede its last observation.
- [Calculating point density for a UAV lidar survey](https://fossettlab.org/guides/uav-lidar-point-density): Instantaneous ground point density under a UAV lidar, at lateral distance x from nadir for a sensor at altitude h emitting R rays per second, is N(x) = R·h / [4π(h² + x²)^(3/2)]. The h/(h² + x²)^(1/2) factor is cos θ, the oblique projection onto horizontal ground; dropping it — the common napkin error — overstates density by 1.41× at one altitude out from nadir and 2.24× at two. The 4π is a normalization convention rather than a measured property of the sensor. Accumulated density along a straight pass is PD = R/(2πhv) · L/(h² + L²)^(1/2), and the integration limit L is set by the sensor, not by a round number: at 50 m AGL a point 100 m along the track is at a slant range of 111.8 m, past the 100 m maximum range, so L = (r² − h²)^(1/2) = 86.6 m. For a GeoSLAM ZEB Horizon at 300,000 points/s, 50 m AGL and 5 m/s that gives 165 points/m² on the centerline and a 7.8 cm nominal spacing. Accumulated density 50 m to the side needs its own along-track integral, PD(y) = RhL/[2π(h²+y²)(h²+y²+L²)^(1/2)v] = 67 points/m², which is 41 % of the centerline value — not the 35 % that comes from scaling by the instantaneous ratio N(y)/N0, an operation that mixes an instantaneous ratio with an integrated density. Emitted rays, returns and ground returns are three different quantities: on one of the lab's own flown runs (Tisch Park, 19 April 2024) the delivered rate was 77,596 points/s, 26 % of the spec rate, and those are returns of every kind rather than ground returns. The manufacturer's 300,000/s is specified as a scan rate in points, while the model's R is an emitted rate. Returns are physically possible across 173 m at 50 m AGL, the 45° cone's 100 m is a planning convention, and flight-line spacing (25–30 m for detailed work) is neither. Sanity check: in the isotropic model exactly half the rays go downward, so the average across a swath of width W at speed v cannot exceed R/(2Wv) — 300 points/m² at W = 100 m and v = 5 m/s; that half is a property of the 4π convention, not a bound on the instrument.
- [Reading a CSRS-PPP report](https://fossettlab.org/guides/csrs-ppp-report): An annotated NRCan CSRS-PPP result set from a two-hour static GNSS occupation, block by block. The .sum file's POS block gives the same answer twice, Cartesian and geodetic: A_PRIORI is the receiver's own autonomous fix copied from the RINEX header's APPROX POSITION XYZ, ESTIMATED is the PPP solution, DIFF is the move between them in meters (0.92 m south, 0.16 m west, 2.13 m down in this run), and SIGMA(95%) is a 95 % value — about 1.96 times one standard deviation, so halve it before comparing with any figure quoted at 1 sigma. The SYST and EPOCH columns give the frame and the epoch of the coordinate, here NAD83 at 23:256:65265, which is 18:07:45 UTC on day 256 of 2023, the midpoint of the observation; NRCan's PDF writes the same thing as NAD83(CSRS) (2023.7). Outside Canada the OHT and GHT lines return OUTSIDE_GRID_LIMIT, because the CGVD28 vertical datum and the HT2_1997 geoid model are Canadian, so the only height the service returns is ellipsoidal. In the .pos file each FWD row is one epoch and its SD columns are that epoch's 95 % standard deviations, not the session's; the session answer is the last two rows, FIX and SCA, and the summary's SIGMA(95%) matches the scaled SCA row.
- [Reading a LAS/LAZ header, and what COPC adds](https://fossettlab.org/guides/las-header-anatomy): LAS stores every header value at a fixed byte offset with no field names, so reading a header means knowing which offset holds what: bytes 24-25 the version (227-byte header for LAS 1.2, 375 for LAS 1.4), byte 104 the point data record format with LAZ's compression flag in the high bits, byte 105 the record length, bytes 107-110 the legacy 32-bit point count (zero in LAS 1.4 files using point formats 6-10, where the real count is the 64-bit field at offset 247), and offsets 131-226 the scale factors, offsets and bounds. X, Y and Z are signed 32-bit integers: coordinate = (record value x scale) + offset, and the offset exists because a UTM northing divided by a 0.001 m scale exceeds the 2,147,483,647 integer ceiling; the test is two-sided about the offset, -2,147,483,648 <= (coordinate - offset) / scale <= 2,147,483,647 at both ends of every axis, not extent divided by scale. The coordinate reference system is stored in a variable length record and not in the header at all - an ordinary VLR or an extended one, LASF_Projection record 2112 for WKT, with Global Encoding bit 4 saying whether WKT or GeoTIFF keys are used; a file with no such record has no CRS and no units, which is normal for a handheld SLAM export delivered in its own scanner-local frame and means the transform to a mapping frame has to travel separately. COPC is a LAZ 1.4 file carrying the same cloud in octree order plus two records: a 160-byte copc info VLR that must be the first VLR, beginning at byte 375, and a copc hierarchy extended VLR after the point data giving each octree node's file offset, byte length and point count. Worked against the lab's published Greenwood Cemetery cloud (947,352 points of USGS 3DEP lidar, LAS 1.4 point format 6, EPSG:6344+5703): the plain LAZ is 5,014,321 bytes and the COPC 4,990,292 with 1,074 bytes of index, and the COPC's coarser 0.01 m scale requantizes every coordinate so its decoded extent (735,472.07 to 736,256.83 in X) no longer matches the millimeter bounds its header inherited from the LAZ, and three HTTP range requests totaling 184,716 bytes - 3.7 percent of the file - return a 19,776-point whole-site overview.

