📦 Contents

Chapter 2: The Science of Remote-Sensing and Field Indicators

Chapter 1 argued that desertification can only be managed if it is measured repeatedly, comparably and honestly. This chapter looks at the measurements themselves. Every number that later travels through the WIA-ENE-057 records — an NDVI value in vegetationData, a bulk density in soilData, a soil loss rate in erosionData, an SPI value in rainfallData, a stocking rate in landUseData — has a physical meaning, a way of being observed, and a set of ways in which it can mislead. An engineer who stores these numbers without understanding them will build a system that validates perfectly and still reports the wrong story.

The chapter follows the four observable pillars introduced in Chapter 1: vegetation (2.1–2.2), soil and erosion (2.3–2.4), climate in the form of rainfall and drought (2.5), and human pressure through land use (2.6). It ends with the problem that ties them together — deciding whether a change over time is real (2.7). The running example is, as throughout the book, the location SAHEL-REGION-001 (latitude 14.5, longitude −4.0, 10,000 ha). All values quoted from the specification are illustrative example values chosen by the standard's authors; they are not measurements of any real place. There is no WIA API code in this chapter. The data-format tab of the simulator (panel 0) shows the same fields in record form for readers who want to look at them alongside the explanations.

2.1 Vegetation Indices

Why red and near-infrared

Green leaves treat two neighbouring parts of the spectrum very differently. Chlorophyll absorbs strongly in the red band (roughly 620–690 nm) to drive photosynthesis, so a healthy canopy looks dark in red. The internal cell structure of the leaf scatters near-infrared light (roughly 760–900 nm) very efficiently, so the same canopy looks bright in NIR. Bare soil, by contrast, reflects red and NIR at broadly similar, moderate levels, and water absorbs NIR almost completely. A vegetation index exploits that contrast: the larger the gap between NIR and red, the more photosynthetically active vegetation the pixel is likely to contain.

The inputs are surface reflectances — the fraction of incoming light reflected by the ground, after correction for the atmosphere — expressed as numbers between 0 and 1. Top-of-atmosphere values are a common source of silent bias, because haze brightens the red band and depresses the index.

NDVI

NDVI = (NIR − Red) / (NIR + Red)

range: −1 … +1   (dimensionless)

The normalisation by the sum makes NDVI less sensitive to overall illumination: a pixel in partial shade has lower NIR and lower red, but the ratio changes much less than either band. The theoretical range is −1 to +1. In practice, open water gives negative values, bare soil and rock give values close to zero, sparse dryland vegetation sits in the low tenths, and dense green canopy approaches the upper end of the scale. Table 2-1 summarises the typical behaviour; the ranges are approximate and depend on sensor, soil colour and season.

The worked example below uses made-up reflectances, chosen so that the result is close to the value used in the specification's vegetationData example (0.35). It is a teaching example, not an observation.

Worked example (made-up surface reflectances)
  Blue = 0.06   Red = 0.12   NIR = 0.25

  NDVI = (0.25 − 0.12) / (0.25 + 0.12)
       = 0.13 / 0.37
       ≈ 0.351

For comparison (also made up):
  bare sandy soil  Red 0.25, NIR 0.30  →  NDVI ≈ 0.09
  open water       Red 0.05, NIR 0.02  →  NDVI ≈ −0.43
  dense canopy     Red 0.04, NIR 0.45  →  NDVI ≈ 0.84

EVI

EVI = G · (NIR − Red) / (NIR + C1·Red − C2·Blue + L)

with G = 2.5, C1 = 6, C2 = 7.5, L = 1:

EVI = 2.5 · (NIR − Red) / (NIR + 6·Red − 7.5·Blue + 1)

The Enhanced Vegetation Index adds two corrections. The blue band, which is strongly affected by aerosols, is used to compensate for residual atmospheric scattering in the red band. The constant L in the denominator dampens the influence of the soil background showing through a sparse canopy. The result is an index that saturates less over dense vegetation — where NDVI flattens out close to its maximum — and that is less driven by soil brightness. Applying the formula to the same made-up reflectances gives:

