Machine learning and landscape metrics track forest fragmentation in Naqamte City, Ethiopia

In the highlands of western Ethiopia, a quiet ecological collapse is unfolding, and for the first time scientists have captured its full anatomy in unprecedented detail. A new study of Naqamte City, a rapidly growing secondary city in the East Wollega Zone of the Oromia Regional State, has documented a landscape transformation so complete that the interior habitat of its forests — the dark, cool, undisturbed core that many forest species require to survive — has vanished entirely. According to research published in the journal Discover Forests, the proportion of land covered by buildings and roads in Naqamte surged from just 9.2 percent in 1991 to a staggering 51.1 percent by 2025, while forest cover crashed from 25.8 percent to a mere 6.3 percent. But the most alarming finding is not how much forest remains; it is what kind of forest remains. By 2025, core forest habitat — the ecologically precious interior of forest tracts — had been reduced to zero, meaning that every last hectare of woodland in the study area now qualifies as fragmented edge habitat, exposed to the drying winds, invasive species, human disturbance, and microclimatic stress that define forest margins.

The research, led by Milkessa Dangia Nagasa and Birhanu Tadese Edosa, stands out for its methodological rigor, combining machine learning classification of satellite imagery with standardized landscape ecology metrics and on-the-ground socio-economic investigation. The team analyzed Landsat satellite images from three pivotal years — 1991, 2011, and 2025 — drawn from three successive generations of sensors: Landsat 5 Thematic Mapper, Landsat 7 Enhanced Thematic Mapper Plus, and Landsat 9’s Operational Land Imager and Thermal Infrared Sensor. Each scene, captured at a spatial resolution of 30 meters along Path/Row 170/54, was downloaded from the USGS EarthExplorer archive and subjected to careful preprocessing, including radiometric correction, geometric registration to the Universal Transverse Mercator Zone 37N projection with a root mean square error below half a pixel, and cloud masking using Quality Assessment bands — a step that proved especially important for the 2011 Landsat 7 imagery, which is prone to scan-line artifacts.

At the heart of the classification pipeline was a Support Vector Machine, or SVM, a supervised machine learning algorithm that has increasingly displaced older statistical classifiers such as maximum likelihood and ISODATA in land cover mapping. The SVM’s advantage lies in its mathematical construction: rather than modeling the statistical distribution of each land cover class, it finds the optimal hyperplane — or, in non-linear cases, the optimal decision boundary in a high-dimensional kernel-transformed feature space — that maximally separates classes with complex, overlapping spectral signatures. This capability matters enormously in tropical highland landscapes like Naqamte’s, where mixed pixels containing both crops and grasses, or shadows from scattered trees, routinely confuse conventional classifiers. The researchers collected roughly 100 training polygons per epoch from historical high-resolution Google Earth Pro imagery and field observations, and optimized the SVM’s two critical hyperparameters — the penalty parameter C, which controls the tolerance for misclassification, and the kernel coefficient gamma, which defines the influence radius of individual training points — through a grid-search procedure.

The results validated the choice emphatically. Classification accuracy, assessed with 100 independently collected validation points per map year using stratified random sampling, climbed from an overall accuracy of 81.3 percent with a Kappa coefficient of 0.82 in 1991, to 88.0 percent and 0.86 in 2011, to an almost perfect 94.7 percent and 0.93 in 2025. Waterbodies, when detectable, achieved perfect user’s and producer’s accuracies thanks to their spectrally distinctive signature, while grasslands proved the most troublesome class throughout, reflecting persistent spectral confusion with cropland in this smallholder farming landscape where coffee, cereals, and khat are cultivated under rainfed conditions. Because the same classifier and validation protocol were applied consistently across all three epochs, the researchers argue the maps remain directly comparable, allowing robust long-term trend analysis despite the modest variation in accuracy.

