What question did this study set out to answer?

This study aims to evaluate the impact of different land uses on soil carbon levels in Eritrea's drylands.

April 3, 2026Open Access

Estimating carbon sequestration potential of dryland enclosures: A comparative assessment assisted by Sentinel 2 time-series and machine learning framework

Key Points

This study aims to evaluate the impact of different land uses on soil carbon levels in Eritrea's drylands.
Compared four land uses including enclosures and grazing
Utilized Sentinel-2 time-series data and Google Earth Engine
Developed soil carbon predictive models using machine learning techniques
Applied Kruskal-Wallis and Dunn tests for statistical assessment
Permanent enclosures showed the highest median SOC (1.28%) and SOCs (37.66 Mg ha-1).
Conservation enclosures increased soil carbon significantly compared to grazing and rainfed cropping.
Predictive accuracy for SOC was excellent (R2 = 0.85) using Ridge Regression.
Optimal season for SOC prediction was winter, with Boruta features yielding the best results.

Abstract

In Eritrea, in response to widespread land degradation, rehabilitation works are underway. However, the impacts of such works on soil carbon remain unknown. This study assessed the impacts of enclosures by comparing four land uses: permanent enclosure, oxen enclosure, communal grazing, and rainfed farming. Seasonal Sentinel-2 time-series data were extracted using Google Earth Engine and soil organic carbon (SOC) concentration and stock (SOCs) predictive models were developed using soil, topographic, and Sentinel 2 data using machine learning. We compared Random Forest versus Ridge Regression models across four seasons with correlation-based versus Boruta algorithm-based selected features. Kruskal-Wallis H and Dunn post-hoc tests showed that conservation enclosures significantly enhanced soil carbon compared to grazing and rainfed cropping lands. The highest median SOC (1.28%) and SOCs (37.66 Mg ha-1, up to 30 cm soil layer) were found in the permanent enclosure followed by the oxen enclosure, while rainfed farming had the lowest values (0.45% and 14.85 Mg ha-1). These improvements correspond to potential carbon sequestration gains of up to 153.6% when converting degraded land to permanent enclosure. Modelling revealed that SOC was predicted with excellent accuracy (R2 = 0.85, RPD = 2.57) using Ridge Regression and with very good accuracy (R2 = 0.80, RPD = 2.22) with Random Forest with Boruta selected winter season features. For SOCs, the best models of both models achieved similar fair accuracy (R2 = 0.60, RPD = 1.58). Key predictors for SOC were bulk density, Sentinel 2 bands B5 and B8A, while for SOCs B5, Soil Organic Carbon Index (SOCI), B11, and soil pH. Winter (Jan-Mar) was the optimal window season and Boruta-selected features gave the best results for SOC and SOCs prediction. The study demonstrates that enclosure systems are effective for soil carbon restoration in dryland landscapes. This framework highlights the potential for cost effective satellite based monitoring of restoration outcomes. These findings provide evidence based support for scaling up enclosure based restoration in Eritrea and similar dryland regions, contributing to both land degradation neutrality and climate change mitigation goals.

KI fragen

Bookmark

View Full Paper