EVI = 2.5 · 0.13 / (0.25 + 6·0.12 − 7.5·0.06 + 1)
    = 0.325 / (0.25 + 0.72 − 0.45 + 1)
    = 0.325 / 1.52
    ≈ 0.214

In drylands, saturation is rarely the limiting problem; soil background is. A bright, pale sandy soil can shift NDVI by several hundredths from one season to the next simply because its moisture and therefore its colour changed. That is why the standard carries both NDVI and EVI in indices rather than one of them. An implementation should not compute one from the other: they respond differently to the same land surface, and a divergence between them is itself informative.

Leaf Area Index

Leaf Area Index is the one-sided green leaf area per unit ground area, in m²/m². An LAI of 1.2, as in the specification example, means that if every leaf above one square metre of ground were laid flat, it would cover 1.2 square metres. Unlike NDVI, it is a physical quantity that can be measured on the ground. From orbit it is estimated by inverting a canopy radiative-transfer model, for example in the MODIS LAI/FPAR product MCD15A2H (500 m, 8-day). LAI matters for desertification analysis because it links to transpiration, rainfall interception and the protection the canopy offers against raindrop impact and wind.

How the indices map onto vegetationData

Each index in vegetationData.indices carries its value together with source (the sensor) and resolution. Two details deserve attention. First, the specification example writes "scale": "0-1" for NDVI and EVI, while the PHASE-1 validation rules accept NDVI between −1 and 1. The physical range is −1 to 1, and implementations should declare and validate that range; clipping negative values to zero would erase water bodies and flooded pixels. Second, many products store indices as scaled integers. MOD13Q1, for example, stores NDVI and EVI as 16-bit integers with a scale factor of 0.0001, so a stored value of 3514 means an NDVI of 0.3514. Forgetting the scale factor produces values in the thousands, which the validation rule will correctly reject; applying it twice produces values near zero, which it will not.

2.2 Satellite Products, Resolution and Cadence

PHASE-4 §4 of the specification lists the Earth-observation sources the standard expects to be integrated: MODIS on the Terra and Aqua satellites, Landsat 8/9, Sentinel-2 and SMAP. They differ in the three properties that decide what a product can and cannot tell you: spatial resolution, revisit cadence and the variables they deliver.

Resolution against the size of the site

The running example covers 10,000 ha, or 100 km². How many independent observations each product contributes to that area determines how much spatial detail the resulting record can honestly claim.

The implication is practical. A single SMAP cell cannot say anything about the difference between a restored half-moon field and the degraded rangeland next to it; it can only say whether the region as a whole is wetter or drier than usual. A 250 m MODIS pixel is often a mixture of fields, tracks and fallow; its NDVI is an area-weighted blend. A restoration project covering a few hundred hectares may be visible as a clear patch in Sentinel-2 while being lost within a handful of mixed MODIS pixels. The resolution field in the record is therefore not decoration: it tells the consumer at what scale the number may be interpreted.

Clouds, quality flags and compositing

Optical sensors cannot see through cloud, and the Sahel's wet season — the season in which vegetation is growing and most worth observing — is also the cloudiest. Products deal with this in two ways. Quality flags mark each pixel as usable or not: MOD13Q1 carries a pixel reliability layer (with codes for good data, marginal data, snow or ice, and cloud) and a more detailed VI quality bit field. Compositing combines many daily observations into one value per period; MOD13Q1 uses a constrained-view maximum-value approach over 16 days, favouring clear, near-nadir observations. The consequence is that a composite value is not an observation made on a single day: the actual acquisition date can vary from pixel to pixel within the 16-day window, and MOD13Q1 records it in a separate composite-day layer.

