Sensors
- Sentinel-2 MSI: 10 m resolution in red and NIR bands (used to compute NDVI); 5-day revisit at the equator with both satellites active. Dense enough to resolve individual smallholder parcels and to capture the green-up pulse within a growing season window of two to three weeks.
- MODIS MOD13Q1: 250 m, 16-day composite NDVI product with an archive stretching back to 2000. Coarse for parcel-level work but indispensable for building the multi-year phenological baselines against which Sentinel-2 anomalies are judged.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit per satellite (8-day when both are operational). Bridges the resolution gap between MODIS and Sentinel-2 and extends usable time series back to 1984, which matters when distinguishing long-term abandonment from recent fallow.
- Sentinel-1 SAR C-band: 12-day repeat, all-weather. Backscatter and coherence change through a growing season differ markedly between tilled bare soil and untouched fallow; SAR is used to confirm optical classifications when cloud cover breaks the Sentinel-2 time series.
What a missing green-up pulse actually means
Every cultivated field has a phenological fingerprint: a predictable rise in NDVI as the canopy closes, a plateau during grain fill, and a decline at senescence or harvest. The timing varies by crop and climate zone, but within any given region the pattern repeats year after year with modest variation. A parcel that lies inside a validated cropland mask yet shows no NDVI rise above background soil reflectance during the expected growing window is, by definition, not being cultivated that season.
The complication is that 'not cultivated this season' covers a wide range of situations. A farmer may have planted late due to delayed rainfall. The field may be in a deliberate one-season fallow as part of a rotation. Or the land may have been abandoned years ago and is now reverting to scrub. Each scenario has different policy implications, and satellite data alone cannot always resolve the cause. What it can do is document the fact with precision and flag it for ground follow-up.
Baselines make the classification, not single images
A single low-NDVI image proves nothing. Cloud shadow, smoke, or a late planting date can all depress a pixel's value on any given acquisition date. The analytical approach that holds up is trajectory-based: stack all cloud-screened observations across a growing season, fit a smoothed phenology curve, and compare that curve against the distribution of curves observed over the preceding five to ten years for the same parcel or land-cover class.
MODIS MOD13Q1 provides the long archive needed to build those baselines. Its 250 m resolution means that for fields smaller than roughly one hectare, MODIS pixels are mixed and the baseline reflects an average of several parcels. Sentinel-2 then supplies the spatial precision: its 10 m NDVI time series, once harmonised to the MODIS baseline using methods like those described in the NASA Harmonized Landsat Sentinel-2 (HLS) product, can flag individual parcels as anomalous relative to the regional multi-year norm.
The threshold for 'fallow' classification is typically set where the peak growing-season NDVI falls more than one standard deviation below the multi-year parcel mean during the expected green-up window. Persistent absence across three or more consecutive years is the operational definition of abandonment used in several published European land-use studies. Both thresholds are adjustable and should be calibrated against local ground truth before operational deployment.
Cloud cover is the honest limiting factor
Sentinel-2 revisits every five days, but usable observations depend on cloud-free acquisitions. In humid tropical or monsoonal zones, the entire growing season can fall within a persistent cloud deck. A 10 m optical time series that has four clear observations across a 90-day season is marginal for reliable phenology fitting; eight or more is the practical minimum for confident curve-fitting.
Sentinel-1 SAR partly compensates. C-band backscatter from bare soil differs from that of a vegetated canopy, and coherence between repeat passes degrades rapidly over growing crops but stays high over undisturbed bare or fallow ground. SAR-based indicators are noisier than NDVI for phenology work, but they provide an independent, cloud-independent signal that can confirm or challenge the optical classification. In persistently cloudy regions, a combined optical-SAR approach is not optional; it is the only way to achieve acceptable accuracy.
Separating intentional fallow from abandonment
The one-year versus multi-year distinction matters enormously for food-balance accounting. A field in planned fallow will return to cultivation the following season; an abandoned field will not. The satellite record resolves this straightforwardly over time, but a government ministry that needs an estimate now, mid-season, cannot wait for next year's data.
Several contextual signals help. Tilled bare soil has a different spectral signature from untilled bare soil: tillage disrupts surface crust structure and changes soil moisture retention, producing a detectable shift in shortwave infrared reflectance (Sentinel-2 bands 11 and 12, centred at 1610 nm and 2190 nm respectively). Fields showing recent tillage but no subsequent green-up are almost certainly in intentional fallow or failed establishment, not abandonment. Fields showing no tillage signal, no green-up, and progressive encroachment of shrub or grass spectral signatures over multiple years are strong candidates for permanent abandonment.
Parcel-level land registry data, where available, adds a further layer. A parcel with no registered ownership change and no cultivation signal for five years is a different policy problem from one recently transferred to a new owner who has not yet planted.
