Agrivoltaic system crop-stress co-monitoring
Bifacial solar arrays create spatially heterogeneous shading that ground sensors miss entirely. Satellite-derived vegetation indices from Sentinel-2 red-edge bands and Planet SuperDove resolve inter-row and under-panel zones separately, but only after correcting for the spectral contamination that panel shadows introduce into surface-reflectance retrievals.
Sensors
- Sentinel-2 MSI: 10 m visible and NIR bands; 20 m red-edge bands (B5 705 nm, B6 740 nm, B7 783 nm) used for NDRE and red-edge LAI retrieval; 5-day revisit at mid-latitudes with both satellites. The 20 m red-edge pixels are coarse relative to typical 2–4 m inter-row spacing, so sub-pixel mixing between shaded and sunlit zones is unavoidable and must be modelled.
- Planet SuperDove: 8-band imagery at 3.7 m native resolution including a dedicated red-edge band near 705 nm. Daily revisit over most land surfaces. The finer ground sampling resolves individual inter-row strips in installations with row spacing above roughly 4 m, reducing mixed-pixel contamination substantially compared with Sentinel-2 red-edge.
- Landsat 8/9 OLI: 30 m multispectral; useful for long time-series baselines (Landsat 8 from 2013, Landsat 9 from 2021) and for establishing pre-installation crop-health benchmarks. Too coarse to distinguish inter-row from under-panel zones in most agrivoltaic layouts; best used for trend context rather than spatial disaggregation.
- MODIS MOD13 (Terra/Aqua): 250 m NDVI and 500 m EVI products at 16-day composites. Entirely unable to resolve individual agrivoltaic rows. Useful only as a regional agricultural background reference or for detecting severe, field-scale stress events.
What the panels do to the pixels
An agrivoltaic installation is not a uniform surface. Elevated bifacial panels, typically mounted 2–4 m above the crop, cast moving shadow bands whose width and position shift with solar angle throughout the day. At the moment a satellite passes over, some crop rows sit in direct sun, others are fully shaded, and the panel backs themselves occupy a fraction of each pixel. A Sentinel-2 red-edge pixel at 20 m may contain all three surface types simultaneously.
The practical consequence is spectral mixing. Reflectance recorded by the sensor is a weighted average of sunlit crop canopy, shaded crop canopy, and panel material. Panel backs are typically dark polymer or glass with low NIR reflectance, which suppresses the apparent red-edge step that NDRE and red-edge LAI estimators depend upon. Without correction, indices computed naively from top-of-atmosphere or even atmospherically corrected reflectance will underestimate vegetation vigour in shaded zones and produce artefacts that look like crop stress even when the plants are healthy.
Correcting for shadow geometry before computing indices
Shadow correction in agrivoltaic scenes requires knowing the panel geometry precisely: row azimuth, tilt angle, height, and inter-row pitch. These parameters, combined with the solar zenith and azimuth at the satellite overpass time, allow the shadow footprint to be modelled deterministically for each acquisition. The shaded fraction of each pixel can then be estimated and used to unmix the composite reflectance signal before index computation.
Several published approaches exist. Linear spectral unmixing assigns end-member spectra to sunlit crop, shaded crop, and panel material, then solves for fractional cover. A simpler but less accurate alternative is to mask pixels where the modelled shadow fraction exceeds a threshold, typically 30–40 %, and interpolate from neighbouring sunlit pixels. Both methods introduce their own errors. Unmixing requires accurate end-member spectra that vary with crop type and phenological stage. Masking discards data and can leave entire rows uncharacterised on days when the sun angle is low. Neither approach eliminates uncertainty entirely; analysts should report confidence intervals alongside derived indices rather than treating corrected values as ground truth.
Planet SuperDove's 3.7 m resolution changes the geometry of the problem. In installations with inter-row spacing of 5 m or more, individual sunlit and shaded strips can be addressed as separate pixels rather than mixed fractions. This allows direct comparison of NDVI or NDRE between zones without unmixing, though the panel shadow boundary still contaminates pixels that straddle the edge.
Which indices carry useful signal in this setting
NDVI (using red and NIR bands) remains the most widely computed index, but it saturates at moderate-to-high canopy cover, typically above a leaf area index of around 3. For dense crops beneath panels, NDRE computed from the Sentinel-2 red-edge bands at 705 nm and 783 nm is more sensitive in the upper range of vigour and less susceptible to soil background effects in sparse canopies. Published work on precision agriculture has shown NDRE to correlate more strongly with chlorophyll content than NDVI across a range of crop types.
LAI retrieval using the Sentinel-2 red-edge bands is supported by ESA's biophysical processor, which applies a neural-network inversion of canopy radiative transfer models trained on PROSAIL simulations. The processor outputs LAI, FAPAR, and canopy chlorophyll content at 20 m. In agrivoltaic contexts, these outputs should be treated as indicative rather than precise: the processor was not trained on mixed panel-crop scenes, and shadow contamination that survives atmospheric correction will propagate into the biophysical estimates.
The most diagnostic use of multi-index time series is comparative: tracking NDRE or LAI in the inter-row zone against the under-panel zone across a growing season. Persistent divergence, where under-panel values remain 15–25 % below inter-row values through periods when water stress is unlikely, suggests light limitation rather than water or nutrient stress. Temporary convergence after rainfall, by contrast, may indicate that the shaded zone is benefiting from reduced evapotranspiration, an effect documented in several published agrivoltaic trials in semi-arid climates.
