TERRA

Overview

Open-source desktop instrument that turns the SBrT 2026 land-cover pipeline into an interactive workflow: draw or import an AOI, preview Planetary Computer scenes, run classifiers, and inspect overlays, confidence, phenology, and side-by-side run comparison — without downloading full Sentinel-2 products. Releases ship LITE (bring your own Python) and FULL (embedded Python 3.12 for spectral Random Forest) zip builds for macOS, Windows, and Linux.

Research

Agricultural land-cover mapping at parcel scale still often requires bespoke notebooks: STAC queries, spectro-temporal feature stacks, and classifiers that are hard to re-run outside the lab. Following the SBrT 2026 study, the research line behind TERRA develops automatic classification from Sentinel-2 L2A time series with MapBiomas (Collection 10) as reference labels, using a Random Forest on spectro-temporal descriptors and spatial block cross-validation. Reported overall accuracy ≈ 0.76 and κ ≈ 0.62 quantify agreement with MapBiomas under that protocol — not pixel-level field accuracy. Landowner crop calendars, used as a temporal coherence check, showed soybean retention above 84% as the temporal stack consolidated, consistent with capture of the soybean–winter cereal rotation signature. Evaluation used three private properties with distinct crop rotations; TERRA packages the method (and later model variants) as desktop software so any AOI can be classified, saved, and compared without re-authoring the pipeline.

• Sentinel-2 L2A spectro-temporal features + MapBiomas Collection 10 reference labels
• Random Forest with spatial block cross-validation on three properties (distinct rotations)
• OA ≈ 0.76 and κ ≈ 0.62 vs MapBiomas under block CV (agreement, not field accuracy)
• Soybean retention > 84% vs landowner calendars as a temporal coherence check

Melo, J. L. S., Magalhães, D. K., Kolodziej, J. E., Kuhn, E. V. Automatic Land Cover Classification with Sentinel-2 and MapBiomas Time Series.

XLIV Brazilian Symposium on Telecommunications and Signal Processing (SBrT 2026), Salvador, BA.

Software

TERRA is the desktop shell around that pipeline: interactive AOI tooling, model selection, map overlays, analysis persistence, and compare views — so the research method can be exercised outside the notebook.

• Draw, search, or import AOIs (KML / GeoJSON); example areas from the reference study
• Planetary Computer STAC discovery with optional data-cube preview before Classify
• Models: spectro-temporal Random Forest, Temporal Transformer, Prithvi-EO 2.0 (pixel / patch)
• Map and Analysis views: prediction & confidence overlays, MapBiomas reference, VI & phenology
• Local persistence of runs and dedicated Compare mode (side-by-side overlays and class bars)
• Release flavors: FULL (embedded Python for spectral RF) and LITE (system Python + requirements.txt)
• Cloud-Optimized GeoTIFF windowed reads — no full-scene download required
• Public roadmap: change detection, crop diagnostics, surface-water mapping, richer phenology-by-class views