Machine Learning Maps Fragile Zones in Vietnam’s Hinh River Basin

Deep in the mountains of south-central Vietnam, the Hinh River winds through one of the most ecologically important and environmentally sensitive landscapes in Phu Yen Province. A new study published in the journal One Ecosystem has combined satellite imagery, machine learning and expert-driven spatial analysis to track how this landscape has changed over more than a decade and to predict where it is heading by 2030. The findings offer a rare, data-rich window into the environmental fragility of a basin that supplies vital water, hydropower and climate-regulation services to downstream communities across the entire south-central coastal region.

The research team, working on the Google Earth Engine cloud platform, classified land cover across the basin for four separate years: 2010, 2015, 2018 and 2023. They drew on surface reflectance images from the Landsat 5 and Landsat 8 satellites, each with a spatial resolution of 30 metres. Before classification, the images were pre-processed with a cloud mask that used the pixel quality band to strip out clouds and cloud shadows, producing clean, cloud-free composites. The scenes were then clipped to the boundary of Song Hinh District so that the analysis remained strictly focused on the study area.

To train the classification, the researchers sampled 978 ground-truth points, drawn from a 2010 land-use map supplied by the Phu Yen People’s Committee and from field surveys in which every point was photographed, described and geotagged. They applied the Random Forest algorithm, using 70 percent of the points to train the model and the remaining 30 percent to test it. The results were strong: a mean area under the curve value of 0.85 and a Kappa coefficient of 0.8, indicating highly reliable classification. Seven land-cover types were mapped: annual industrial crops, urban and built-up land, water, crop-shrub-grass mixtures, rice, scattered trees, and forest covering both natural and plantation stands.

The classified maps revealed a striking pattern. Between 2010 and 2023, only two land-cover types expanded: forest, including plantation forests, and annual industrial crops. Every other category shrank. Urban and built-up areas declined, rice paddies and crop-shrub-grass land contracted, and scattered trees thinned out. The researchers interpret the growing forest area as evidence that provincial forestry policies are working, pointing to the implementation plan for the forestry development strategy for 2021 to 2030 and the provincial project to plant 15 million trees between 2021 and 2025.

Looking forward, the team used a Markov chain model to project land-cover change to 2030. The forecast suggests forest area will climb from roughly 25,000 hectares in 2010 to nearly 40,000 hectares by 2030, an increase of around 20 percent for forest and annual industrial crops combined. Annual industrial crops are expected to rise modestly from about 15,000 hectares to just under 20,000 hectares. Urban and built-up land is projected to slip from around 8,000 hectares to nearly 7,000, while rice cultivation should hold steady near 10,000 hectares and scattered trees remain close to 5,000. Crop-shrub-grass land, meanwhile, is expected to fall from about 26,000 hectares in 2010 to below 20,000 by 2030.

But mapping where land cover is going is only half the story. The study’s second major contribution is a zoning system that grades the basin by environmental fragility. The researchers first built a map of potential environmental fragility using five natural factors: terrain slope, annual mean precipitation, river hierarchy derived from a digital elevation model using the Strahler method, the percentage of sand in the top five centimetres of soil from the SoilGrid database, and geological domains supplied by the Vietnam Institute of Geosciences and Mineral Resources. A sixth factor, the land-cover type mapped in this study, was then added to produce a map of emergent environmental fragility, which captures the combined influence of natural susceptibility and human activity.

To weigh these factors, the team convened a focus group of six experts who scored each criterion on a nine-point scale, from extremely high to low resilience. The weights were then refined through the Analytic Hierarchy Process, a multi-criteria decision-making technique that compares factors pairwise on a scale from one, meaning equal importance, to nine, meaning extreme importance. The consistency of the expert judgements was checked using the consistency index and consistency ratio, both of which came in below the accepted threshold of 0.10, confirming that the weighting scheme was internally coherent. Weighted overlays of the scored factor maps produced the final fragility zones, reclassified into five classes using the Jenks natural-breaks method.

The results show that land cover itself is the single most influential driver of fragility, carrying a weight of 40 percent, followed by slope at 27.5 percent, precipitation at 14 percent, fluvial hierarchy at 11 percent, soil sand content at 4.5 percent and geology at just 3 percent. On the potential fragility map, roughly 16.3 percent of the basin falls into the extremely high fragility class, concentrated in the northwest and around large reservoirs, while a further 17.7 percent is highly fragile. Low-fragility zones, covering about 19 percent of the territory, sit mainly in the northeast and southwest, where extensive natural and plantation forests dominate.

The emergent fragility map sharpens the picture further. Areas requiring extremely high protection cover 26 percent of the basin and are dominated by low-lying terrain, dense river networks, lakes and immature scattered trees. High-protection zones, accounting for 29 percent, lie mainly in the mountainous south and east, where natural forests and mature plantations stand. Among the land-cover types, water emerged as the most sensitive category, followed by annual industrial crops, forest, scattered trees, rice, crop-shrub-grass and urban land. The researchers note that aquatic ecosystems are particularly vulnerable because they depend on multiple environmental factors to maintain ecological balance, and fragmentation or pressure from agriculture and industry can quickly degrade water quality, alter flow regimes and threaten the species that depend on these habitats.

The study’s conclusions carry practical weight for local planners. The authors recommend prioritising the protection of river, lake, irrigation and canal ecosystems, optimising land-resource use and improving the economic efficiency of agriculture and forestry. They also highlight the potential of scattered-tree areas for integrated agroforestry projects that could enhance land-cover sustainability. By making the Google Earth Engine workflow openly available, the team argues that researchers, managers and ecologists can share, verify and extend the approach, enabling continuous monitoring of land-cover change. For a basin that remains one of the region’s most vulnerable localities, the combination of machine-learning classification, Markov forecasting and expert-weighted fragility zoning offers a replicable template for safeguarding the natural ecosystems on which the south-central coast depends.

Subject of Research: Land-cover dynamics and environmental fragility zoning in the Hinh River Basin, Vietnam

Article Title: Landscape Dynamics and Environmental Fragility Zoning in Hinh River Basin: Insights for protecting natural ecosystems

Article References: Nguyen, Q. K., Tong, H., Nguyen, L., Nguyen, T. N., Ngo, T. D., Hong Quang, N., Dinh, A. T., & Pham, M.-P. (2024). Landscape Dynamics and Environmental Fragility Zoning in Hinh River Basin: Insights for protecting natural ecosystems. One Ecosystem, 9, Article e134088. https://doi.org/10.3897/oneeco.9.e134088

Image Credits: AI Generated

DOI: 10.3897/oneeco.9.e134088

Keywords: Hinh River Basin, environmental fragility, land cover change, Google Earth Engine, Random Forest, Markov chain, Analytic Hierarchy Process, Vietnam, forest restoration, remote sensing, Phu Yen Province, ecosystem protection