Satellites and Algorithms Race to Map the World’s Hidden Dryland Forests

Dryland forests are easy to overlook. Scattered across some of the driest inhabited regions of the planet, their trees grow in open canopies, separated by bare soil, grasses, and shrubs, so that from the vantage point of an orbiting satellite a single pixel of imagery may blend trees, shrubs, and ground together into one ambiguous signal. Yet these sparse woodlands matter enormously. They store carbon, regulate water cycles, shelter biodiversity, and sustain rural livelihoods across vast tracts of Africa, Asia, Australia, and the Americas. They are also among the most threatened forest types on Earth, cleared for agriculture and fuelwood and degraded by grazing and fire far faster than most monitoring systems can register. A new study published in Environmental Monitoring and Assessment offers a careful, technically grounded answer to a deceptively simple question: when you try to find these forests from space, should you classify the image pixel by pixel, or should you first group pixels into meaningful objects?

The research, led by Bárbara Guida-Johnson of the Argentine Institute of Snow, Glaciology and Environmental Sciences (IANIGLA) together with Leandro M. Alvarez, Juan A. Alvarez, Anabela Laudecina, and Pablo E. Villagra, focused on the Monte region of Mendoza Province in western Argentina. The Monte is a desert biome dominated by open woodlands of algarrobo trees, chiefly Neltuma flexuosa, formerly known as Prosopis flexuosa. These woodlands are structurally diverse: in some places the tree canopy forms relatively continuous patches, while in others, particularly where fire or roller-chopping has disturbed the vegetation, the trees survive only as scattered individuals embedded in a matrix of shrubs and bare ground. That structural variability is precisely what makes the Monte an ideal natural laboratory for testing remote sensing methods meant to generalize across contrasting forest architectures.

The team analyzed imagery from Sentinel-2, the European Space Agency’s pair of satellites that sweep the entire land surface every five days at a resolution of ten to twenty meters per pixel. Sentinel-2 has become the workhorse of medium-resolution land monitoring worldwide, not least because its data are free and flow into cloud-computing platforms such as Google Earth Engine, where continental-scale analyses have become routine. But the very features that make dryland forests interesting ecologically make them difficult targets for Sentinel-2. A ten-meter pixel over an open algarrobal typically contains a mixture of tree crown, shrub canopy, herbaceous cover, and soil, and the resulting spectral signature is a composite that may resemble none of those components individually. The question of how to handle that mixture sits at the heart of the pixel-versus-object debate in remote sensing.

The pixel-based approach is the classical method: each image pixel is assigned to a class, in this case forest or non-forest, based on its own spectral values. Guida-Johnson and colleagues fed the classifier pixel-level spectral information along with the Soil-Adjusted Total Vegetation Index, or SATVI, an index designed for arid environments that helps separate green vegetation from soil background, a persistent challenge where bright ground dominates the reflectance signal. The object-based approach, by contrast, first segments the image into contiguous groups of similar pixels, producing discrete objects that ideally correspond to real landscape units such as woodland patches, shrublands, or bare fields. Each object is then described not only by its average spectral values and SATVI but also by texture descriptors, which capture the spatial roughness and internal variability of the canopy, a property that open woodlands express distinctly.

Both approaches relied on Random Forest, a machine learning algorithm that builds hundreds of decision trees, each trained on random subsets of the reference data and random subsets of the predictor variables, and then lets them vote on the final class. Random Forest has become a standard tool in land-cover classification because it handles nonlinear relationships between spectral features and land cover, resists overfitting, and requires relatively little tuning. The authors applied it under a deliberately simple binary scheme, forest versus non-forest, at four study sites in the Monte chosen to represent contrasting forest structures, from sites with greater canopy continuity to sites where disturbance had fragmented the woody vegetation into a fine-grained mosaic.

