Journal of Remote Sensing submission package

Boundary information without parcel objects

A controlled probing study tests whether the AlphaEarth annual embedding preserves parcel-boundary information for small-parcel coffee delineation in Minas Gerais, Brazil, compared with seasonal Sentinel-1/2 features.

Study Summary

AlphaEarth carries a stronger threshold-free boundary signal than the seasonal baselines, but the advantage does not reliably become better parcel objects under a fixed watershed extraction. The failure is concentrated in small-parcel tiles, where smoother boundary fields merge adjacent coffee parcels.

308,293 CONAB/fiboa coffee parcels (≥ 1 pixel) in Minas Gerais, 2017
1.00 ha median evaluated reference parcel area
+0.120 median Boundary AP gain for AlphaEarth over S2 seasonal features
0.179 best object recall at IoU >= 0.5 across tested configurations

Figures

Experimental design from reference parcels through feature representations, probing, watershed extraction, and paired comparison.
Experimental design with fixed labels, probe, extraction, and paired tests.
Paired tile-level effects showing Boundary AP gains but no consistent object metric gains.
Boundary information improves while object-level metrics remain comparable or worse.
Size-stratified performance plots for mean IoU, Boundary F1, and Boundary AP.
The information-object gap is strongest in the small-parcel stratum.
Study area map for Minas Gerais with evaluation tiles and parcel-area distribution.
Study area, spatial split, and reference parcel-area distribution.

Artifacts

Code

Public reproduction repository: paper-alphaearth-mg.

Archive

Permanently archived on Zenodo: 10.5281/zenodo.21222710.

Data Sources

CONAB/fiboa br_conab, Sentinel-1, Sentinel-2, and AlphaEarth Satellite Embedding V1 via Google Earth Engine.

Reproduction Commands

The full run requires Docker, Google Earth Engine authentication, and an Earth Engine project configured through .env. 

cp .env.example .env
# edit EE_PROJECT in .env
docker compose build
docker compose run --rm pipeline python code/gee_check.py
docker compose run --rm pipeline python code/01_extract_reference.py
docker compose run --rm pipeline python code/02_tiles.py --n 80
docker compose run --rm pipeline python code/03_gee_stacks.py
docker compose run --rm pipeline python code/04_probe.py
docker compose run --rm pipeline python code/05_validate.py
docker compose run --rm pipeline python code/05_validate.py --variant lin
docker compose run --rm pipeline python code/06_figures.py
docker compose run --rm pipeline python code/08_pipeline_figure.py
docker compose run --rm pipeline python code/check_artifacts.py

Citation

BibTeX for the archived research package:

@misc{abreufilho2026alphaearthmg,
  title = {Boundary information without parcel objects: probing the AlphaEarth annual embedding against seasonal Sentinel-1/2 features for small-parcel coffee delineation},
  author = {Abreu Filho, Antonio Campos de and Timbo Elmiro, Marcos Antonio and Nero, Marcelo Antonio and Leitao, Rodrigo Machado Fernandes},
  year = {2026},
  doi = {10.5281/zenodo.21222710},
  url = {https://alphaearth-mg.abreufilho.com.br}
}

Figure Provenance

Figures and tables on this page are generated from the study data and code. No AI-generated images or multimedia are used in the manuscript figures or on this site.