Record 2 — Open Satellite Imagery (Aligned with the Oil Palm Label Grids) Overview Per-grid satellite imagery chips and learned embeddings covering the same 6 × 6 km annotation grids as the label record, derived from three openly licensed sources. Every chip is provided as a label-date composite, aligned to the Sentinel-2 10 m geotransform and the native UTM coordinate reference system (UTM CRS) of its corresponding grid, ensuring that imagery pixels, reference labels, and learned embeddings spatially coincide. Planet NICFI imagery is not included in this record. Because Planet NICFI imagery is distributed under a non-commercial license, it is provided separately in Record 3 – Planet NICFI Basemap Imagery (record3_planet_nicfi/). This record contains only openly redistributable imagery. Component-wise Licensing (All Open) Each data source retains its own license, as the licensing terms differ among providers. Accordingly, this record is deposited under "Other (Open)" in Zenodo, with a dedicated LICENSE.txt and NOTICE.txt included in each subfolder. All three data sources permit redistribution, including commercial reuse, subject to their respective license terms. Subfolder Source License Required attribution sentinel2/ ESA/EC Copernicus Sentinel-2 L2A Copernicus Sentinel Data Terms "Contains modified Copernicus Sentinel data [year]" landsat8/ USGS/NASA Landsat 8/9 Collection 2 Level-2 Public domain (no restrictions) "Courtesy of the U.S. Geological Survey" (requested) alphaearth/ Google Satellite Embedding V1 (AlphaEarth) CC BY 4.0 "Produced by Google and Google DeepMind" Contents record2_satellite_imagery/├── README.md├── imagery_year_manifest.csv new filename → source file + per-grid year (S2 & L8)├── metadata.json├── sentinel2/│ ├── LICENSE.txt NOTICE.txt│ └── label_S2_10m_grid_XX_YYYY.tif (B2,B3,B4,B8 @ 10 m)│ label_S2_20m_grid_XX_YYYY.tif (B5,B6,B7,B8A,B11,B12 @ 20 m)├── landsat8/│ ├── LICENSE.txt NOTICE.txt│ └── label_L8_grid_XX_YYYY.tif (7 SR bands @ 30 m; resampled variants tagged _10m_/_20m_)└── alphaearth/ ├── LICENSE.txt NOTICE.txt ├── fetch_alphaearth.py reproducible Earth Engine fetch script ├── alphaearth_manifest.csv grid → year → CRS → output file └── alphaearth_embedding_grid_XX_YYYY.tif (64-band annual embedding @ 10 m) Same-Year Naming (for Training) Filenames encode each grid's label year (YYYY). Sentinel-2 and Landsat imagery therefore share the same year for each grid, ensuring that every image–label training pair is temporally consistent. The original source filenames (including per-sensor acquisition dates and, where applicable, 3-month-median composites) are preserved in imagery_year_manifest.csv. The manifest also documents one source correction: the Sentinel-2 files for Grid 16 were incorrectly named grid_03 in the original archive and have been restored to grid_16 after verification using the image geotransform. Coverage This record is scoped to the 72 benchmark grids used in the associated publication. Not every grid contains data from every sensor. Layer Grids Missing cells Sentinel-2 10 m 69 03, 18, 46 (no 10 m composite available in the source) Sentinel-2 20 m 70 18, 46 (no Sentinel-2 imagery available) Landsat 8/9 71 (73 files) 06 (no Landsat imagery); Grid 46 includes three resolution variants AlphaEarth 69 03, 18, 46 (derived from the Sentinel-2 10 m grids) Planet NICFI (Record 3) 72 — All 72 benchmark grids have corresponding Planet NICFI imagery and reference labels. Users can join all datasets using the common grid identifier and should account for the missing imagery listed above. Notes on Alignment and Years Each benchmark grid is composited using its own label year (2019–2025). Because oil palm landscapes change rapidly, using a fixed-year mosaic would introduce temporal discrepancies between imagery and reference labels. Sentinel-2 and Landsat imagery therefore share the same label year for each grid. Additional details are provided in imagery_year_manifest.csv and grid_index.csv (Record 1). AlphaEarth embeddings were generated independently for each grid using Google Earth Engine and exported on the exact Sentinel-2 10 m geotransform. The complete workflow can be reproduced using fetch_alphaearth.py. Related Records Record 1 — Geospatial Labels (CC BY 4.0): Reference annotation polygons for the benchmark grids. Record 3 — Planet NICFI Basemap Imagery (Non-commercial): Planet NICFI basemap image chips aligned with the same benchmark grids.