The record's metadata.cloudCover (5 in the example) and metadata.quality fields summarise this for the whole location. An ingest pipeline should compute them from the pixel flags rather than copy a scene-level cloud estimate, and it should decide explicitly whether pixels flagged as marginal are included in the zonal statistics. Both choices change the resulting NDVI.

From product to record: the ingest metadata

PHASE-3 §3.2 defines how a raster enters the system: a multipart upload to /api/v1/ingest/satellite with a JSON metadata part (satellite, instrument, product, acquisitionDate, resolution, bands, cloudCover, crs) followed by the GeoTIFF itself, with NetCDF as an alternative. The protocol mechanics belong to Chapter 5; here the scientific point is what that metadata must capture. Three observations follow from the preceding discussion:

• Native projection differs from the declared CRS. MODIS land products are distributed in HDF-EOS files on a sinusoidal grid, while the example metadata declares EPSG:4326. Reprojection is step 3 of the PHASE-4 §4.2 pipeline and involves resampling; the record should reflect the grid on which statistics were actually computed.
• A composite has a period, not an instant. An acquisitionDate of 25 December for a 16-day product is ambiguous. It is good practice to record the composite start and end dates alongside it.
• Sensor continuity must be planned. The MODIS instruments are long past their design lifetimes, and the VIIRS instruments on newer platforms provide continuity vegetation-index products. A multi-year trend that switches source must record the switch (see 2.7).

Satellites also need ground truth: field plots with photographs and cover estimates are the only way to check that an index trend reflects vegetation rather than soil colour, dust or crop residue. The same vegetationData shape carries both kinds of source, distinguished by metadata.sensorType.

2.3 Soil Indicators

Vegetation responds within weeks; soil changes over years and recovers over decades. That asymmetry makes soil indicators central to deciding whether land is degrading, even though they are observed far less often. Table 2-4 maps the soil indicators in PHASE-1 §2.1 to what they signal and how they are usually obtained.

Texture

Texture describes the proportions of sand (2–0.05 mm in the USDA system), silt and clay. The example's 65 % sand, 25 % silt and 10 % clay falls in the sandy loam class, which matches its classification. Texture barely changes under management, so it is best treated as a site property: sandy soils drain quickly, hold little water, and are easily moved by wind once the protective vegetation and crust are broken; silty soils in drylands tend to form surface seals under raindrop impact.

Bulk density and compaction

Bulk density is the oven-dry mass of soil divided by its total volume, including pores. As trampling by livestock and repeated traffic squeeze pores shut, bulk density rises, infiltration falls and roots find it harder to penetrate. The level at which root growth is restricted depends on texture — sandy soils tolerate higher densities than clays — so a bare threshold without texture is not meaningful. The example value of 1.45 g/cm³ should be read in the light of the 65 % sand content recorded alongside it.

Organic matter and soil organic carbon

Organic matter binds particles into aggregates, holds water and nutrients, and is the reservoir of soil carbon. Because the UNCCD's Land Degradation Neutrality framework uses soil organic carbon (SOC) stocks as one of its three sub-indicators, the relationship between the two is worth stating. Organic matter is often converted to organic carbon with the conventional factor 1.724 (equivalently, assuming organic matter is about 58 % carbon); the true ratio varies between soils, so direct SOC measurement is preferable where available. A stock per hectare then needs bulk density and depth. Using the example values purely as arithmetic:

Illustrative SOC arithmetic (example values, conventional factor)
  SOC (%)       ≈ organic matter (%) / 1.724 = 1.8 / 1.724 ≈ 1.04 %
  SOC stock     = SOC (%) × bulk density (g/cm³) × depth (cm)   [t C/ha]
  0–30 cm layer ≈ 1.04 × 1.45 × 30 ≈ 45 t C/ha
  (ignores stones and gravel; real inventories correct for coarse fragments)

The 0–30 cm depth is the reference layer used in IPCC default methods for SOC stock change and in LDN reporting. A record that gives organic matter without depth and bulk density cannot be turned into a stock, which is why the three belong together.