The results were nuanced and, in their way, quietly revolutionary for monitoring practice. The object-based method achieved higher overall accuracy and lower allocation disagreement, a metric that measures how often the classifier places a unit in the wrong class regardless of how much area it assigns to each class, at sites where tree cover forms spatially coherent patches. Where the canopy is continuous, or where the woodland contrasts strongly with the surrounding land covers, grouping pixels into objects before classification pays off: the objects capture the shape, texture, and context of forest patches that individual pixels cannot. Texture, in particular, proved valuable, because the mottled pattern of crowns and gaps in an open woodland is a signature that survives even when the spectral signal is diluted by soil background.

But the story reversed at the most heterogeneous sites. In disturbance-driven systems, where fire histories and land use had shattered the vegetation into a patchwork of scattered trees, resprouting shrubs, and bare ground, the pixel-based method performed better. The explanation lies in the geometry of segmentation: when forest patches are smaller than or comparable to the size of image objects, segmentation merges forest with surrounding non-forest elements, blurring the very distinction the classifier is asked to make. Individual pixels, for all their mixed signatures, at least preserve the local variability of such landscapes. The pixel-based approach also carries a practical advantage: it demands far less computational effort, an important consideration when the goal is mapping forests across entire dryland regions rather than across a handful of study sites.

The broader significance of these findings extends well beyond Mendoza. A landmark 2017 study in Science estimated that drylands contain substantially more forest than conventional maps acknowledged, prompting both celebration and controversy, with critics arguing that differing definitions of forest and the difficulty of detecting sparse tree cover introduced large uncertainties. Subsequent work has warned that underestimating tree cover in drylands can bias global restoration planning, sometimes steering ambitious tree-planting initiatives toward ecosystems that are naturally open grasslands or shrublands. Accurate, structure-aware detection of dryland forests is therefore not a technical nicety; it shapes conservation policy, national forest inventories, and the allocation of restoration funding on a planetary scale. In Argentina, the stakes are concrete: national and provincial native forest mapping under the country’s forest law guides land-use planning and the regulation of clearing in woodlands such as the algarrobales of the Monte.

What the new study delivers is practical guidance rather than a single winner. Classification performance, the authors conclude, is strongly influenced by forest structure and spatial organization, and the appropriate strategy depends on the characteristics of the landscape being mapped. For regions where woodlands form coherent patches with clear boundaries, object-based analysis with spectral and texture descriptors offers the sharper picture. For regions where disturbance has produced fine-grained, heterogeneous mosaics, pixel-based classification with a soil-adjusted vegetation index may be both more accurate and more efficient. Monitoring programs, the study suggests, should diagnose the structural character of their target landscapes before committing to a method, rather than assuming that one approach fits all drylands.

There is also a forward-looking dimension. The entire workflow was designed to run in cloud-computing environments, meaning the same comparisons can be scaled from four sites in Mendoza to entire biomes, and the binary forest/non-forest framework can be extended with additional indices, radar data, or deep learning architectures as they mature. As climate change intensifies aridity and land-use pressure across the world’s drylands, the forests that survive there will become ever more critical refuges, and ever more important to count correctly. This study reminds the remote sensing community that the forest, in drylands especially, is not always easy to see, and that seeing it well requires matching the algorithm to the architecture of the trees themselves.

Subject of Research: Comparison of pixel-based and object-based Sentinel-2 classification methods for detecting open dryland forests in the Monte region of Argentina

Article Title: Detecting dryland forests across contrasting vegetation structures: object- and pixel-based classification methods for forest monitoring

Article References: Guida-Johnson, B., Alvarez, L. M., Alvarez, J. A., Laudecina, A., & Villagra, P. E. (2026). Detecting dryland forests across contrasting vegetation structures: object- and pixel-based classification methods for forest monitoring. Environmental Monitoring and Assessment, 198(11), Article 1183. https://doi.org/10.1007/s10661-026-16025-2

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16025-2

Keywords: dryland forests, remote sensing, Sentinel-2, Random Forest, object-based image analysis, pixel-based classification, Monte desert, Argentina, SATVI, forest monitoring, Google Earth Engine, open woodlands