EarthView / README.md
gera-richarte's picture
Update README.md
05b6c30 verified
metadata
dataset_info:
  - config_name: default
    features:
      - name: 1m
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: chm
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: rgb
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: metadata
        struct:
          - name: bounds
            sequence: float64
          - name: epsg
            dtype: string
          - name: siteID
            dtype: string
          - name: timestamp
            sequence: string
    splits:
      - name: train
        num_bytes: 303349477315
        num_examples: 35501
    download_size: 240895951943
    dataset_size: 303349477315
  - config_name: satellogic
    features:
      - name: 1m
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: rgb
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: metadata
        struct:
          - name: bounds
            sequence: float64
          - name: crs
            sequence: string
          - name: timestamp
            sequence: string
    splits:
      - name: train
        num_bytes: 3675346598134
        num_examples: 2967663
    download_size: 3528764568282
    dataset_size: 3675346598134
  - config_name: sentinel_1
    features:
      - name: 10m
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 1762041095796
        num_examples: 1049466
    download_size: 1487586838960
    dataset_size: 1762041095796
  - config_name: sentinel_2
    features:
      - name: 10m
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: 20m
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: 40m
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: rgb
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: scl
        sequence:
          sequence:
            sequence:
              sequence: uint8
      - name: metadata
        struct:
          - name: s3Path
            sequence: string
          - name: solarAngles
            sequence: string
          - name: tileGeometry
            sequence: string
          - name: timestamp
            sequence: string
          - name: viewIncidenceAngles
            sequence: string
    splits:
      - name: train
        num_bytes: 14967110614854
    download_size: 12926621389874
    dataset_size: 14967110614854
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
  - config_name: satellogic
    data_files:
      - split: train
        path: satellogic/**/train-*
  - config_name: sentinel_1
    data_files:
      - split: train
        path: sentinel_1/train-*
  - config_name: sentinel_2
    data_files:
      - split: train
        path: sentinel_2/**/train-*

EarthView dataset

EarthView

Overview

The EarthView Dataset is a comprehensive collection of multispectral earth imagery. The dataset is divided into four distinct subsets sourced from Satellogic, Sentinel-1, Sentinel-2 (coming soon), and NEON imagers, each providing unique data.

Dataset Viewer

Check the EarthView Dataset Viewer and it's code for examples on how to access the images and navigate the dataset.

Data Sources

Each subset (AKA configuration) in the EarthView dataset includes samples representing specific patches of the Earth. Each source (satellite type) has different characteristics, so the details for the samples in each of the subsets are subtly different. We provide a very simple library to access the images in the subsets.

Available Subsets

Name Samples Unique locations Products Image Resolution
Satellogic ~6 million ~3 million RGB 1m
NIR (Near Infrared) 1m
Neon ~1 million ~0.3 million RGB 0.1m
Canopy Height Model (Lidar) 1m
Hyperspectral (369 bands) 1m
Sentinel-1 ~5.2 million ~1 million SAR (mapped to RGB) 10m (from 20m)
Sentinel-2 ~10 million ~1 million RGB 10m
NIR 10m
NIR / Red Edge / SWIR 20m
Scene Classification Layer 20m
Coastal-Aerosol / Water Vapour / Cirrus 40 (from 60m)

Data Format

Each subset has some peculiarities and a specific data format in the dataset. Each item (sample) in the dataset is a dictionary with a metadata field and one or more entries for the different image products available, such as rgb, chm, 1m, 10m (see below). All of the image fields are 4D arrays where dimensions are REVISITS, BANDS, HEIGHT, WIDTH.

