TERRA Earth Observation

Overview

Local-first research software for farm-to-landscape studies. TERRA brings Sentinel-2 land-cover classification, surface water, field delineation, vegetation health, public-register overlap, Sentinel-1 radar, management zones and EMIT mineral mapping into one Studio. Arrange multiple areas side by side, combine outputs in a node compositor, inspect them over a 3D globe, and generate PDF reports. Imagery is fetched on demand as Cloud-Optimized GeoTIFF windows; computation runs locally without an application server.

Research

TERRA grew from the SBrT 2026 land-cover study: Sentinel-2 L2A spectro-temporal descriptors, MapBiomas Collection 10 reference labels, and Random Forest with spatial block cross-validation on three western Paraná properties. The original study reported overall accuracy ≈ 0.76 and κ ≈ 0.62 against MapBiomas, with soybean retention above 84% against landowner calendars as a temporal coherence check. These are results of that study, not validation of every current product or every new area. The current Studio extends the workflow with per-class agreement, spatial disagreement, domain-shift diagnosis and complementary observation products.

• Land-cover models: 80-feature spectro-temporal Random Forest, Temporal Transformer and Prithvi-EO 2.0 frozen embeddings with a Random Forest head
• Per-class producer’s and user’s accuracy, Wilson intervals, quantity/allocation errors and spatial agreement blocks against MapBiomas
• Domain-shift diagnostics from NDVI divergence, change-vector magnitude, RBF MMD and per-feature shifts; spectral response and library comparisons
• EMIT surface mineralogy through Tetracorder rules and the USGS spectral library, with uncertainty stability and EMIT L2B comparison

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

• Draw or import areas; arrange runs and multiple areas in persistent Studio workspaces for Layout, Compare, Diagnose and Data
• Node compositor to filter, mask, mix and tabulate outputs; Globe and Viewer nodes; evaluate the same graph across a field set
• Surface-water persistence from NDWI, MNDWI or AWEI; field delineation from the Fields of The World baseline
• Vegetation health from NDVI and NDRE departures against the same time of year in three earlier seasons
• Public-register overlap: PRODES, DETER, IBAMA/ICMBio embargoes, indigenous lands, conservation units and CAR registrations
• Sentinel-1 RTC VV/VH series by relative orbit, open-water screening and canopy-loss indicators; fuzzy c-means management zones from seasonal NDVI
• EMIT mineral maps, absorption-band positions and uncertainty stability; Earthdata token required
• Typst PDF reports with method, scope, maps, product tables and limitations; CSV tables and partial research-pack exports
• FULL packages embed Python for the spectral Random Forest; LITE uses local Python. Optional neural and field-segmentation dependencies are managed in Settings

Scope & interpretation

• Crop models were fitted on western Paraná areas and emit only MapBiomas codes 3, 21, 25, 39 and 41. Confidence outside that domain is not evidence of correct class identity.
• MapBiomas agreement is concordance with an annual reference map (Collection 10, 2023), not field truth. Pixel-count areas are not adjusted for classification error.
• Vegetation departures and radar canopy losses are screening indicators; management zones show differences, not their causes. Register overlap is not a legal finding.
• Mineral band depth is not abundance. Vegetation, water and clouds suppress the mineral answer; no unmixing is performed.
• Save a Studio board under a name to preserve its contents. Exports currently omit accuracy and domain-shift reports; independent validation and fuller exports remain on the roadmap.