Honest limits of the satellite view
Cloud cover is the dominant operational constraint. Sentinel-2 acquires cloud-free imagery over a given site roughly every 5–10 days in temperate climates, fewer in persistently cloudy regions. Planet's higher revisit rate improves this, but commercial tasking costs rise quickly when dense time series are required over multiple sites.
The 20 m floor on Sentinel-2 red-edge resolution means that agrivoltaic installations with inter-row spacing below about 6 m will always produce mixed pixels in those bands, regardless of correction method. For smaller-scale or experimental installations, satellite indices can characterise field-level averages but cannot reliably attribute stress to specific structural zones. Ground-based or UAV hyperspectral surveys remain necessary for sub-metre spatial diagnosis.
Satellite data also cannot distinguish between the causes of spectral anomalies. Low NDRE in an under-panel zone could reflect light limitation, water stress (if drainage is altered by panel runoff), pest pressure, or soil compaction from installation equipment. Satellite indices narrow the list of hypotheses; they do not resolve it. Integrating soil moisture products, precipitation records, and agronomic knowledge is necessary before any management intervention is justified.
From indices to operational insight
The practical output of a well-designed monitoring programme is a seasonal map series showing the spatial distribution of crop vigour across the agrivoltaic installation, with inter-row and under-panel zones disaggregated where pixel resolution permits. Overlaying this with panel-layout geometry allows operators to identify whether stress patterns align with structural shading or are spatially random, which points toward different causes.
Satellize applies this type of vegetation-index time-series analysis using open Sentinel-2 data and commercial Planet tasking, drawing on methods developed through its crop-estimation work including the Kingdom of Tonga programme. For agrivoltaic operators, the concrete deliverable is a GIS layer updated on each cloud-free acquisition, showing NDRE and shadow-corrected LAI by zone, with anomaly flags where values fall more than one standard deviation below the site's own historical baseline for that calendar week.
Operators who want to benchmark a new installation against pre-deployment crop performance can use the Landsat archive, which extends to 2013 for OLI-quality data, to establish what the field produced before panels were installed. That baseline comparison is often the most commercially useful output: it tells an agrivoltaic developer whether the crop yield penalty from shading is within the range the system was designed to tolerate.
Typical figures
| Spatial resolution (visible/NIR) | 10 m (Sentinel-2), 3.7 m (Planet SuperDove), 30 m (Landsat 8/9) |
| Spatial resolution (red-edge) | 20 m (Sentinel-2 B5/B6/B7), 3.7 m (Planet SuperDove red-edge band) |
| Revisit frequency | 5 days at mid-latitudes (Sentinel-2A+B combined); daily (Planet SuperDove); 16-day (Landsat 8/9) |
| Key spectral bands | Red-edge 705 nm, 740 nm, 783 nm (Sentinel-2); red-edge ~705 nm (SuperDove); NIR, red, green |
| Minimum resolvable zone width | ~6 m inter-row strip for Sentinel-2 red-edge; ~4 m for Planet SuperDove |
| Biophysical products | NDVI, NDRE, LAI, FAPAR, canopy chlorophyll content (ESA biophysical processor at 20 m) |
| Archive depth | Sentinel-2 from 2015; Landsat from 1972 (OLI quality from 2013); Planet from 2016 (variable) |
| Cloud-free acquisition frequency (temperate) | Typically every 5–10 days for Sentinel-2; higher with Planet but subject to tasking |
| Latency from acquisition to product | Sentinel-2 L2A available within 1–3 hours of overpass via Copernicus Data Space |
| Delivery formats | GeoTIFF raster layers, GeoPackage vector zone statistics, CSV time-series tables |
Analytics Satellize can run
| Shadow-fraction map per acquisition | Solar geometry modelling using panel tilt, azimuth, height, and satellite overpass solar angles to compute shaded pixel fractions | GeoTIFF raster at native sensor resolution, updated each cloud-free pass |
| Shadow-corrected NDRE layer | Linear spectral unmixing using sunlit-crop, shaded-crop, and panel end-member spectra derived from site-specific imagery | GeoTIFF NDRE layer with per-pixel uncertainty estimate; zone-average statistics in CSV |
| LAI time series by structural zone | ESA biophysical processor (neural-network PROSAIL inversion) applied to Sentinel-2 red-edge bands, with shadow-contaminated pixels flagged | Seasonal LAI time-series chart and GIS layer disaggregated by inter-row and under-panel zones |
| Crop-stress anomaly alerts | Z-score comparison of current NDRE against site historical baseline for the same calendar week; threshold set at minus one standard deviation | Automated alert report with flagged zones, magnitude of deviation, and candidate cause list |
| Pre/post-installation yield-proxy benchmark | Landsat 8/9 OLI NDVI time series from archive years before panel installation compared with current Sentinel-2 NDRE series using phenology-matched composites | PDF benchmark report with seasonal vigour curves and estimated canopy-cover change |
| Spatial stress attribution map | Overlay of vegetation-index anomaly zones with panel-layout geometry to classify stress as structurally correlated (shading) or spatially random (other causes) | GeoPackage layer with stress-attribution classification; input to agronomic management recommendations |
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.