If Geospatial MRV Cannot Work for Smallholder Agroforestry, It Has Not Solved the Real Problem Yet

The easiest landscape for geospatial MRV is a large, continuous forest block with clear boundaries, uniform vegetation, and limited human interference.

The real test is almost the opposite: thousands of small farms, scattered trees, mixed crops, narrow boundaries, different planting dates, seasonal harvesting, pruning, grazing, and management decisions that rarely follow a clean spatial pattern.

That is where a large share of the future opportunity for agroforestry and farmer-led restoration may sit. It is also where earth observation begins to struggle in commercially important ways.

A 2026 stakeholder report from Project CREDIBLE and the European Association of Remote Sensing Companies identifies cloud gaps, mixed pixels, young forest stands, and heterogeneous smallholder and agroforestry landscapes as persistent constraints on scalable MRV. Carbon Direct’s 2025 landscape analysis reaches a complementary conclusion: remote sensing can improve forest-carbon monitoring, but adoption is still constrained by unclear standards, limited technical capacity, expensive or inaccessible data, and inconsistent treatment of uncertainty. (Project CREDIBLE/EARSC; Carbon Direct)

This does not mean earth observation is failing. It means the industry has often demonstrated its strongest results in the landscapes that are easiest to observe.

Why this is the real commercial test

Smallholder agriculture is not a marginal land-use category. A global analysis published in World Development estimated that farms smaller than two hectares account for approximately 84 percent of all farms worldwide, although they operate a much smaller share of total agricultural land. (Lowder et al., 2021)

Trees are also already widespread within agricultural landscapes. A global remote-sensing analysis found that, in 2010, approximately 43 percent of agricultural land had at least 10 percent tree cover. The estimate was based on relatively coarse MODIS data and should not be interpreted as a precise cadastral census of agroforestry, but it demonstrates the scale at which trees and agriculture already coexist. (Zomer et al., 2016)

The future restoration opportunity may be even larger. Shyamsundar and colleagues estimated that, by 2050, low-cost tree-cover restoration could be feasible across approximately 280 million hectares of tropical cropland, 200 million hectares of pastureland, and 60 million hectares of degraded forestland. Their analysis suggested that these opportunities could affect around 291 million people, with most of the potential concentrated in populated agricultural landscapes. (Shyamsundar et al., 2022)

In other words, the landscapes that matter for scaling restoration finance are often not neat polygons. They are lived-in farm mosaics.

If MRV works only where tree cover is continuous, parcel boundaries are obvious, and field conditions are relatively uniform, it risks excluding precisely the landscapes where carbon finance could support large numbers of farmers.

Where conventional EO pipelines break

The first problem is spatial resolution: A 10-metre pixel may contain a tree crown, crop vegetation, exposed soil, shadow, and part of an adjacent parcel. The satellite records the combined signal. It does not automatically know which farmer owns the tree, whether it was planted by the project, or whether a change occurred inside the eligible boundary.

Research on farmland trees in India illustrates the problem clearly. Brandt and colleagues found that Sentinel-2-scale imagery can detect groups of trees but generally cannot identify individual farmland trees. Even imagery at three-to-five-metre resolution was primarily able to detect larger adult trees, while smaller trees and crowns below the study’s detection threshold remained difficult to capture. (Brandt et al., 2024)

The second problem is temporal noise: Clouds, shadows, seasonal crop cycles, pruning, harvesting, drought stress, soil moisture, and changes in viewing conditions can all alter the signal recorded from the same farm. A short-term increase in greenness is not necessarily tree growth. A temporary decline is not necessarily tree mortality.

Recent pan-tropical tree-mapping research found that cloud cover, acquisition date, dust, sensor differences, viewing angle, seasonal crown condition, and model-transferability issues all affected tree detection. The researchers required large and geographically diverse training datasets, repeated imagery, automated sampling, and explicit uncertainty analysis to produce consistent results. (Liu et al., 2025)

The third problem is that satellites do not directly measure carbon: Optical sensors observe reflected energy. Radar measures backscatter. LiDAR measures aspects of vegetation structure. Carbon estimates are derived by relating these signals to tree height, crown dimensions, wood density, field measurements, and allometric models.

That relationship can vary across species, climatic zones, management systems, tree ages, and biomass ranges. Understory vegetation is especially difficult to observe from above, while below-ground biomass, dead organic matter, and soil carbon cannot be reliably quantified from canopy imagery alone. (Huang et al., 2022)

A 2023 review of 33 agroforestry biomass studies found that imagery below two metres was widely used because it helped delineate heterogeneous agroforestry features. It also found that accuracy depended on spatial and spectral resolution, covariate selection, delineation methods, and the type and size of the agroforestry system. Importantly, the review found no statistically significant evidence that machine-learning models automatically outperform other modelling approaches. (Thapa, Lovell and Wilson, 2023)

Machine learning can improve interpretation. It cannot manufacture information that the sensor never captured.

The problem is operational as much as technical

Even a technically strong biomass model does not solve parcel registration, farmer identification, tenure documentation, planting records, species information, monitoring dates, data permissions, or the allocation of credits and payments.

This is where many geospatial MRV systems quietly become expensive again.

Very-high-resolution imagery may improve tree detection, but repeated acquisitions over large areas can be costly. Drone and LiDAR surveys provide valuable structural information, but they require equipment, trained operators, permissions, processing capacity, and field access. Ground plots remain necessary for calibration and validation.

The 2026 Project CREDIBLE report notes that EO-based forest inventories may become technically and economically viable only after projects reach several thousand hectares. It further suggests that pooling or group-based project structures may be a precondition for making EO affordable across small holdings.

