# Downstream Tasks

Each sample consists of a 12-band Sentinel-2 Surface Reflectance image (B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, and B12), a four-band NAIP image (Red, Green, Blue, and Near-Infrared), and a task-specific label image.

## ChesapeakeRSC

- **Task type:** Land cover segmentation
- **Source dataset:** Chesapeake Remote Sensing Classification (RSC): https://github.com/isaaccorley/ChesapeakeRSC
- **Classes:** Water, Emergent Wetlands, Tree Canopy, Scrub/Shrub, Low Vegetation, Barren, Impervious Structures, Other Impervious, Impervious Roads, Tree Canopy over Structures, Tree Canopy over Other Impervious, Tree Canopy over Impervious Roads, Aberdeen Proving Ground
- **Label format:** Semantic segmentation masks

## RoadDetection

- **Task type:** Binary semantic segmentation
- **Source dataset:** Microsoft RoadDetections dataset derived from high-resolution images: https://github.com/microsoft/RoadDetections
- **Classes:** Road / background
- **Label format:** Binary segmentation masks

## USBuildings

- **Task type:** Binary semantic segmentation
- **Source dataset:** Microsoft US Building Footprints: https://www.usgs.gov/data/a-national-dataset-rasterized-building-footprints-us
- **Classes:** Building / background
- **Label format:** Binary segmentation masks

## VermontLC

- **Task type:** Land cover segmentation
- **Source dataset:** Vermont land cover dataset: https://geodata.vermont.gov/pages/land-cover
- **Classes:** Tree Canopy, Grass/Shrub, Bare Soil, Water, Buildings, Roads, Other Paved, Railroads, others
- **Label format:** Semantic segmentation masks

## CHM-NAIP

- **Task type:** Regression
- **Source dataset:** Canopy height model and NAIP imagery pairs across CONUS: https://github.com/allredbw/chm-naip
- **Target:** Continuous canopy height values
- **Label format:** Raster regression targets
- **Note:** NAIP GeoTIFFs in this dataset are stored as five-band 8-bit images: RGBN plus a mask band. A mask value of 0 indicates an invalid pixel, while a value of 1 indicates a valid pixel. CHM GeoTIFFs are scaled by 100 and stored as single-band 16-bit images, with 65535 used as the no-data value.