Solar farm construction progress monitoring from satellite imagery
High-resolution optical and SAR time series let lenders and EPC managers verify ground-clearing, racking, and panel placement at utility-scale solar sites independently of contractor reporting. Change-detection on Sentinel-2 and sub-metre commercial imagery resolves individual panel rows at sites above roughly 50 MW.
Sensors
- Sentinel-2 MSI: 10 m multispectral resolution, 5-day revisit at mid-latitudes with both satellites. Sufficient to detect cleared ground and large panel blocks, but individual rows at typical 5–7 m row pitch are below the resolution floor. Free and open archive from 2015.
- Planet SuperDove: 3 m resolution, daily revisit over most land areas. Resolves panel rows clearly at sites above roughly 20 MW. Eight spectral bands including red-edge. Commercial tasking required; archive depth varies by site.
- Maxar WorldView-3: 0.3 m panchromatic, 1.24 m multispectral. Resolves individual panel rows and racking hardware. Revisit is irregular, typically 1–4.5 days depending on latitude and tasking priority. Best suited to milestone verification snapshots rather than continuous monitoring.
- Sentinel-1 SAR-C: C-band SAR at 10 m IW mode resolution, 6–12 day revisit. Acquires through cloud and at night. Backscatter change detects ground clearing and the appearance of metallic racking structures reliably, but cannot distinguish monocrystalline from thin-film panels or confirm electrical commissioning.
Why lenders want a second set of eyes
Project-finance drawdowns for utility-scale solar are typically tied to construction milestones: site cleared, foundations complete, racking installed, panels placed, grid connection ready. EPC contractors self-report against these milestones. Independent engineers visit periodically. Neither mechanism is continuous, and neither is cheap at sites that may span several hundred hectares across difficult terrain.
Satellite monitoring does not replace the independent engineer. It does give lenders and owners a dated, georeferenced record of physical progress between site visits, which is useful when a contractor's reported percentage-complete diverges from what the imagery shows. That divergence happens. It is worth having evidence.
What each construction phase looks like from orbit
Ground-clearing is the easiest phase to detect. Vegetation removal and topsoil disturbance produce a sharp increase in bare-soil reflectance visible in Sentinel-2 Band 4 (red) and a drop in NDVI. Even at 10 m, a cleared hectare is unambiguous. Sentinel-1 SAR backscatter drops when dense vegetation is removed, providing a cloud-independent confirmation.
Racking installation is the first phase where resolution matters. Aluminium or galvanised-steel racking structures are highly reflective and produce strong SAR backscatter returns. At 3 m resolution (Planet SuperDove), individual row lines become visible. At 0.3 m (WorldView-3), you can count rows and estimate row pitch. This matters because racking density is a proxy for installed capacity.
Panel placement darkens the surface sharply. Silicon panels have very low reflectance in the visible (typically 3–5%) and a characteristic spectral signature in the near-infrared. Change detection between a pre-panel and post-panel image isolates newly commissioned sections with reasonable accuracy. The limit: partial shading, soiling on newly placed panels, and cloud shadow all produce similar darkening artefacts, so any automated detection needs manual review at the margins.
Choosing the right sensor for the milestone
Not every milestone needs sub-metre imagery. Ground-clearing verification is adequately handled by Sentinel-2 at no cost, with a 5-day revisit. The economics of commercial tasking only make sense when the question requires row-level resolution, typically from racking installation onward.
SAR deserves more credit than it usually gets in this application. Monsoon-belt and equatorial sites can lose weeks of optical coverage to cloud. A Sentinel-1 pass every 6–12 days cuts through that completely. The trade-off is real: SAR cannot confirm panel type, colour, or electrical configuration. It detects metal on the ground. For lenders whose primary concern is whether physical assets exist at the claimed location, that is often enough.
A practical monitoring stack for a 200 MW site might combine Sentinel-2 for weekly change detection, Planet SuperDove for monthly row-count verification, and a WorldView-3 tasking at three or four defined milestones. Sentinel-1 fills cloud gaps. This is not a theoretical configuration; it follows directly from the published capabilities of these systems.
Change-detection methods and their honest limits
The standard approach is bitemporal differencing: subtract a baseline image (pre-construction) from each subsequent acquisition and threshold the result. Applied to NDVI or a bare-soil index, this cleanly separates cleared from uncleared land. Applied to a panel-darkness index in the red and near-infrared, it separates panelled from unpanelled sections.