We encourage you to use the supplied earthview library to simplify accessing the dataset, metadata and images. (For now, download the single python file and place in the same directory as your scripts, or in Python's path)

Bellow you'll find details for each subset

Satellogic

Metadata

key description
bounds [x_min, y_min, x_max, y_max] (bottom-left corner and top-right corner coordinates in (easting, northing) format, WGS 84 / UTM).
[178191.0, 8248444.0, 178575.0, 8248828.0]
[[171507.0, 7874670.0, 171891.0, 7875054.0], [174795.0, 7878747.0, 175179.0, 7879131.0]]
crs EPSG code
['EPSG:32723']
timestamp list of timestamps corresponding to the capture dates (only date is valid)
['2022-07-21T00:00:00']
['2022-07-21T00:00:00', '2022-07-25T00:00:00']
count Number of re-visits for the location (not present in the dataset, generated by items_to_images())
1

Images

Satellogic images are captured with Satellogic's MarkIV satellite fleet. The payload produces RGB and NIR images at 1m native resolution (no PAN sharpening). Each sample in the dataset has 1 or 2 re-visits per location.

key Product Image Resolution Size (pixels) Bands Re-samples
rgb RGB 1m 384 x 384 3 1 or 2
1m NIR 1m 384 x 384 1 1 or 2

Example (Jupyter Notebook)

import numpy as np
import earthview as ev

data = ev.load_dataset("satellogic", shards=[10])  # shard is optional
sample = next(iter(data))

print(sample.keys())
print(np.array(sample['rgb']).shape)      # RGB Data
print(np.array(sample['1m']).shape)       # NIR Data
dict_keys(['1m', 'rgb', 'metadata'])
(1, 3, 384, 384)
(1, 1, 384, 384)
sample = ev.item_to_images("satellogic", sample)
display(sample)
display(*sample["rgb"])
display(*sample["1m"])
{'1m': [<PIL.Image.Image image mode=L size=384x384>],
 'rgb': [<PIL.Image.Image image mode=RGB size=384x384>],
 'metadata': {'bounds': [[178191.0, 8248444.0, 178575.0, 8248828.0]],
  'crs': ['EPSG:32723'],
  'timestamp': ['2022-08-13T00:00:00'],
  'count': 1}}

Satellogic RGB

Satellogic RGB

Example (iterate)

from itertools import islice
import earthview as ev

data = ev.load_dataset("satellogic", shards=[10]) # shard is optional
datai = iter(data)
for sample in islice(datai, 10):
  sample = ev.item_to_images("satellogic", sample)
  print(sample["metadata"]["bounds"])
  sample["rgb"][0].show()
### Sentinel-1

Metadata

key description
type 'Polygon' Indicates the coordinates are the vertices of a Polygon
coordinates Five coordinates of a closed Polygon.
[[[434520.0, 8715520.0], [438360.0, 8715520.0], [438360.0, 8711680.0], [434520.0, 8711680.0], [434520.0, 8715520.0]]]
crs EPSG code
'epsg:32736'
count Number of re-visits for the location (not present in the dataset, generated by items_to_images())
6

Images

Sentinel-1 carries a Synthetic Aperture Radar (SAR) payload. The data (imagery) produced has two channels, for vertical and horizontal polarization. The data in the dataset contains just the two channels, the images returned by item_to_images() implements a standard mapping to return RGB images. (see the example below). Samples in the dataset contain varied numbers of re-visits per location.

key Product Image Resolution Size (pixels) Bands Re-samples
10m RGB 10m 384 x 384 3 1 or 2

Example (Jupyter Notebook)

import numpy as np
import earthview as ev

# data = ev.load_dataset("sentinel_1", shards=[88])  # shard is optional
data = ev.load_parquet("dataset/sentinel_1/train-00088-of-01763.parquet")

sample = next(iter(data))

print(sample.keys())
print(np.array(sample['rgb']).shape)      # RGB Data
print(np.array(sample['10m']).shape)       # NIR Data
dict_keys(['10m', 'metadata'])
(6, 2, 384, 384)
sample = ev.item_to_images("sentinel_1", sample)
display(sample)
display(*sample["10m"][:2])
{'10m': [<PIL.Image.Image image mode=RGB size=384x384>,
  <PIL.Image.Image image mode=RGB size=384x384>,
  <PIL.Image.Image image mode=RGB size=384x384>,
  <PIL.Image.Image image mode=RGB size=384x384>,
  <PIL.Image.Image image mode=RGB size=384x384>,
  <PIL.Image.Image image mode=RGB size=384x384>],
 'metadata': {'type': 'Polygon',
  'crs': 'epsg:32736',
  'coordinates': [[[434520.0, 8715520.0],
    [438360.0, 8715520.0],
    [438360.0, 8711680.0],
    [434520.0, 8711680.0],
    [434520.0, 8715520.0]]],
  'count': 6}}

Sentinel-1 RGB

Sentinel-1 RGB

Neon

In the dataset, the subset/configuration for NEON is called default, but when using the earthview library you should call it neon.