What the numbers can and cannot support
At 10 m resolution, Sentinel-2 can reliably delineate parcels larger than roughly 0.1 hectares in area, assuming the field boundary layer is accurate. Below that size, mixed pixels introduce classification noise. In fragmented smallholder landscapes, where fields of 0.05 to 0.2 hectares are common, the effective detection floor is higher than the sensor's nominal resolution implies.
Accuracy figures from published studies on fallow detection using Sentinel-2 time series typically range from 80 to 92 percent overall accuracy, depending on the complexity of the cropping system, the quality of the cropland mask used as the starting layer, and the density of cloud-free observations. Those figures are honest benchmarks, not guarantees. A poorly constructed cropland mask will propagate errors regardless of how well the phenology analysis is executed.
Satellize's crop-estimation programme in the Kingdom of Tonga demonstrated the value of combining dense Sentinel-2 time series with local agronomic calendars to distinguish active cultivation from bare ground in a fragmented island landscape. The same methodological framework applies directly to fallow and abandonment detection in any zone where a validated cropland mask exists.
From classification to policy use
The primary deliverables are a seasonal fallow map and a multi-year abandonment layer, both expressed as GIS polygons or raster grids aligned to the national parcel fabric. Change detection between consecutive seasons quantifies how much cultivated area has shifted in or out of production, which feeds directly into food-balance models and early warning systems.
For rural land-use policy, the abandonment layer is often more valuable than the seasonal fallow map. It identifies where agricultural land is leaving production permanently, whether due to rural depopulation, soil degradation, water scarcity, or economic marginalisation. Governments can cross-reference the abandonment layer against census data, irrigation infrastructure maps, and market access indicators to understand the drivers and target interventions. The satellite record does not explain why a field was abandoned. It documents that it was, with a date range and a spatial extent that no field survey at national scale could match for cost.
Typical figures
| Spatial resolution (optical) | 10 m (Sentinel-2 red/NIR); 30 m (Landsat 8/9); 250 m (MODIS MOD13Q1) |
| Revisit frequency | 5 days (Sentinel-2, both satellites); 8–16 days (Landsat 8/9); 16-day composite (MODIS MOD13Q1) |
| Minimum detectable parcel | ~0.1 ha at 10 m resolution in open landscapes; higher in fragmented or cloudy conditions |
| Key spectral bands | Red (~665 nm), NIR (~842 nm) for NDVI; SWIR (~1610 nm, ~2190 nm) for bare-soil and tillage discrimination |
| Archive depth | Sentinel-2 from 2015; Landsat from 1984; MODIS from 2000 |
| Baseline period recommended | 5–10 years of same-season observations for reliable anomaly detection |
| Cloud limitation | Minimum ~8 cloud-free observations per growing season for reliable phenology curve fitting; SAR fills gaps |
| Typical classification accuracy | 80–92% overall accuracy in published Sentinel-2 fallow detection studies, dependent on cropland mask quality |
| Latency after season close | 2–4 weeks for a seasonal fallow map once the growing season window is complete |
| Delivery formats | GeoTIFF raster, GeoPackage or Shapefile polygon layer, CSV parcel-level summary table |
Analytics Satellize can run
| Seasonal fallow map | NDVI time-series phenology fitting (Savitzky-Golay or TIMESAT-class smoothing) compared against multi-year parcel baselines from MODIS MOD13Q1 and Sentinel-2 | GeoTIFF and polygon GIS layer showing cultivated vs. fallow parcels for the current season, updated at season close |
| Multi-year abandonment layer | Persistent absence of green-up signal across three or more consecutive seasons, cross-checked against Landsat archive back to 2015 or earlier | Polygon GIS layer with estimated first year of non-cultivation per parcel, delivered as annual update |
| Fallow area statistics by administrative unit | Zonal aggregation of parcel-level classification against administrative boundary fabric | CSV or Excel report of fallow and abandoned area (hectares and percentage of cropland mask) per district or region |
| Season-on-season change detection | Differencing of consecutive seasonal fallow maps to identify fields entering or leaving cultivation | Change polygon layer and summary table for food-balance and early warning inputs |
| SAR-confirmed fallow classification | Sentinel-1 C-band coherence and backscatter change analysis used to validate or override optical fallow flags in cloud-affected zones | Revised fallow GIS layer with confidence score per parcel, noting optical vs. SAR-confirmed status |
| Tillage signal layer | SWIR reflectance (Sentinel-2 bands 11 and 12) change detection to distinguish tilled bare soil from untilled fallow or abandoned ground | Binary raster or polygon layer indicating recent tillage activity, used to separate intentional fallow from abandonment |
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.