Smallholder field boundary delineation in fragmented landscapes
Deep-learning segmentation on very-high-resolution imagery can delineate farm parcels well below one hectare in landscapes where no official boundary records exist, enabling farm-level insurance, subsidy targeting and yield estimation across sub-Saharan Africa and South Asia.
Sensors
- Planet SuperDove: 3 m native resolution, 8 spectral bands including red-edge and NIR, daily revisit at the equator. Sufficient for fields down to roughly 0.1 ha in open terrain; canopy overhang at boundaries degrades edge precision.
- Maxar WorldView-3: 0.3 m panchromatic, 1.24 m multispectral (8 VNIR bands plus 8 SWIR). The sharpest commercially available optical archive for boundary delineation; revisit is 1–4.5 days depending on latitude and tasking priority, not guaranteed daily.
- Airbus Pléiades Neo: 0.3 m panchromatic, 0.7 m multispectral, 4 bands plus a dedicated red-edge channel. Constellation of two satellites gives roughly 1–2 day revisit with tasking; stereo acquisition supports canopy-height estimation at field edges.
- ESA Sentinel-2: 10 m in visible and NIR bands, 5-day revisit with both satellites. Too coarse to delineate sub-hectare parcels individually, but dense time series are used to build seasonal NDVI composites that sharpen spectral contrast between crop types at boundaries, improving segmentation when fused with VHR imagery.
Why half a hectare is invisible to most satellites
The average smallholder plot in Ethiopia, Bangladesh or Malawi is often between 0.2 and 0.8 hectares. At Sentinel-2's 10 m resolution, a 0.5 ha square field occupies roughly 50 pixels, which sounds adequate until you account for the mixed pixels along every boundary. A single pixel straddles the crop, the footpath and the neighbour's sorghum simultaneously. Segmentation algorithms trained on European field data, where parcels run to tens of hectares, fail badly when applied here.
Very-high-resolution imagery at 0.3 to 3 m changes the geometry. A 0.5 ha field now contains hundreds to thousands of pixels, and the boundary zone, even at 3 m, is only two or three pixels wide rather than the dominant feature. That is still a hard problem, but a tractable one.
What a spectral gradient actually reveals at a field edge
Field boundaries in fragmented landscapes are not sharp lines. They are zones: a berm, a footpath, a strip of unmanaged vegetation, sometimes just a change in tillage direction. Deep-learning segmentation models trained on these environments learn to read spectral gradients across those transition zones rather than looking for a hard reflectance step.
Seasonal compositing helps considerably. A single-date image may catch two adjacent fields at the same crop stage, making them spectrally identical. A composite built from four or five acquisitions across a growing season captures the phenological divergence between, say, maize and groundnuts, even when both look green in any single image. Red-edge bands, present in WorldView-3, Pléiades Neo and SuperDove, are particularly useful here because chlorophyll absorption features shift with crop type and growth stage in ways that broadband NIR does not resolve.
Homesteads, trees and paths are the main confounders. Acacia canopies cast shadows that mimic boundary transitions. Compound walls produce edges that a boundary model will happily trace as field limits. Post-processing rules, typically minimum-area filters and shape-regularity constraints, suppress most of these artefacts, but manual review remains necessary in densely settled areas.
The architecture behind the segmentation
Instance segmentation models, particularly variants of U-Net and Mask R-CNN adapted for remote sensing, are the current workhorses for this task. They are trained on labelled datasets where field boundaries have been digitised by hand, often from WorldView or Pléiades imagery, and then applied to new areas. Transfer learning matters enormously: a model trained on West African compound fields performs poorly on the long strip fields of the Ethiopian highlands without fine-tuning.
Training data scarcity is the binding constraint, not compute. Labelling sub-hectare boundaries in fragmented landscapes is slow, skilled work. Published open datasets, including those released under the NASA Harvest programme and the AI4EO Food Security challenge, have helped, but coverage of South Asian landscapes remains thin. Accuracy figures from peer-reviewed studies using these methods typically report intersection-over-union scores of 0.6 to 0.8 on held-out test tiles, which translates to boundary position errors of a few metres at best and tens of metres at worst near dense canopy.
Minimum mappable unit in practice sits around 0.05 to 0.1 ha with WorldView-3 or Pléiades Neo, and 0.2 to 0.3 ha with Planet SuperDove. Below those thresholds, boundary uncertainty exceeds parcel size and the output becomes unreliable for individual farm attribution.