Metadata

key description
bounds [x_min, y_min, x_max, y_max] (bottom-left corner and top-right corner coordinates in (easting, northing).
[-82.04138011662944, 29.634596313943526, -82.04071312100113, 29.635179014437973]
epsg EPSG code
'EPSG:32617' (this is an error, coordinates are actually 'ESPG:4326', the library (earthview) solves it)
timestamp list of timestamps corresponding to the capture dates (only date is valid)
['2018-01-01T00:00:00', '2019-01-01T00:00:00', '2021-01-01T00:00:00']
siteID 'OSBS' for Ordway-Swisher Biological Station
count Number of re-visits for the location (not present in the dataset, generated by items_to_images())
3

Images

The NEON subset is composed of very high resolution RGB images at 0.1m, 1m hyperspectral data (369 bands), and 1m Canopy Height Model out of a LIDAR sensor. Every sample in the dataset contains 3 re-visits.

key Product Image Resolution Size (pixels) Bands Re-samples
rgb RGB 0.1m 640 x 640 3 3
chm Canopy Height Model 1m 64 x 64 1 3
1m Hyperspectral 1m 64 x 64 369 3

When using item_to_images() the Hyperspectral images are mapped, from the 369 bands to RGB using a meaningless mapping. Please, don't use it for anything else than an example.

Example (Jupyter Notebook)

import numpy as np
import earthview as ev

data = ev.load_dataset("neon", shards=[100])  # shard is optional
# data = ev.load_parquet("neon", batch_size = 10)

sample = next(iter(data))

print(sample.keys())
print(np.array(sample['rgb']).shape)       # RGB Data
print(np.array(sample['chm']).shape)       # Canopy Height
print(np.array(sample['1m']).shape)        # Hyperspectral
dict_keys(['1m', 'chm', 'rgb', 'metadata'])
(3, 3, 640, 640)
(3, 1, 64, 64)
(3, 369, 64, 64)
sample = ev.item_to_images("neon", sample)
display(sample)
display(sample["rgb"][0])
display(sample["1m"][0])
display(sample["chm"][0])
{'1m': [<PIL.Image.Image image mode=RGB size=64x64>,
  <PIL.Image.Image image mode=RGB size=64x64>,
  <PIL.Image.Image image mode=RGB size=64x64>],
 'chm': [<PIL.Image.Image image mode=L size=64x64>,
  <PIL.Image.Image image mode=L size=64x64>,
  <PIL.Image.Image image mode=L size=64x64>],
 'rgb': [<PIL.Image.Image image mode=RGB size=640x640>,
  <PIL.Image.Image image mode=RGB size=640x640>,
  <PIL.Image.Image image mode=RGB size=640x640>],
 'metadata': {'bounds': [-82.04138011662944,
   29.634596313943526,
   -82.04071312100113,
   29.635179014437973],
  'epsg': 'ESPG:4326',
  'siteID': 'OSBS',
  'timestamp': ['2018-01-01T00:00:00',
   '2019-01-01T00:00:00',
   '2021-01-01T00:00:00'],
  'count': 3}}

NEON RGB

NEON Canopy Height

NEON Hyperspectral meaningless mapping

Sentinel 2 (Coming Soon)

Check the EarthView Dataset Viewer and it's code to get an idea on how to use it.

Known Issues

  • The time component of timestamps is not correct.
  • Bounds for some images is not correct. We are evaluating what has happened.. The new version of Satellogic data has correct bounds.

We intend to update the dataset to fix issues when possible