Data model

DHIDB v1 is a dense TileDB array with dimensions (time, y, x). The array stores annual global layers on the CLMS 300 m raster grid. TileDB Embedded reads only the selected ranges, so a point, bounding-box, or polygon query does not require downloading the global archive.

Dimension	Size	Meaning
time	12	Annual layers from 2014 through 2025
y	47,040	Global raster rows
x	120,960	Global raster columns

Each annual layer contains 5,689,958,400 grid cells. The complete logical array contains 68,279,500,800 cells before compression and exclusion of unwritten fragments.

The spatial dimensions use zero-based raster row and column indices. The array domain is time=0..11, y=0..47039, and x=0..120959 for the current 2014--2025 build. The mapping from a time index to a calendar year is stored in the array metadata as time_index; clients should use db.years rather than assuming that every future build has the same labels.

DHI variables

The three primary variables describe complementary aspects of vegetation dynamics. Let p be a pixel and y a year. The source GPP time series is denoted by g(p, y, d), where d indexes the available 10-day scenes.

Cumulative productivity: dhi_cum

Where the annual LSP product is available, dhi_cum is the sum of the two seasonal LSP total-productivity values:

dhi_cum(p, y) = TPROD(p, y, season 1) + TPROD(p, y, season 2)

TPROD is the growing-season integral of PPI between the detected start and end of a season. It is an LSP seasonal productivity summary, not a direct sum of the 10-day GPP scenes. Its stored unit is m2 m-2 day, as documented for the PPI-integral representation. If the LSP values are unavailable, the production code can use the finite GPP time series as a cumulative fallback.

Minimum productivity: dhi_min

dhi_min is the lower seasonal productivity level across the two LSP seasons:

dhi_min(p, y) = min(MINV(p, y, season 1), MINV(p, y, season 2))

MINV is the average of the PPI minima on the left and right sides of the season. The stored unit is m2 m-2. This is an LSP-derived seasonal minimum-value baseline proxy, not the pixel-wise minimum of the 10-day GPP scenes. A finite GPP minimum is available as a fallback when LSP values cannot be used.

Temporal variability: dhi_var

dhi_var is calculated from the accepted finite 10-day GPP values. Let V(p, y) be the valid scene indices and mean_g(p, y) their arithmetic mean:

mean_g(p, y) = sum(g(p, y, d) for d in V(p, y)) / count(V(p, y))

dhi_var(p, y) =
  population_standard_deviation(g(p, y, d) for d in V(p, y)) / mean_g(p, y)

This is a coefficient of variation and is dimensionless. It represents relative intra-annual variation in GPP; it is not a variance in GPP units.

Variables

Variable	Definition	Unit
dhi_cum	Sum of LSP TPROD over two growing seasons	PPI integral (m2 m-2 day)
dhi_min	Minimum of scaled LSP MINV over two growing seasons	PPI (m2 m-2)
dhi_var	Coefficient of variation of accepted 10-day GPP	dimensionless
dhi_combined	Geometric mean of normalized component layers	dimensionless
observed_count	Available GPP observations	scenes
valid_count	Accepted GPP observations	scenes
qflag_any_count	Observations carrying any source QFLAG	scenes
qflag_rejected_count	Rejected observations	scenes

dhi_cum and dhi_min summarize the smoothed Plant Phenology Index trajectory represented by Copernicus LSP. They are not direct annual GPP sums or minima. dhi_var is calculated from the 10-day GPP time series.

The count attributes are stored as uint16; the DHI attributes are stored as float32. Pixels with no finite observations use NaN for floating-point DHI attributes and zero for the count attributes. An unwritten TileDB cell can instead expose the uint16 fill value 65,535, so consumers should check for plausible counts and finite DHI values rather than treating every count as an observation.

Combined score

The schema includes dhi_combined, but the current production archive does not populate it unless global normalization bounds are supplied. For components X_k, with lower and upper bounds L_k and U_k, the planned normalization is:

Z(k) = clip((X(k) - L(k)) / (U(k) - L(k)), 0, 1)
dhi_combined = (Z(cumulative) * Z(minimum) * Z(variation)) ** (1 / 3)

This remains disabled by default because tile-local normalization would make values incomparable between locations and years. Treat dhi_combined as unavailable in v1 unless the dataset metadata explicitly provides the normalization bounds.

TileDB storage layout

The array is dense, not sparse. Its dimensions, tile extents, and stored attributes are fixed by the schema:

Schema element	Configuration
Dimensions	time, y, x
Dimension domains	0..11, 0..47039, 0..120959 in the current build
Tile extents	time=1, y=1024, x=1024
Spatial processing tiles	1,024 x 1,024 cells, approximately 307.2 x 307.2 km on the 300 m grid
DHI attribute type	float32
Count attribute type	uint16
Compression	Zstandard on dimensions and attributes
Cell order	Row-major

One annual layer therefore has 119 tiles across and 46 tiles down, or 5,474 spatial tiles. The one-cell time tile aligns storage with the dominant access pattern: selected annual layers over a spatial window. The physical size on disk is smaller than the logical cell count because TileDB stores compressed attribute chunks and only materialized fragments.

The array metadata records the grid width and height, affine transform, CRS WKT, spatial tile height and width, and the time_index mapping. These metadata are required to convert array coordinates into geospatial outputs.

Reproducibility note

The source LSP value layers contain product-specific scaling information. Before treating dhi_min as a calibrated physical quantity, users should check the release metadata and scale/offset handling for the exact source version. The ecological interpretation of MINV as a seasonal baseline proxy is separate from this numerical calibration issue.