This is a critical commercial point. The objective is not simply to achieve the lowest possible model error. It is to achieve defensible uncertainty at a cost the project can carry without absorbing an unreasonable share of farmer revenue.

What a workable MRV stack looks like

The solution is not one better satellite image. It is a layered MRV architecture.

The first layer is parcel and farmer intelligence: verified boundaries, farmer identifiers, tenure or use-right documentation, activity dates, species information, planting density, management history, and links between each parcel and its supporting evidence.

The second layer is object-level mapping. Where individual or dispersed trees matter, the system needs to move beyond pixel averages and identify crowns, tree rows, shelterbelts, hedges, and other discrete features. Research in Rwanda has demonstrated that sub-metre imagery, deep learning, extensive hand-labelled training data, and field plots can support national-scale mapping of individual overstory trees. The same research also shows how demanding such an exercise is. (Mugabowindekwe et al., 2023; Mugabowindekwe et al., 2024)

The third layer is time-series monitoring. Frequent optical imagery can track vegetation and land-cover change, while radar can reduce dependence on cloud-free optical observations. Multisensor studies in West African agroforestry systems have already combined Sentinel-1, Sentinel-2, ALOS and GEDI data for biomass estimation. (Kanmegne Tamga et al., 2023)

The fourth layer is targeted field measurement. Field plots, tree censuses, drone surveys, or LiDAR should be deployed where they add the most information: during model calibration, within unusual strata, in young plantations, where imagery is inconclusive, or where the financial consequence of an error is high.

The final layer is auditability. Every boundary revision, model version, input dataset, exclusion decision, uncertainty estimate, and manual correction needs a traceable record. A dashboard is useful only if the evidence behind it can survive verification.

The operating model should therefore be exception-based. Remote sensing should monitor the full portfolio and identify where the signal is clear, where it is uncertain, and where field investigation is required. Field resources can then be directed towards the uncertain and high-risk cases rather than spread uniformly across every parcel.

Why methodology trends matter

Carbon methodologies are beginning to reflect this distinction.

Verra’s VM0047 v1.1 provides two different quantification routes. The area-based approach combines remote sensing with plot-based sampling and uses a dynamic performance benchmark. Changes in a vegetative stocking index are compared between project areas and matched control plots to assess additionality and determine the crediting baseline at each verification.

The census-based approach is designed for dispersed planting activities, including certain agroforestry, shelterbelt, urban forestry, and revegetation projects that do not result in land-use change and where a complete census is feasible. Under v1.1, Verra describes this route as applying to projects planting 50 or fewer trees per hectare, while denser planting generally falls under the area-based approach. (Verra, VM0047 v1.1)

That distinction matters. It recognises that not every agroforestry project should be forced into the same remote-sensing architecture.

The operational requirements are also becoming more explicit. In May 2026, Verra announced its first three vetted data service providers for VM0047 stocking-index data: Sylvera, Kanop, and Chloris Geospatial. The associated requirements include annual data at 30-metre resolution or finer, consistent seasonal treatment, evidence that the stocking index correlates with above-ground biomass in the relevant forest type or ecoregion, and auditable documentation of project and control-plot inputs. (Verra, 2026)

This is a meaningful step towards standardisation. It is not proof that every fragmented smallholder landscape is now easy to monitor.

The breakthrough that still matters

The next breakthrough in geospatial MRV is unlikely to be a marginal improvement in prediction accuracy inside an already well-mapped forest.

It will be a system that can identify and monitor dispersed trees, connect spatial observations to the correct farmer and parcel, separate real growth from seasonal noise, quantify uncertainty honestly, direct field teams towards exceptions, and remain affordable across thousands of small holdings.

That is the point at which geospatial MRV becomes more than a remote-sensing product. It becomes infrastructure for inclusive climate finance.

Until it can work reliably in fragmented smallholder agroforestry landscapes, it has not solved the real problem yet.

References

• Lowder, S. K., Sánchez, M. V. and Bertini, R. (2021). Which farms feed the world and has farmland become more concentrated? World Development, 142, 105455.

• Zomer, R. J. et al. (2016). Global Tree Cover and Biomass Carbon on Agricultural Land. Scientific Reports, 6, 29987.

• Shyamsundar, P. et al. (2022). Scaling smallholder tree cover restoration across the tropics. Global Environmental Change, 76, 102591.

• Thapa, B., Lovell, S. and Wilson, J. (2023). Remote sensing and machine learning applications for aboveground biomass estimation in agroforestry systems. Agroforestry Systems, 97, 1097–1111.

• Kanmegne Tamga, D. et al. (2023). Estimation of Aboveground Biomass in Agroforestry Systems over Three Climatic Regions in West Africa. Sensors, 23, 349.

• Mugabowindekwe, M. et al. (2023). Nation-wide mapping of tree-level aboveground carbon stocks in Rwanda. Nature Climate Change, 13, 91–97.

• Mugabowindekwe, M. et al. (2024). Trees on smallholder farms and forest restoration are critical for Rwanda to achieve net zero emissions. Communications Earth & Environment, 5.

• Brandt, M. et al. (2024). Severe decline in large farmland trees in India over the past decade. Nature Sustainability.

• Liu, S. et al. (2025). Mapping previously undetected trees reveals overlooked changes in pan-tropical tree cover. Nature Communications.

• Carbon Direct (2025). Remote Sensing for Forest Carbon: Challenges and Opportunities.

• Project CREDIBLE and EARSC (2026). Earth Observation across the Carbon Value Chain.

• Verra (2025). VM0047 Afforestation, Reforestation, and Revegetation, v1.1.

• Verra (2026). Vetted Data Service Providers for the VM0047 ARR Methodology.