1Department of Environmental Management, Institute of Environmental Engineering, RUDN University, 6 Miklukho-Maklaya St., 117198Moscow, Russia
2National Higher Education and Research Institute, Hamelmalo Agricultural College, Keren, Eritrea
3V.V. Dokuchaev Soil Science Institute, Pyzhevsky per. 7, Building 2, 119017Moscow, Russia
4National Authority for Remote Sensing and Space Sciences, Cairo1564, Egypt
*Corresponding Author: Tumuzghi Tesfay, Department of Environmental Management, Institute of Environmental Engineering, RUDN University, 6 Miklukho-Maklaya St., 117198Moscow, Russia. Email: tumuzghitesfay7@mail.com
Soil organic carbon (SOC) is essential for soil health, food security and climate change mitigation. However, reliable data on SOC distribution and its environmental drivers remain limited in data-scarce dryland regions like Eritrea. This hinders effective soil management and restoration planning.
SOC modelling was conducted across an altitudinal gradient landscape using environmental covariates and machine learning. Three predictor sets (46, 28 and 11 variables) were used selected through three approaches: 1) no selection, 2) removal of highly collinear (r≥0.90) and non-significant variables and 3) boruta algorithm-based selection. The predictive performance of cubist, random forest (RF) and partial least squares (PLS) algorithms were evaluated.
SOC levels across the study area were generally low (mean = 0.71%). Rainfed croplands and communal grazing areas showed particularly depleted SOC, attributed to unsustainable land management, while forest and irrigated systems retained significantly higher SOC, indicating greater carbon sequestration potential. The Cubist model with 46 predictors performed best (R2 = 0.7465, RPD = 2.0895), whereas PLS with 11 variables had the lowest accuracy (R2 = 0.5930, RPD = 1.6491). Temperature emerged as the strongest predictor, followed by land use, altitude, Soil Organic Carbon Index, Landsat 8 band B10 and rainfall. The dominance of temperature for SOC prediction was supported by the strong negative correlation of SOC with temperature (r = -0.582) and positive with altitude (r = 0.580). These underscore the role of climate on the spatial-temporal dynamics of SOC and highlight for climate-smart strategies. Thus, we conclude that cost-effective assessments and monitoring of SOC that support evidence-based strategies for enhancing soil health, land restoration and climate resilience are possible through the developed Cubist and RF models from the Eritrean Central Highlands to the Western Midlands and similar environments.
Climate-smart agriculture, Land use, Machine learning, Temperature, Topography