pH, electrical conductivity and salinization

Salinization — the accumulation of soluble salts, often driven by irrigation with poor drainage or by rising saline groundwater — is measured through the electrical conductivity of a soil extract. By a widely used convention from soil salinity science, soils with an EC of the saturated paste extract (ECe) of 4 dS/m or more are classed as saline. The convention applies to the saturated paste method; extracts with a fixed soil-to-water ratio (such as 1:5) give systematically lower readings and need conversion. The specification's ec field does not say which extract was used, so an implementation should record the method in an extension field; otherwise two values of 0.8 dS/m may not mean the same thing.

The chemistry object also contains nutrient values under a single mixed unit string ("ppm for nutrients, percentage for organic matter"). Its nitrogen example is not plausible as a ppm measurement, and it should not be read as data. The general lesson — give each quantity its own unit — returns in Chapter 3.

Water retention and available water

Available water = field capacity − wilting point

spec example:  18.5 % − 8.2 % = 10.3 %

Field capacity is the water content a soil holds after free drainage has stopped; the permanent wilting point is the water content at which plants can no longer extract water (conventionally the content at a suction of 1500 kPa). The difference is the water available to plants. The example's arithmetic is internally consistent, but the unit "percentage" hides an important distinction: water content can be gravimetric (mass of water per mass of dry soil) or volumetric (volume of water per volume of soil). Only the volumetric form converts directly into a depth of water:

If 10.3 % is volumetric, over the 0–60 cm profile:
  0.103 × 600 mm ≈ 62 mm of plant-available water

If 10.3 % is gravimetric, convert first with bulk density:
  10.3 × 1.45 ≈ 14.9 % volumetric  →  0.149 × 600 mm ≈ 90 mm

The same record gives a difference of nearly 50 % depending on an unstated convention. In-situ sensors (capacitance and time-domain reflectometry probes) report volumetric content; laboratory oven drying reports gravimetric content. The record should say which.

Depth profile

The moisture.depthProfile array records moisture and temperature by depth band (0–10, 10–30 and 30–60 cm in the example). In dry environments the surface dries within days after rain while deeper layers retain water for weeks, so a single surface value says little about what deep-rooted shrubs and trees can use. The separate surfaceLevel value (12.5 %) differs from the 0–10 cm band (15.2 %) in the example, which is reasonable if they are measured at different depths or by different methods — but the record should make that explicit rather than leave the reader to guess.

Soil crusting — a thin, dense surface layer formed by raindrop impact on bare soil — belongs to the same family of indicators, because it cuts infiltration and seedling emergence. Physical crusts are typically millimetres thick, so the specification's example thickness should not be taken as a realistic measurement.

2.4 Erosion

Erosion is the process by which degradation becomes hard to reverse: topsoil, which holds most of the organic matter and nutrients, is physically removed. Drylands suffer both water erosion, during the short intense storms of the wet season, and wind erosion, during the long dry season when soils are bare and loose.

Water erosion and RUSLE

A = R · K · LS · C · P

A   long-term average soil loss per unit area (e.g. t/ha/yr)
R   rainfall erosivity         (energy and intensity of storms)
K   soil erodibility           (texture, organic matter, structure)
LS  slope length and steepness
C   cover and management       (vegetation, residue, crusts)
P   support practices          (contouring, bunds, terraces)

The Revised Universal Soil Loss Equation, developed in the United States from plot experiments, estimates long-term average sheet and rill erosion as the product of five factors. It is used here conceptually rather than as a requirement of the standard, because it makes the drivers visible. Rainfall intensity enters through R, which is why rainfallData.intensityMetrics matters and annual totals alone do not. Soil properties from soilData enter through K. Vegetation cover enters through C, which is where restoration has its largest effect. Physical works such as stone bunds and half-moon micro-catchments enter through P. RUSLE estimates long-term averages and does not predict individual storm losses or gully erosion; applying it outside the conditions for which its factors were calibrated needs local validation.

