allenai/s2-naip

AI2-S2-NAIP is a remote sensing dataset consisting of aligned NAIP, Sentinel-2, Sentinel-1, and Landsat images spanning the entire continental US.

Data is divided into tiles. Each tile spans 512x512 pixels at 1.25 m/pixel in one of the 10 UTM projections covering the continental US. At each tile, the following data is available:

• National Agriculture Imagery Program (NAIP): an image from 2019-2021 at 1.25 m/pixel (512x512).
• Sentinel-2 (L1C): between 16 and 32 images captured within a few months of the NAIP image at 10 m/pixel (64x64).
• Sentinel-1: between 2 and 8 images captured within a few months of the NAIP image at 10 m/pixel (64x64).
• Landsat-8/9: 4 images captured in the same year as the NAIP image at 10 m/pixel (64x64).
• OpenStreetMap: a GeoJSON containing buildings, roads, and 30 other categories. It uses pixel coordinates relative to the 512x512 NAIP image.
• WorldCover: the 2021 land cover image at 10 m/pixel (64x64).

AI2-S2-NAIP is applicable to several supervised and unsupervised tasks in remote sensing, including super-resolution (e.g. NAIP -> Sentinel-2), segmentation and detection (e.g. NAIP or Sentinel-2 -> OpenStreetMap or WorldCover), and multi-modal masked autoencoder pre-training.

For questions or feedback about AI2-S2-NAIP, please open an issue on Github at https://github.com/allenai/satlas.

Structure

Once extracted, the dataset contains the different data types in different folders. Each folder contains files named by a tile ID, which consists of the UTM projection, column, and row. The column and row are based on tiles that are 512x512 pixels with pixel coordinates at 1.25 m/pixel, e.g. 32612_960_-6049.png spans (614400, -3871360) to (615040, -3870720) in EPSG:32612 projection units.

Here is an example of NAIP data:

naip/
    32612_960_-6049.png
    32612_960_-6050.png
    32612_960_-6051.png
    ...

And an example of Sentinel-2 data:

sentinel2/
    32612_960_-6049_16.tif
    32612_960_-6049_32.tif
    32612_960_-6049_8.tif
    32612_960_-6050_16.tif
    ...

The Sentinel-2, Sentinel-1, and Landsat images are GeoTIFFS so they contain georeference metadata. Other data does not have georeference metadata, but data at each tile is aligned, so the georeference metadata from the above images is applicable to the other data as well with only a resolution shift.

Mapping Longitude and Latitude to Tile

Here is an example of mapping longitude and latitude to a tile.

First install packages:

pip install rasterio shapely utm

Then launch Python shell:

from rasterio.crs import CRS
from rasterio.warp import transform_geom
import shapely
import utm
# Define source location.
src_crs = CRS.from_epsg(4326)
src_point = shapely.Point(-122.331711, 47.648450)
# Get UTM zone.
_, _, zone_suffix, _ = utm.from_latlon(src_point.y, src_point.x)
epsg_code = 32600 + zone_suffix
dst_crs = CRS.from_epsg(epsg_code)
# Transform to UTM CRS.
dst_point = transform_geom(src_crs, dst_crs, src_point)
dst_point = shapely.geometry.shape(dst_point)
# dst_point is in projection coordinates (meters).
# Now convert to pixel coordinates at 1.25 m/pixel.
col = int(dst_point.x/1.25)
row = int(dst_point.y/-1.25)
# Print the prefix for the image filenames.
print(f"{epsg_code}_{col//512}_{row//512}")
# Print the prefix for the tar filenames to know which one to download.
# These group together many 1.25 m/pixel 512x512 tiles into one tar file.
print(f"{epsg_code}_{col//512//32}_{row//512//32}")

So then you would download the tar file from the second prefix, extract it, and look at the file with name matching the first prefix.

See visualize_tile.py for example of visualizing the data at a particular tile.

Sentinel-2

The 10 m/pixel (_8.tif), 20 m/pixel (_16.tif), and 60 m/pixel (_32.tif) bands are stored separately. Pixel values are the L1C 16-bit values.

The band order is as follows:

• _8.tif (64x64): B02, B03, B04, B08
• _16.tif (32x32): B05, B06, B07, B8A, B11, B12
• _32.tif (16x16): B01, B09, B10

The GeoTIFFs contain multiple images concatenated along the channel axis. The CSV shows the original Sentinel-2 scene ID of each image.

Sentinel-1

The Sentinel-1 bands are 10 m/pixel and ordered VV then VH.

Only IW VV+VH scenes are used.

The pixel values are 32-bit floating point values representing decibels 10*log10(x). We obtain the radiometric-calibrated and terrain-corrected images from Google Earth Engine so see https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD for details.

The GeoTIFFs contain multiple images concatenated along the channel axis. The CSV shows the original Sentinel-1 scene ID of each image.

NAIP

The NAIP image is 512x512 with four 8-bit bands: R, G, B, IR.

It is encoded as PNG but the IR is alpha mask so cannot be visualized correctly in image viewer without removing the alpha mask.

There are two NAIP images available, one under "naip" (2019-2022) and one under "oldnaip" (2015-2018). The CSV shows the original NAIP scene ID of each image.

Landsat

We include OLI-TIRS images from Landsat-8 and Landsat-9. As with Sentinel-2, we select Landsat images that were captured within a few months of the NAIP image.

We store the 15 m/pixel bands (i.e. B8) at 10 m/pixel, and the 30 m/pixel bands (all the others) at 20 m/pixel. There are separate GeoTIFFs for the 10 m/pixel (_8.tif) and 20 m/pixel (_16.tif). All pixel values are 16-bit.

The band order is as follows:

• _8.tif (64x64): B8
• _16.tif (32x32): B1, B2, B3, B4, B5, B6, B7, B9, B10, B11

The GeoTIFFS contain multiple images concatenated along the channel axis. The CSV shows the original Landsat scene ID of each image.