More sophisticated methods use machine-learning classifiers trained on labelled examples of cleared earth, racking, and placed panels. Published work in Remote Sensing (MDPI) and similar journals has demonstrated panel-row detection at 3 m resolution using convolutional neural networks, though training data requirements and site-to-site generalisation remain active research problems. At 10 m, pixel-based classifiers perform well for phase detection but cannot resolve row geometry.
Cloud is the main operational limit for optical methods. A site under heavy monsoon cloud may produce no usable optical image for 4–8 weeks. SAR mitigates this but does not eliminate ambiguity at the panel-placement phase. Reporting latency from tasking to delivered analysis is typically 24–72 hours for commercial constellations, longer for archive requests. Sites below roughly 10 MW are difficult to monitor at Sentinel-2 resolution because the installation footprint approaches the sensor's detection floor.
Turning imagery into a progress certificate
Raw change-detection output is not a progress certificate. It needs to be reconciled against the project's construction schedule, the site boundary polygon, and the milestone definitions in the EPC contract. The analytic product that is actually useful to a lender is a georeferenced map showing which sub-sections of the site have reached each milestone, with a dated image as evidence, a confidence rating, and a flag for any area where cloud or shadow prevented assessment.
Satellize structures this kind of analysis as a periodic report tied to drawdown events, with GIS layers that an independent engineer can overlay on their own site surveys. The workflow is the same one applied in the Tonga crop-estimation programme: open constellation data as the baseline, commercial tasking added where resolution demands it, and a human analyst reviewing algorithmic output before anything goes to a client.
The archive matters too. If a dispute arises eighteen months after commissioning about whether a section was complete at a specific drawdown date, a dated satellite image is admissible evidence in a way that a contractor's spreadsheet is not. Sentinel-2 archive runs from 2015; Landsat from 1972. Commercial archives vary but typically extend three to five years.
Typical figures
| Spatial resolution (optical) | 10 m (Sentinel-2), 3 m (Planet SuperDove), 0.3 m pan / 1.24 m MS (WorldView-3) |
| Spatial resolution (SAR) | 10 m IW mode (Sentinel-1 C-band) |
| Revisit frequency | 5 days (Sentinel-2, mid-latitude); daily (Planet SuperDove); 6–12 days (Sentinel-1); 1–4.5 days (WorldView-3, tasked) |
| Analysis latency | 24–72 hours from image acquisition for commercial tasking; near-real-time for open constellations |
| Spectral bands used | Visible, NIR, red-edge (optical); C-band 5.4 GHz (SAR) |
| Minimum detectable site size | ~1 ha cleared area (Sentinel-2); row-level detection from ~20 MW at 3 m; row counting from ~5 MW at 0.3 m |
| Cloud penetration | SAR only; optical methods are cloud-limited |
| Archive depth | Sentinel-2 from 2015; Landsat from 1972; commercial archives typically 3–5 years |
| Delivery formats | GeoTIFF change maps, GeoJSON milestone polygons, PDF milestone report with dated imagery |
| Coverage | Global land; Sentinel-1 and Sentinel-2 cover most inhabited latitudes systematically |
Analytics Satellize can run
| Ground-clearance progress map | NDVI differencing and bare-soil index thresholding on Sentinel-2 time series | GeoTIFF and GeoJSON showing cleared vs. uncleared area by date, with percentage-complete figure per contract zone |
| Racking installation front detection | SAR backscatter change detection (Sentinel-1) and high-reflectance structure mapping at 3 m (Planet) | Georeferenced polygon showing active installation front with dated imagery; cloud-gap flags where SAR substitutes for optical |
| Panel placement verification | Bitemporal dark-surface change detection in red and NIR bands; CNN-based panel classifier at 3 m where training data available | Sub-section completion map with confidence ratings; flagged ambiguous areas requiring ground check |
| Milestone certificate image package | WorldView-3 tasking at defined contract milestones; analyst-reviewed output | Dated 0.3 m image with annotated row count, installed area estimate, and analyst commentary; suitable for lender file |
| Construction schedule deviation alert | Comparison of observed physical progress against client-supplied Gantt schedule; threshold-based alerting | Email alert with supporting image when observed progress lags schedule by a configurable margin |
| Full construction archive | Systematic ingestion of all available optical and SAR passes over the site boundary from groundbreak to commissioning | Timestamped image archive with change-detection layers; exportable for dispute resolution or insurance purposes |
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.