Wind erosion

Wind moves soil when the wind's shear at the surface exceeds a threshold that depends on particle size, soil moisture, surface crusts and roughness elements such as clods, stones, stubble and plants. Empirical models such as the Wind Erosion Equation and its revised form combine climatic erosivity, soil erodibility, surface roughness, field length along the wind and vegetative cover — the same families of variables that appear in the specification.

The erosionData.factors object records exactly these drivers for a wind-erosion assessment: average and gust wind speeds (25.5 and 45.2 km/h in the example, about 7.1 and 12.6 m/s), vegetation cover (35.2 %), soil moisture (8.5 %) and a qualitative surfaceRoughness. Two points help interpretation. Wind speed depends strongly on measurement height; the meteorological convention is 10 m above ground, and a value measured at another height should say so. And the vegetation cover in erosionData need not equal vegetationData.coverage.vegetationCover for the same location: the erosion assessment may use a different date, method or definition of cover. Such differences should be documented, not silently reconciled.

Soil loss rate

Soil loss is reported in severity.soilLossRate in t/ha/yr (15.2 in the example). Converting a mass rate into a depth makes it easier to communicate. One millimetre of soil over one hectare is 10 m³; at the example bulk density of 1.45 t/m³ that is 14.5 t. The example rate therefore corresponds to roughly 1 mm of topsoil per year. That may sound small, but soil forms very slowly, which is why soil conservation literature compares loss rates with a locally defined "tolerable" rate rather than with zero.

The related field soilData.degradationIndicators.erosionRisk gives a score from 0 to 10 with a level such as "high". The specification does not define how the score is calculated, so it is a qualitative index. Implementations that populate it should document their method, for example which RUSLE or wind-erosion inputs they combine, so that scores from different providers are not compared as if they were the same quantity.

2.5 Rainfall and Drought Indices

Totals, seasonality and intensity

In the Sahel, as in many drylands, the annual total tells only part of the story. What matters for vegetation and erosion is when the rain falls and how it falls. rainfallData therefore splits the year into seasonalDistribution (in the example 285.2 mm in the June–September wet season, 87.6 % of the 325.4 mm annual total) and characterises individual events in intensityMetrics: number of rain days, the largest single event, the number of days above a heavy-rain threshold (25 mm in the example), and the average event size. Heavy events drive runoff and water erosion and recharge micro-catchments; small events wet only the surface and evaporate.

Cross-field consistency is worth checking at ingest. In the example, the seasonal totals sum correctly to the annual total, but 42 rain days at an average of 18.5 mm would imply about 777 mm, not 325.4 mm (325.4 / 42 ≈ 7.7 mm per rain day). The example is illustrative and the field definitions leave room for different conventions (for instance, averaging only events above a minimum size), but a pipeline should either derive such averages from the event list or state its definition.

Normals and deviation

A departure from "normal" is only meaningful once "normal" is defined. The World Meteorological Organization defines climatological standard normals as averages over 30-year periods, currently 1991–2020. The example record carries two references — a 10-year average of 425.8 mm and a 30-year average of 485.2 mm — and a single deviation of −23.6 %. Working it through shows which reference was used:

vs 10-year average: (325.4 − 425.8) / 425.8 ≈ −23.6 %
vs 30-year average: (325.4 − 485.2) / 485.2 ≈ −32.9 %

The two baselines tell different stories: a 10-year average that is itself lower than the 30-year average suggests a drier recent decade, so a deviation against the shorter window understates the shortfall relative to the longer climate. Records should name the reference period explicitly.