Cloud cover is not a scheduling problem, it is a data problem
Humid sub-Saharan landscapes and South Asian monsoon zones are cloud-covered for large fractions of the growing season. This is the single largest operational constraint on VHR boundary mapping. WorldView-3 and Pléiades Neo are optical-only systems; a cloud deck at acquisition time means no usable data, full stop. Planet's daily revisit improves the odds of finding a clear acquisition, but even at daily cadence, persistent cloud in the Congo Basin or the Ganges Plain can block a usable window for weeks.
The practical response is to plan acquisitions at the start and end of the growing season, when cloud probability is lower in many target regions, and to use Sentinel-1 SAR backscatter as a gap-filler for detecting field activity when optical imagery is unavailable. SAR cannot replace optical imagery for boundary delineation at sub-hectare scale, but it can confirm which areas were cultivated, narrowing the region where VHR tasking is most urgent.
From polygons to programmes
A delineated boundary layer is not an end product; it is infrastructure. Once individual parcels are mapped and attributed to farmers through ground-truth enrolment, the same geometry supports crop type classification, yield estimation, insurance loss verification and subsidy disbursement. The boundary file becomes the spatial key that joins satellite observations to farmer records.
Accuracy requirements differ by application. Index insurance can tolerate boundary errors of 10 to 20 m if the area estimate is unbiased at the portfolio level. Individual subsidy verification requires much tighter attribution, often better than 5 m, to prevent double-counting adjacent plots. Knowing the downstream use case before acquisition planning is therefore not optional: it determines whether SuperDove is sufficient or whether WorldView-3 tasking is justified.
Satellize's crop-estimation work in the Kingdom of Tonga demonstrated how parcel geometry, even in a small island context, anchors every downstream analytic. The same principle applies at continental scale in sub-Saharan Africa, where governments and development programmes are beginning to treat boundary layers as sovereign data assets rather than project deliverables. Discuss your target geography and minimum parcel size with our team before committing to a sensor or acquisition schedule.
Typical figures
| Spatial resolution (VHR optical) | 0.3 m pan / 0.7–1.24 m multispectral (Pléiades Neo, WorldView-3); 3 m multispectral (Planet SuperDove) |
| Revisit (tasked VHR) | 1–4.5 days depending on constellation and latitude; not guaranteed daily |
| Revisit (Planet SuperDove) | Daily at the equator under clear sky; cloud gaps can extend effective revisit to weeks in humid tropics |
| Minimum mappable unit | ~0.05–0.1 ha with WorldView-3 or Pléiades Neo; ~0.2–0.3 ha with Planet SuperDove |
| Spectral bands used | Visible (RGB), red-edge, NIR, SWIR (sensor-dependent); seasonal NDVI composites from Sentinel-2 used as auxiliary input |
| Boundary position accuracy | Typically 2–15 m RMS depending on canopy cover, image resolution and training data quality; degrades significantly under tree canopy |
| Archive depth | WorldView/Pléiades archive from ~2010; Planet SuperDove from ~2021; Sentinel-2 from 2015 (open access) |
| Delivery format | GeoPackage or Shapefile polygon layer with area attributes; GeoTIFF probability raster optionally included |
| Typical processing latency | 3–10 days from clear-sky acquisition to validated boundary layer, depending on area and QA requirements |
Analytics Satellize can run
| Parcel boundary polygon layer | Instance segmentation (U-Net or Mask R-CNN variants) applied to VHR multispectral imagery with spectral-gradient edge detection | GeoPackage polygon file with per-parcel area, centroid coordinates and confidence score |
| Seasonal composite for boundary enhancement | Median or percentile compositing of Sentinel-2 time series across growing season to maximise phenological contrast at field edges | Multi-band GeoTIFF composite used as auxiliary model input; optionally delivered as standalone product |
| Minimum mappable unit assessment | Resolution and signal-to-noise analysis against target parcel size distribution derived from field survey samples | Technical memo specifying recommended sensor, acquisition timing and expected accuracy for the target landscape |
| Change detection between boundary epochs | Polygon overlay and area-difference analysis between two dated boundary layers to detect field subdivision or consolidation | Change layer (GeoPackage) flagging parcels with area change exceeding threshold, with before/after imagery chips |
| Accuracy assessment report | Stratified random sampling of delineated parcels against independent reference digitisation or GPS field measurements; intersection-over-union and boundary displacement statistics | PDF accuracy report with confusion statistics, error maps and recommended use-case suitability rating |
| Farmer-parcel attribution table | Spatial join of delineated polygons to ground-truth enrolment coordinates collected via mobile survey; duplicate and overlap resolution | CSV or geodatabase table linking parcel IDs to farmer identifiers, ready for programme management system ingestion |
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.