Separating Forest from Farmland: Mapping Mixed Land Cover with EO Data
Why Mixed Land Cover Is Hard to Map
Earth observation (EO) has made it possible to classify land cover across entire regions, but the hardest areas are rarely the pure ones. A dense forest interior is easy to label, and so is a large uniform field. The real difficulty begins where the two meet: along forest edges, hedgerows, riparian strips, and the mosaic of small plots that characterises many agricultural landscapes.

In these transition zones, a single satellite pixel can contain tree crowns, crop rows, grass, and bare soil at the same time. The result is a spectral mixture that resembles neither forest nor farmland. This ambiguity is not a marginal problem; in many regions, mixed land cover occupies a substantial share of the land surface and shapes how carbon, water, and biodiversity are accounted for.
The Cost of Confusing Forest with Farmland
When forest edges are classified as agricultural land, forest area is underestimated and carbon stock calculations become unreliable. When farmland with scattered trees is classified as forest, agricultural production can be overstated as forest loss. Both errors propagate into policy, reporting, and supply chain monitoring.
A boundary is not a line on the ground; it is a decision about where one system ends and another begins.
Conventional classifiers often force each pixel into a single class. That design choice works against mixed landscapes, where the most honest answer is frequently a proportion rather than a label.
How SMFM Approaches the Problem
The SMFM project treats the forest-farmland boundary as a modelling problem rather than a labelling shortcut. Instead of assigning one class per pixel, SMFM methods estimate fractional cover and use spatial context to separate genuine forest edges from agricultural plots that merely contain trees.
- Harmonise EO imagery from multiple sensors and seasons.
- Extract features that describe texture, phenology, and structure.
- Estimate sub-pixel fractions of tree, crop, and other cover.
- Validate the resulting maps against field and reference data.
This layered approach keeps the method transparent and lets users inspect where the uncertainty lies, rather than hiding it behind a single hard class.
From Pixels to Decisions
Mixed land cover is not a nuisance to be smoothed away. It is a signal about how people use the landscape. A farm with shade trees, a forest edge that is gradually cultivated, and a plantation that resembles a natural forest all require different interpretations even when their spectral signatures overlap.
SMFM therefore reports both the most likely class and the confidence attached to it. Planners can then decide whether a parcel should be counted as forest, farmland, or a managed transition zone, and they can revisit that decision as new EO data arrive.
What This Means for Users
For national reporting, the practical benefit is a map that does not force false certainty at the forest-farmland interface. For researchers, it offers a reproducible way to compare landscapes. For local stakeholders, it provides a clearer picture of where cultivation and tree cover actually overlap.
| Challenge | Common outcome | SMFM response |
|---|---|---|
| Mixed pixels at forest edges | Misclassification | Fractional cover estimation |
| Scattered trees on farms | Overstated forest | Context-aware features |
| Seasonal crop changes | Unstable labels | Multi-season EO inputs |
Separating forest from farmland is ultimately about respecting complexity. EO data can capture that complexity, and SMFM methods are designed to make it usable.