The Standardized Precipitation Index

The SPI, introduced by McKee and colleagues in 1993 and promoted by WMO as a standard meteorological drought index, turns precipitation accumulated over a chosen time scale (1, 3, 6, 12 months and so on) into a standard normal value. A long record of accumulations is fitted with a probability distribution — usually the gamma distribution — and each observation's cumulative probability is mapped to the equivalent z-score. An SPI of −1.0 means the accumulation was about as unusual, on the dry side, as a value one standard deviation below the mean of a normal distribution. Because it is standardised, SPI is comparable between a wet and a dry climate. McKee's classes are:

Under this scheme the example value spi: -1.8 falls in the severely dry class. The category label printed next to it in the specification example does not match McKee's classes, and implementations should derive the category from the value using a documented table rather than copying a label. Two further fields are needed to interpret an SPI value at all: the accumulation time scale (a 3-month SPI describes agricultural stress; a 12-month SPI describes hydrological conditions) and the calibration period over which the distribution was fitted. Neither is in the example record; both belong in it.

The Palmer Drought Severity Index (PDSI), which the standard also mentions, is built differently: it runs a simple soil-water balance driven by precipitation and temperature and expresses the result as an index. It captures heat-driven evaporative demand, which SPI ignores, but was calibrated for the central United States; where the two disagree, keep both.

2.6 Land Use and Human Pressure

The fourth pillar is human pressure. landUseData.classification divides the location into classes — agriculture, grazing, forest, barren, water and built-up land — each with an area and a percentage. In the example they add up to the full 10,000 ha (4,500 + 3,200 + 1,200 + 800 + 200 + 100), which is the first check any implementation should perform. The classification is normally derived from supervised classification of Landsat or Sentinel-2 imagery combined with field knowledge, and it is only as good as the class definitions: "barren" land and "degraded grazing" land may be spectrally similar and distinguishable only by use.

Stocking rate

Stocking rate = animal units / grazing area

spec example: 2,400 AU / 3,200 ha = 0.75 AU/ha
              (equivalently ≈ 1.33 ha per animal unit)

Stocking rate expresses how many animals graze a given area over a period. It is only meaningful with a defined animal-unit convention: rangeland science in North America uses an animal unit based on a mature cow of a reference weight, while work in sub-Saharan Africa often uses the Tropical Livestock Unit (TLU), defined as an animal of 250 kg live weight, with conversion factors for cattle, sheep, goats and camels. A record that says 0.75 without the convention cannot be compared with another. What stocking rate means for the land depends on carrying capacity, which in drylands varies strongly from year to year with rainfall: the same herd can be sustainable after a good wet season and damaging after a poor one. That is why the rainfall and land-use records are kept separate but joinable through locationId.

Population, fuelwood and overall pressure

The pressure object carries populationDensity, livestockDensity, fuelwoodCollection and overallPressure. The density fields in the example have no unit, so it is not clear whether they refer to people or animals per km² or per hectare; implementations should add explicit units. Fuelwood collection is given qualitatively ("high"), which is realistic, since wood harvesting is rarely measured directly. Management practices (tillage, irrigation, soilConservation) complete the picture and are what restoration projects change.

2.7 Trend Detection

Every indicator above is ultimately used to answer one question: is this land getting worse, staying the same or recovering? In the record this answer appears in vegetationData.trendAnalysis: monthlyChange, annualChange, fiveYearTrend, a status of declining, stable or improving, and a confidence. The specification does not prescribe a method, so each implementation must choose one and document it.

State the unit of change

The example gives fiveYearTrend: -0.42 for a current NDVI of 0.35. If that were an absolute change in NDVI units, the site would have had an NDVI around 0.77 five years ago — high for a Sahelian landscape. If it is a relative change, it means a 42 % decline. The specification does not say which. An implementation should state whether change values are absolute index units or fractions, and over which window they are computed.

Seasonality and compositing

Dryland vegetation has a strong annual cycle: green-up with the rains, peak in late wet season, senescence in the dry season. A month-to-month change of −0.05 in December is ordinary senescence, not degradation. Trends should therefore be computed on deseasonalised data — for example annual integrals of NDVI over the growing season, anomalies against the same period in previous years, or a seasonal variant of the trend test. Mixing 16-day composites with daily or monthly data in one series without aligning periods produces spurious steps.

