Hyperspectral inland water constituent retrieval

Narrow-band hyperspectral imagery from PRISMA and EnMAP can separate chlorophyll-a, phycocyanin, coloured dissolved organic matter and suspended sediment in lakes and reservoirs where standard multispectral sensors produce ambiguous or meaningless results.

Sensors

PRISMA (ASI): Italian Space Agency hyperspectral mission, 30 m ground sampling distance, 239 spectral bands from 400 to 2500 nm at roughly 10 nm spectral resolution, single-track revisit approximately 29 days at the equator but targetable for off-nadir acquisitions. Operational since 2019.
EnMAP (DLR/GFZ): German hyperspectral mission, 30 m GSD, 228 spectral bands across 420 to 2450 nm at 6.5 to 10 nm spectral resolution, swath 30 km, revisit 27 days at nadir but steerable to 4 days. Launched April 2022, science data publicly available from late 2022.
DESIS (DLR on ISS): VNIR-only hyperspectral sensor mounted on the International Space Station, 235 bands from 400 to 1000 nm at approximately 2.5 nm spectral sampling, 30 m GSD, 30 km swath. ISS inclination of 51.6 degrees limits coverage to mid-latitudes. Revisit is irregular and ISS-orbit-dependent.
PACE OCI (NASA): Ocean Colour Instrument on the PACE satellite launched February 2024, hyperspectral from 340 to 890 nm at 5 nm resolution but at 1 km pixel size. Suited to large lakes and coastal transitions rather than small inland water bodies; daily global coverage.

Why broadband sensors fail in optically complex water

Inland lakes, reservoirs and rivers are what bio-optical scientists call Case 2 waters: the optical signals of multiple constituents overlap so severely that a sensor with only a handful of broad spectral bands cannot disentangle them. Chlorophyll-a absorbs strongly near 440 nm and 676 nm but also scatters. Phycocyanin, the pigment diagnostic of cyanobacterial blooms, has a distinctive absorption feature near 620 nm that sits between chlorophyll peaks. Coloured dissolved organic matter (CDOM) absorbs exponentially from the UV into the visible, mimicking turbidity. Suspended sediment scatters broadly across the visible and near-infrared. A Sentinel-2 or Landsat band straddles all of these simultaneously.

Hyperspectral instruments resolve this by sampling the full reflectance spectrum at 6 to 10 nm intervals, giving enough spectral degrees of freedom to fit a bio-optical model to the data rather than relying on a two-band ratio that assumes one constituent dominates. The difference is not cosmetic. A reservoir with a moderate sediment load and an early cyanobacterial bloom can return a Sentinel-2 chlorophyll index that is inflated by a factor of three or more, leading water managers to miss or misclassify a developing public-health event.

The adjacency effect: when the surrounding land contaminates the water pixel

At 30 m GSD, the atmosphere scatters photons reflected from the land surrounding a water body into the instantaneous field of view of water pixels near the shoreline. This adjacency effect is proportionally larger for water than for land because water is dark: the contaminating land signal can represent 30 to 50 percent of the total at-sensor radiance in the first one or two pixels inward from the shore. Standard atmospheric correction algorithms designed for open ocean or land surfaces do not correct for this.

Dedicated correction schemes, notably SIMEC (Similarity Environment Correction) and the iterative approaches embedded in ACOLITE and POLYMER, use the spatial structure of the scene to estimate and subtract the adjacency contribution. For PRISMA and EnMAP data, ESA's SNAP toolbox and the dedicated EnMAP processing chain both implement variants of these methods. Even so, pixels within roughly 60 to 90 m of a vegetated or bright shoreline should be treated with caution regardless of the correction applied. This is not a solvable problem at 30 m; it is a physical constraint of the geometry.

Bio-optical inversion: from reflectance spectrum to constituent concentration

Once surface reflectance is retrieved, the inversion step maps the spectrum to concentrations of chlorophyll-a, phycocyanin, CDOM absorption coefficient and total suspended matter. Three broad families of methods are in operational use. Physics-based inversion fits a parameterised radiative-transfer model (typically the quasi-analytical algorithm, QAA, or the SIOP-based matrix inversion approach) to the measured spectrum. It is interpretable and transferable across water bodies but requires accurate specific inherent optical properties (SIOPs) for the target lake, which vary seasonally. Machine learning methods, particularly Gaussian process regression and neural networks trained on in-situ spectral libraries, can outperform physics-based approaches when training data are available but generalise poorly outside the range of conditions they were trained on.

Mixture methods, which combine a forward model with a learned correction, represent the current research frontier. For phycocyanin specifically, the 620 nm absorption feature is narrow enough that only hyperspectral data can reliably detect it at concentrations below roughly 20 micrograms per litre. Sentinel-2 band 3 is 60 nm wide and simply cannot resolve the feature. Published studies using PRISMA over Italian lakes and EnMAP over German reservoirs have demonstrated retrieval uncertainties of approximately 20 to 35 percent for chlorophyll-a and 25 to 45 percent for phycocyanin under clear-sky conditions, with performance degrading significantly when the water surface is wind-ruffled or when thin cirrus is present.

The minimum water body problem

Thirty metres is the practical floor for current spaceborne hyperspectral instruments. A pixel is usable for water-quality retrieval only when the water fraction within it exceeds roughly 90 percent, otherwise mixed-pixel effects dominate. That threshold places the practical minimum water body width at approximately 90 to 120 m, meaning lakes smaller than around 1 to 2 hectares are effectively invisible to PRISMA and EnMAP for constituent retrieval purposes. Many ecologically important farm dams, urban ponds and small highland lakes fall below this limit.