What those maps revealed is a story of two very different decades. Between 1991 and 2011, the landscape changed moderately, with built-up land gaining 397 hectares at a rate of about 20 hectares per year and land exchanging mainly among cropland, grassland, and forest in complex bidirectional flows — cropland converting to grassland across 16.13 percent of the studied transitions, and grassland reverting to cropland across 13.68 percent. Then came the acceleration. Between 2011 and 2025, built-up land gained 3,543 hectares, expanding roughly ten times faster than in the previous interval, while forests lost 944 hectares at a rate of 67.4 hectares per year, croplands shed 1,309 hectares, and grasslands declined by 633 hectares. The single largest transitions in this period were forest converted to built-up land, accounting for 23.78 percent of all changes, and cropland converted to built-up land at 23.70 percent — a clear signature of urbanization consuming both natural and agricultural land simultaneously.

To translate these land cover statistics into statements about ecological integrity, the team turned to FRAGSTATS version 4.2, the standard software of landscape ecology, after reclassifying their maps into a binary forest-versus-non-forest layer. The fragmentation framework they employed distinguishes forest pixels into four structural categories: patch forest, which consists of small, isolated fragments too small to contain any interior; edge forest, the roughly 100-meter band along forest boundaries where edge effects dominate; perforated forest, which lines internal clearings within larger tracts; and core forest, the protected interior lying beyond the reach of edge influence. The trajectory of these categories over 34 years reads like an ecological death certificate. Core forest fell from 165 hectares in 1991 to 61 hectares in 2011, and by 2025 it had disappeared completely. Patch forest — scattered, disconnected fragments — became the dominant forest structure, meaning the remaining woodland exists only as ecological islands.

The implications of losing core habitat extend far beyond aesthetics. Interior forest environments maintain higher humidity, more stable temperatures, and reduced light penetration, conditions on which shade-adapted plants, interior-dwelling birds, small mammals, fungi, and countless invertebrates depend. When a forest is subdivided until only edge remains, these microclimatic refuges disappear even if the raw acreage of trees stays constant on paper. This is precisely the nuance the study exposes: although the net change accounting showed forests with an apparent cumulative gain over the full study period, that statistical stability concealed a catastrophic structural degradation, driven by relentless reciprocal conversions between forest, grassland, and cropland that steadily chopped continuous woodland into ever smaller, more isolated pieces. Landscape metrics confirmed the pattern, with rising patch density, shrinking mean patch size, and increasing edge density all indicating intensifying fragmentation.

Why did this happen? The researchers triangulated their satellite analysis with 24 key informant interviews conducted between March and May 2025 — drawing on farmers, community elders, agricultural extension officers, forestry experts, and municipal officials — along with four focus group discussions of six to eight participants each, all analyzed through thematic content analysis and cross-checked against spatial overlay analysis of roads, settlements, and agricultural lands. The picture that emerged is a classic coupling of proximate and underlying drivers. Directly responsible were urban expansion, agricultural encroachment, fuelwood and charcoal production, timber extraction, and overgrazing. But beneath these immediate pressures lie deeper forces: rapid population growth, entrenched poverty, insecure land tenure that discourages long-term stewardship, weak governance of land use, and limited environmental awareness. Ethiopia’s dependence on biomass energy and its unregulated urbanization, the authors note, have accelerated land cover change nationwide over the past three decades, and Naqamte’s moist Afromontane forests — vital for carbon sequestration, hydrological regulation, and local cultural value — have paid a disproportionate price.

The study’s broader significance lies in its demonstration that forest loss and forest fragmentation are not the same phenomenon, and that measuring only the former can seriously mislead conservation policy. A forest can persist in statistics while dying in structure, its ecological functions dismantled piece by piece. For Naqamte, the authors argue, the landscape has now crossed a threshold into a highly urbanized, fragmented state in which fragmentation outweighs outright forest loss as the central conservation challenge. Their prescriptions are correspondingly ambitious: integrated land-use planning that treats remaining forests as critical infrastructure, stronger governance with enforced land-use regulation, alternative energy programs to relieve fuelwood pressure, active forest restoration and corridor creation to reconnect isolated patches, and community-based forest management programs that give local residents a stake in conservation. Whether such interventions arrive in time may determine whether western Ethiopia’s highland forests survive the twenty-first century — or survive only as scattered trees in a sea of concrete.

Nagasa, M. D., & Edosa, B. T. (2026). Forest fragmentation and land use land cover change analysis in Naqamte City Western Ethiopia using machine learning and landscape metrics. Discover Forests, 2(38). https://doi.org/10.1007/s44415-026-00093-w