- Home
- Science
SCIScience
Methods developed and validated in research
TERRA's classifiers deliver methods developed and validated in research, and published as they mature. This page lists the publications, then gives the method, the validation, and every condition under which the outputs hold.
01Publications
Publications
Methods reach TERRA once they are published. The list grows as the research does.
2 entries · 1 published
- PUB-01
Conference paper · 2026 · Salvador, BA, Brazil
Automatic Land Cover Classification with Sentinel-2 and MapBiomas Time Series
XLIV Brazilian Symposium on Telecommunications and Signal Processing (SBrT 2026)
Portuguese title: Classificação Automática de Cobertura de Solo com Séries Temporais do Sentinel-2 e do MapBiomas.
@inproceedings{melo2026sbrt, author = {Melo, Jo{\~a}o Leonardi da Silva and Magalh{\~a}es, Dimmy Karson Soares and Kolodziej, Javier Ernesto and Kuhn, Eduardo Vin{\'i}cius}, title = {Automatic Land Cover Classification with {Sentinel-2} and {MapBiomas} Time Series}, booktitle = {XLIV Brazilian Symposium on Telecommunications and Signal Processing (SBrT 2026)}, year = {2026}, address = {Salvador, BA, Brazil} } - PUB-02
Software paper
TERRA: Desktop land-cover classification from Sentinel-2 time series
Journal of Open Source Software
Describes the application itself: its design, packaging and the research workflow it makes reproducible.
02Validation
What the reference
paper reports
The SBrT 2026 paper validated the default classifier by spatial block cross-validation against MapBiomas, and by temporal coherence against the crop calendars kept by landowners.
Overall accuracy
Agreement with MapBiomas under spatial block cross-validation.
Cohen's κ
Agreement beyond what chance alone would give.
Properties
In western Paraná, checked against landowner crop calendars.
NOTEFigures as reported in the SBrT 2026 paper. They apply to the spectral Random Forest on its study areas, not to any area TERRA is pointed at.
03Method
Spectro-temporal features
The default path reads six Sentinel-2 bands (B02, B03 and B04 at 10 m; B8A, B11 and B12 at 20 m, placed on the B04 grid) across every acquisition in the window.
From that stack it builds 80 features per pixel: band statistics, temporal descriptors of NDVI, EVI and SAVI, and 22 raw NDVI dates. A Random Forest of 300 trees at depth 20 classifies them into the five MapBiomas classes.
The Temporal Transformer and Prithvi-EO 2.0 are there for controlled comparison on the same polygon and date stack, not as replacements for the published baseline.
| Model | Reads | Artifact |
|---|---|---|
| Spectro-temporal Random Forest Default | 80 features per pixel: band statistics, NDVI/EVI/SAVI temporal descriptors, 22 raw NDVI dates | 300 trees at depth 20 |
| Temporal Transformer Optional | Six bands padded to 22 dates, mean-pooled over time | PyTorch model |
| Prithvi-EO 2.0 300M Optional | One acquisition: frozen foundation-model embeddings with a Random Forest head | Weights fetched on first run |
04Reporting
Where, not only how much
Every run is checked against MapBiomas cell by cell. Producer's and user's accuracy are given per class with Wilson intervals, and quantity error and allocation error are reported apart: getting how much soybean there is wrong is a different failure from getting where it is wrong.
Agreement is also broken into spatial blocks, because an average hides whether the disagreement sits in one corner or runs throughout.
Compare sets two runs side by side, in the same region or in different states, to show how a model's performance holds up from one area to another. Diagnose measures the domain shift behind a difference (symmetric KL on NDVI, change-vector magnitude and RBF MMD) and reads each class's spectral response in detail.
05Flight rules
Limitations
Read these before trusting an output. They are grouped by where each limit comes from: most belong to the current models, which research is working to improve.
The current models
These are limits of the models TERRA ships today, not of the application. Improving the models is the subject of ongoing research; a better model reaches TERRA once it has been validated and published.
- MD-01
The output legend is fixed.
The current crop models emit classes 3, 21, 25, 39 and 41 and nothing else. They cannot predict pasture, savanna or any other MapBiomas code. An area in another biome can return a confident and semantically wrong result, which is why domain-shift diagnosis exists.
- MD-02
The models were fitted for western Paraná.
The current classifiers learned from study areas in western Paraná, Brazil. The training data are not distributed with the application.
In researchDomain-shift diagnosis ships today and says how far an area sits from that ground. Adapting the models to it remains research.
- MD-03
Class 41 is a residual bucket.
High overall accuracy against MapBiomas does not imply fine crop identity.
In researchDecomposing class 41 into crops that can be named is in research.
How results are measured
Limits of the reference an output is checked against, and of the data sources it is read from.
- MS-01
MapBiomas is not field truth.
Agreement is concordance with an annual map, often offset in time from the Sentinel-2 series. The reference is Collection 10, year 2023.
- MS-02
Areas in hectares are pixel counts.
They are not corrected for classification error.
- MS-03
Point energy figures do not resolve the field.
NASA POWER radiation is on a one-degree cell, and the other meteorology, wind included, on a 0.5 × 0.625° MERRA-2 cell. The terrain and siting maps are the exception: those resolve within the area at 30 m.
The application
Limits of TERRA itself, as of the current release.
- AP-01
The export package is partial.
It carries the run's tables, the AOI geometry and the classification raster. The accuracy assessment and the domain-shift report are not in it.
- AP-02
A board's contents do not survive closing.
Unless the board is saved under a name. The arrangement survives either way.
06Citation
Cite TERRA
Cite the method paper, and the software record for the version you used.
The repository carries a CITATION.cff, so GitHub offers the same record from its sidebar.
@inproceedings{melo2026sbrt,
author = {Melo, Jo{\~a}o Leonardi da Silva and Magalh{\~a}es, Dimmy Karson Soares
and Kolodziej, Javier Ernesto and Kuhn, Eduardo Vin{\'i}cius},
title = {Automatic Land Cover Classification with {Sentinel-2} and {MapBiomas} Time Series},
booktitle = {XLIV Brazilian Symposium on Telecommunications and Signal Processing (SBrT 2026)},
year = {2026},
address = {Salvador, BA, Brazil}
}
@software{melo2026terra,
author = {Melo, Jo{\~a}o Leonardi da Silva},
title = {{TERRA}: Desktop land-cover classification from {Sentinel-2} time series},
version = {0.7.0},
year = {2026},
url = {https://github.com/rexionmars/TERRA},
license = {GPL-3.0-only}
}