- [How long to occupy a GNSS point](https://fossettlab.org/guides/gnss-occupation-time): How long a static GNSS occupation needs to be when the results will be postprocessed by PPP (precise point positioning, one receiver and no base station), measured on a two-hour campus occupation solved by NRCan CSRS-PPP. In that run the 95% uncertainty in height — the slowest component, because every satellite is above the antenna — first fell below 10 cm at 20 minutes, 5 cm at 43 minutes and 2 cm at 87 minutes, and never reached 1 cm in two hours; north reached the same marks at 12, 19 and 38 minutes. Each doubling of occupation length divided the height uncertainty by about 2.2. The number to quote at the end is not the last forward-filter epoch's: CSRS-PPP resolves carrier-phase ambiguities on a backward pass, which moved the east coordinate 2.5 cm, and it scales static uncertainties before reporting, so the quoted figures are ±0.87 cm north, ±0.71 cm east and ±2.93 cm height at 95% for the full two hours. NRCan's own guidance is a few centimeters in an hour and millimeter level for sessions of 24 hours and longer.

- [How point density falls off with distance from a laser scanner](https://fossettlab.org/guides/scanner-point-density): A laser scanner samples on a fixed angular grid, so density on a surface facing it falls as 1/r² — for a Trimble SX12 at its fitted effective spacing of 1.274 mrad, about 1,540 points/m² on a wall at 20 m. Flat ground follows a different law, because the grazing-incidence cosine costs one more power: ρ(d) = h / ((d² + h²)^{3/2} · δ²), so ground density falls as 1/d³ far from the instrument and drops through 5 points/m² at 57 m from a 1.5 m tripod. Occlusion by stems steepens it further in forest, where a kernel fitted to the lab's ForestGEO scan, k(r) = C·(r+0.5)⁻²·exp(−(r/14 m)^0.84), drops through 5 points/m² at about 62 m; the scan itself puts about 85 % of a station's points within 10 m and 96 % within 25 m. Two corrections are easy to miss. The effective angular spacing is coarser than the datasheet's, since an SX12 set to 1.00 mrad behaves like 1.274 mrad on grass and returns 0.616 of the ideal grid, so nominal figures overstate density by about 1.6× and understate the station count by about 1.38×, since ground density goes as the cube of station spacing and stations per unit area as its square. And a scanner walked along a track integrates along its path, which gives one power back and leaves ground density falling as 1/x².

## Data Access

The lab maintains a public STAC (SpatioTemporal Asset Catalog) for programmatic access:

- STAC Catalog landing: https://fossettlab.org/api/stac
- STAC Search (GET + POST): https://fossettlab.org/api/stac/search
- Collections: https://fossettlab.org/api/stac/collections
- Static STAC mirror on S3: s3://bradleylab-public/stac/ (also reachable as https://bradleylab-public.s3.us-east-2.amazonaws.com/stac/)
- Quickstart with Python (pystac-client) and R (rstac): https://fossettlab.org/data#stac-quickstart

Three collections: imagery, pointclouds, radar. Datasets include drone orthomosaics (RGB, multispectral, thermal), terrestrial and aerial lidar point clouds, and ICEYE SAR scenes. Default license is CC-BY-4.0; some items carry restricted-access flags — contact the lab before redistributing.