Airborne hyperspectral systems such as HySpex or APEX operate at 1 to 5 m GSD and have no such constraint, but they require dedicated flight campaigns and cannot provide the repeat coverage that satellite data offers. The next generation of spaceborne hyperspectral missions, including the planned SBG (Surface Biology and Geology) instrument from NASA and the CHIME mission from ESA, are expected to maintain the 30 m class GSD. Pixel size is unlikely to improve dramatically in the near term from operational free-data sources.

Cloud, revisit and the operational reality for water managers

A 27 to 29-day revisit cycle is the single largest operational limitation of PRISMA and EnMAP for inland water monitoring. Cyanobacterial blooms can develop, peak and partially dissipate within days. A monthly acquisition cadence means that a sensor may capture the before and after but miss the event entirely. DESIS improves this somewhat for mid-latitude sites through ISS orbital geometry but cannot be relied upon for scheduled revisit. PACE OCI provides daily global coverage but at 1 km resolution, which is useful for large lakes such as the African Great Lakes or Lake Baikal but not for most managed reservoirs.

Cloud cover compounds the revisit problem. In temperate and tropical regions, cloud-free acquisitions over a specific water body may average only four to eight per year even with tasking flexibility. Operational programmes therefore typically combine hyperspectral acquisitions for calibration and detailed constituent mapping with higher-frequency multispectral data from Sentinel-2 or Landsat for trend monitoring between hyperspectral overpasses, accepting the reduced accuracy of the broadband data in exchange for temporal density.

Satellize's analytics work for the Kingdom of Tonga crop-estimation programme demonstrated this kind of sensor fusion logic in an agricultural context; the same multi-source temporal infilling approach applies directly to inland water monitoring programmes.

What a practical monitoring programme looks like

A credible inland water constituent retrieval programme starts with in-situ sampling to characterise the SIOPs of the specific water body. Without local optical property measurements, physics-based inversions carry large and poorly quantified uncertainties. That field work is not optional; it is the calibration anchor for everything that follows.

Processing pipelines then apply atmospheric correction (ACOLITE or the EnMAP L2 processor), adjacency correction, quality flagging for cloud, sun glint and adjacency-contaminated shoreline pixels, and bio-optical inversion. Outputs are typically delivered as georeferenced GeoTIFF maps of chlorophyll-a concentration in micrograms per litre, phycocyanin index, CDOM absorption at 440 nm, and total suspended matter in grams per cubic metre, together with per-pixel uncertainty estimates. Alert thresholds for bloom detection can be set against WHO or national drinking-water guidelines, with automated notifications triggered when pixel-averaged chlorophyll-a exceeds a defined threshold across a minimum contiguous area.

Typical figures

Spatial resolution (PRISMA, EnMAP, DESIS)	30 m ground sampling distance
Spectral range	PRISMA: 400–2500 nm (239 bands); EnMAP: 420–2450 nm (228 bands); DESIS: 400–1000 nm (235 bands)
Spectral sampling interval	EnMAP: 6.5–10 nm; PRISMA: ~10 nm; DESIS: ~2.5 nm (VNIR only)
Nominal revisit (nadir)	PRISMA: ~29 days; EnMAP: 27 days (steerable to ~4 days); DESIS: irregular, ISS-dependent
Minimum detectable water body width	~90–120 m (3–4 pixels) for reliable constituent retrieval; smaller bodies are mixed-pixel dominated
Phycocyanin detection limit (published studies)	~20 µg/L under clear-sky, low-wind conditions; higher in optically complex or wind-ruffled water
Chlorophyll-a retrieval uncertainty	~20–35% for well-characterised water bodies with local SIOP data; higher without in-situ calibration
Cloud sensitivity	Optical only; cloud and thin cirrus render acquisitions unusable; no all-weather capability
Archive depth	PRISMA: from 2019; EnMAP: from late 2022; DESIS: from 2018 (ISS deployment)
Delivery formats	L2 surface reflectance (HDF5, GeoTIFF); derived constituent maps as GeoTIFF with uncertainty layers

Analytics Satellize can run

Chlorophyll-a concentration map	Bio-optical inversion (QAA or SIOP-based matrix inversion) applied to atmospherically corrected surface reflectance	Georeferenced GeoTIFF in µg/L with per-pixel uncertainty estimate, per acquisition
Phycocyanin index and cyanobacterial bloom alert	620 nm absorption feature fitting; threshold alert against WHO recreational or drinking-water guideline values	Bloom extent polygon (GIS layer) and automated alert report with affected area in hectares
CDOM absorption coefficient map	Spectral deconvolution of exponential UV-visible absorption; Gaussian process regression where training data available	Raster map of a(CDOM) at 440 nm (m⁻¹), delivered per scene
Total suspended matter concentration	Near-infrared scattering inversion; empirical or semi-analytical model calibrated to local in-situ data	GeoTIFF in g/m³; time-series chart for defined monitoring zones
Temporal trend report (multi-sensor infill)	Hyperspectral acquisitions used as calibration anchors; Sentinel-2 or Landsat time series infilled using cross-calibrated band ratios	Monthly PDF summary with constituent trend plots and anomaly flags for each monitored water body
Adjacency-corrected shoreline quality assessment	SIMEC or ACOLITE adjacency correction; pixel-level flagging of shoreline contamination zone	Quality-flagged raster mask indicating which pixels meet retrieval confidence thresholds

Sources

Who does the work

We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.