Non-parametric tests

Two non-parametric methods are widely used in practice for vegetation and climate time series, because they do not assume normally distributed values and are robust to outliers. They are described here as common practice, not as a WIA requirement.

Mann–Kendall statistic for a series x1 … xn:
  S = Σ over all i < j of sign(xj − xi)
  S > 0 suggests an increasing trend, S < 0 a decreasing one;
  its significance is judged against the distribution of S
  under the hypothesis of no trend.

Theil–Sen slope:
  slope = median over all i < j of (xj − xi) / (tj − ti)

Mann–Kendall answers whether a monotonic trend is present; the Theil–Sen estimator gives its magnitude. Serial correlation in the series inflates apparent significance and needs correction, and a short series — five annual values, say — has little power whichever test is used. A reasonable interpretation of the confidence field is the complement of the test's p-value or a similar measure; whatever it is, the method should be recorded, because 0.87 from one method is not comparable with 0.87 from another.

Sensor changes

A long series built from several sensors, or from different processing versions of the same product, can contain steps that look like trends. Band definitions differ between Landsat sensors and between Landsat and Sentinel-2; and data providers periodically reprocess archives with new calibration. Harmonised products, such as NASA's Harmonized Landsat and Sentinel-2 dataset, reduce but do not remove these differences. The source field on each index is what makes it possible to detect a switch after the fact.

Rainfall-driven greening versus real recovery

The hardest problem is attribution. In drylands, vegetation productivity follows rainfall closely, so a run of wetter years produces greening even on land whose condition has not improved, and a drought produces browning even on well-managed land. A common approach in the research literature is to model the expected index from rainfall — for example by regressing annual NDVI integrals on annual or seasonal rainfall — and to analyse the trend in the residuals (often called RESTREND). A negative residual trend suggests that the land is producing less vegetation for the same rain than it used to, which is closer to what degradation means. A related indicator is rain-use efficiency: productivity per unit of rainfall. The UNCCD's good practice guidance for SDG indicator 15.3.1 similarly treats land productivity through several metrics rather than one raw trend.

For a data standard, the lesson is structural. A trend claim is only as credible as the joins behind it: vegetation from vegetationData, rainfall from rainfallData and land use from landUseData, all tied to the same locationId and the same periods. That is why the standard keeps them as separate record types rather than a single summary number, and why Chapter 3 spends so much attention on identifiers and time stamps.

Chapter Summary

• NDVI and EVI measure the contrast between red absorption and NIR scattering by green leaves. Both range from −1 to 1; the specification example's "0-1" scale label should not be followed. EVI corrects for aerosols and soil background and saturates less.
• Product choice is a choice of scale: SMAP's 9 km grid gives regional context, MODIS 250 m gives long and frequent series, Landsat 30 m and Sentinel-2 10 m resolve fields and restoration plots. Record the source, resolution, quality handling and composite period with every value.
• Soil indicators change slowly but decide whether degradation is reversible. Record whether water contents are volumetric or gravimetric, which extract an EC value uses, and the depth and bulk density needed to turn organic matter into an SOC stock.
• RUSLE and wind-erosion models explain why rainfall intensity, soil properties, cover and conservation practices appear in erosionData. A soil loss of 15.2 t/ha/yr at a bulk density of 1.45 g/cm³ is about 1 mm of topsoil per year.
• Rainfall must be judged against a named reference period; SPI must carry its time scale and calibration period, and its category must follow McKee's classes (−1.8 is severely dry).
• Stocking rate requires an animal-unit convention, and land-use areas must add up to the location's total.
• Trend claims need an explicit unit of change, deseasonalisation, a documented statistical method, a record of sensor switches, and a way to separate rainfall-driven greening from real recovery.

Chapter 3 turns from meaning to structure: it walks through the WIA-ENE-057 records field by field, shows how they join on locationId, and implements the validation rules that catch several of the problems identified here.