DisSCube's derivation layer is the conceptual successor of TerraME's fillCellularSpace (the Fill Cells operation), reformulated as a reproducible, catalogued data-cube layer.

In TerraME/LuccME, populating a cellular space with attributes derived from heterogeneous geographic layers is performed cell by cell, in Lua, through fill strategies (area, presence, count, distance, percentage, majority, average, etc.). DisSCube keeps the same semantic vocabulary but expresses each strategy as a typed, auto-registered Operator that runs over a catalogued data cube rather than over an in-memory cellular space.

The advance is not the set of operations — those are meant to be faithful to TerraME, and the parity benchmarks below measure how faithful they are — but the engineering around them:

• Reproducibility. Every derived product carries a deterministic spec_hash; identical specifications always produce the same catalogued product.
• Grid-aligned aggregation. Categorical and standard-deviation operators aggregate from a fine array snapped to the target grid origin, using real per-cell windows, so target resolutions that are not integer multiples of the source (e.g. 30 m → 1000 m) are well defined rather than silently approximated.
• Explicit cell purity. Coverage purity (valid / total pixels) and, for categorical operators, dominance purity (dominant-class fraction) are computed per cell. They are first-class metadata travelling with the variable, ready for a future masking policy. Purity is implicit in TerraME; here it is named and measurable — matching the parameter the INPE e-Cube design also elevates.
• CRS robustness. Aggregation is validated on real projected CRSs, including the BDC Brazil Albers grid, with named-WKT serialization to avoid PROJCS["unknown"] round-trip problems.

A TerraME fill script can also be written as a DisSCube pipeline file — see Pipeline files (TOML), the quickstart pipeline, and the parity cases in the recipes repository (cases/terrame_fill).

"Parity verified" means the operator reproduces TerraME's own output cell by cell in the benchmarks below; "parity measured" means it was compared and the residual difference is explained; the other rows are correspondences by design that have not yet been measured against TerraME.

• Continuous, resampling-expressible (mean, sum, min, max): aligned to the target grid directly via the corresponding rasterio resampling method.
• Categorical (percentage, majority, minority) and std: aligned to a fine, origin-snapped grid with nearest resampling (never averaging a class code), then reduced per target cell over real windows. These operators set needs_fine_alignment = True.
• Vector (presence, attribute, and the vector branch of the categorical operators): reprojected and clipped to the grid bounding box, then rasterized.

TerraME's gis package ships three Fill examples — Itaituba (the Fill tutorial), Emas and Amazônia — each with its input layers and the Lua script. The reference is the cellular space that TerraME 2.0.1 produces from them, archived as the goldens of LambdaGeo/luccme-goldens (v1.0.0, DOI 10.5281/zenodo.23107748): the same inputs are derived with DisSCube on the same grid and compared cell by cell.

Earlier versions of this page compared against a different reference file, whose provenance was not recorded and which TerraME 2.0.1 does not reproduce (coverage as a percentage divided by the whole cell, distance 0 in every cell that contains a feature). The explanations drawn from it — that TerraME measures distance from the cell polygon and divides coverage by the whole cell — do not hold for the goldens and were removed. elevation, population and all of Emas are identical in both files.

The full parity benchmark suite is maintained and executed on CI in the DisSCube Case Studies and Recipes repository (cases/terrame_fill), where input datasets are fetched from TerraME with cryptographic checksum verification (./run.sh all <dataset>).

Below is the summary of the correspondences and cell-by-cell numerical parity verified across the datasets.

A 31 × 20 cellular space in SAD69 / UTM 21S (EPSG:29191), filled from a 923 m elevation raster, a 60 m land-cover raster, roads, localities and census tracts.

Setup. Grid bbox = [547177.35, 9485214.19, 702177.35, 9585214.19], resolution 5000 m, taken from itaituba.shp (whose .prj is wrong — NAD83_Austin — so EPSG:29191 is assigned explicitly). TerraME numbers rows from the south (row = 0 is the southern row); DisSCube is north-up, so TerraME (row, col) maps to DisSCube (19 − row, col). Coverage values are compared in percent (DisSCube fraction × 100).

Reading the results.

• Continuous averages agree. The residual in elevation comes from the averaging rule (area-weighted resampling vs. TerraME's per-pixel average) on a coarse 923 m source.
• Coverage agrees with no correction. TerraME 2.0.1 divides the class area by the valid pixels of the cell, as DisSCube does; the residual is at most 0.64 pp. Multiplying by coverage_purity — right for the earlier reference — would now be wrong.
• TerraME measures distance from the cell centre to the nearest vertex of the feature. For the localities (points) that is the feature itself, so distance is identical. For roads TerraME ignores the middle of the segments, so distance (to the segment) is smaller wherever a road passes between vertices; recomputing the distance to the nearest vertex reproduces distroad in 100 % of the cells, to the millimetre. The explanation is inferred from the output; TerraME's source was not read.
• min_distance is an approximation. It rasterizes the features with all_touched, takes the Euclidean distance transform between cell centres and multiplies by the resolution, so its error reaches about one cell.
• Rasters without a declared nodata. The deforestation raster declares none, and its classes include 255. Earlier versions reprojected it with the uint8 default fill value (255) and then treated that value as nodata, dropping the legitimate class 255. Since the fix, the fill value can no longer collide with the data, and defor_255 agrees with TerraME in every cell.

Cells of the Emas National Park (EPSG:29192) that touch the park limit, filled from firebreak and river lines and a 30 m fire-accumulation raster.

Cells of the Brazilian Amazon (EPSG:29191) that touch its limit, filled from the 5 km PRODES raster, roads, ports and indigenous lands.

• Distance to lines. TerraME 2.0.1 measures from the cell centre to the nearest vertex of a line; distance measures to the segment, which is the geometrically correct value and is smaller where a line passes between vertices. There is no vertex mode yet; for points the two coincide.
• Cell assignment of pixels and lines. min/max count pixels that straddle a cell border, and presence rasterizes lines through cell centres; TerraME assigns each pixel to the cell containing its centre and marks every cell a line touches (Emas).
• In-memory, single-tile by design. The fine-alignment path materializes a fine array in memory; very large tiles at a high fine/target ratio are bounded by available memory — params = {subcells = n} caps the ratio. Distributed/lazy execution is a roadmap item, not a current capability.

The filling of cellular spaces from heterogeneous geographic data — the core operation of TerraME's fillCellularSpace — is reformulated in DisSCube as a reproducible spatial-derivation layer: fill strategies become typed operators over a catalogued data cube, with aggregation on windows aligned to the target grid and explicit control of cell purity. On the three Fill examples shipped with TerraME (Itaituba, Emas, Amazônia), raster averages and class coverage reproduce TerraME's output cell by cell, distance to points is identical and the areal-weighted sum conserves TerraME's totals, while presence, min and max agree in 98–99.7 % of cells and polygon area in ≥ 99 %; distance to lines follows TerraME's vertex rule only approximately.