## What We Do Summary

- **Collect** (https://fossettlab.org/collect): Drone surveys (RGB, multispectral, lidar, thermal); terrestrial laser scanning (Trimble X9, SX12, GeoSLAM Zeb Horizon); GNSS surveying (Trimble R980 + RTX, Emlid Reach RS2 / RS2+, DJI D-RTK 2); robotic total station (Trimble C5); ground-penetrating radar (Sensoft pulseEKKO Pro); coordinated commercial-satellite tasking (ICEYE SAR, Planet, multispectral) for WashU clients only.
- **Process** (https://fossettlab.org/process): Photogrammetric reconstruction (Pix4D, OpenDroneMap); orthomosaic generation; multispectral and thermal indices (NDVI, NDRE, GNDVI, thermal anomaly maps); point-cloud processing (CloudCompare, lidR, PDAL, Trimble Business Center) including classification, ground filtering, canopy height models, registration, and change detection. Automated, version-controlled pipelines; cloud-native archiving (COG, COPC, GeoPackage) with open standards (STAC).
- **Analyze** (https://fossettlab.org/analyze): GIS workflows (ArcGIS Pro, QGIS); statistical analysis with reproducible Python and R pipelines. Containerized deep-learning models for: imagery analysis (SAM 2 + samgeo, DeepForest, thermal-video animal detection); remote-sensing foundation models (Clay, Prithvi-EO, Satlas, RemoteCLIP); point-cloud segmentation (SegmentAnyTree for UAV and airborne lidar, TreeLearn for ground-based lidar, 3DFin as a classical baseline); and models from fields beyond geospatial science — seismic phase picking (SeisBench: PhaseNet, EQTransformer), rainfall-runoff prediction (NeuralHydrology), powder-XRD multi-phase identification (autoXRD), and genome-scale sequence modeling (Evo 2). Compute: H100-class GPUs for training and large inference, AWS (EC2, Fargate, Lambda) for pipeline and processing workloads, burstable neocloud GPUs for larger jobs, all in version-controlled containers.
- **Share** (https://fossettlab.org/share): Publication maps and figures, orthomosaics and DEM hillshades, browser-based 3D point cloud and mesh viewers, Gaussian splats, XR (Meta Quest 3, HoloLens 2), interactive web maps, and the open data catalog with STAC API.
- **Access & Rates** (https://fossettlab.org/access): Free initial scoping; internal (WashU recharge) and external rates structured per project; written estimates after scoping.

## Equipment Categories

Field-hardware equipment (with specs and availability) is listed at https://fossettlab.org/collect; XR hardware at https://fossettlab.org/share.

- **Drones**: DJI Phantom 4 Pro, Phantom 4 Multispectral, M600 Pro, Tello; Quantum-Systems Trinity F90+
- **Sensors**: MicaSense Altum-PT (multispectral); YellowScan Qube 240 (lidar); GeoSLAM Zeb Horizon (mobile lidar); Sony RX1R II (high-resolution RGB); Sensoft pulseEKKO Pro (GPR)
- **Survey & Positioning**: Trimble SX12, X9, R980, C5, TSC5; Emlid Reach RS2 / RS2+; DJI D-RTK 2
- **XR**: Meta Quest 3; Microsoft HoloLens 2
- **Connectivity**: Starlink Mini; Hologram SIM
- **Software**: Pix4D, CloudCompare, Trimble Business Center, Label Studio

## Lab Tools

Auth-gated web tools for lab users and collaborators:

- Documentation: https://docs.fossettlab.org
- Upload Files: https://upload.fossettlab.org
- Label Studio: https://labels.fossettlab.org
- WebODM: https://webodm.fossettlab.org

## MCP Server

The lab runs a Model Context Protocol server at https://mcp.fossettlab.org — a
machine interface rather than a website. It gives AI agents and research tools
governed access to lab research computing: dataset discovery, geospatial
processing, and batch compute.

- Endpoint: https://mcp.fossettlab.org/mcp (HTTP transport)
- Sign-in is via GitHub, and signing in alone grants nothing. Every request is
  authorized against lab project membership, scoped to that project's data and
  resources, and recorded. Non-members are refused.
- Lab members connect with, for example:
  `claude mcp add --transport http fossett-mcp https://mcp.fossettlab.org/mcp`

It is not part of this site and does not appear in the sitemap; sitemaps cover
one origin, and the MCP server serves no indexable pages.

## Location

Washington University in St. Louis, Department of Earth, Environmental & Planetary Sciences, St. Louis, MO 63130, USA.

## Contact

Use the contact form at https://fossettlab.org/contact to reach the lab director (Alex Bradley) or geospatial technology specialist (Bill Winston).