Modelling of soil erosion susceptibility incorporating sediment connectivity and export at landscape scale using integrated machine learning, InVEST-SDR and Fragstats.

ABSTRACT

Evaluating the linkage between soil erosion and sediment connectivity for export assessment in different landscape patterns at catchment scale is valuable for optimization of soil and water conservation (SWC) practices. Present research attempts to identify the soil erosion susceptible (SES) sites in Kangsabati River Basin (KRB) using machine learning algorithm (decision trees, decision trees cross validation, CV, Extreme Gradient Boosting, XGB CV and bagging CV) taken thirty five variables, for investigating the linkage between erosion rates and sediment connectivity to assess the sediment export at sub-basin level employing connectivity index (IC) and Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) sediment delivery ratio (SDR) model. Based on AUC of receiving operating curve in validation test, excellent capacity of extreme Gradient Boosting, XGB CV and bagging CV (0.95, 0.90) than decision tree and decision tree CV (0.78, 0.82), exhibits about 18.58 % of basin areas facing susceptible to very high erosion. Conversely, considering universal soil loss equation (RUSLE) parameters, InVEST-SDR model estimated about 64.24 % of soil loss rate occurred from high SES in where sediment export rate become very high (136.995 t/ha-1/y-1). The IC result show that high sediment connectivity (<-4.4) measured in high SES of laterite and bare land in upper catchment, and double crop agricultural areas in lower catchment, while least connectivity (>-7.1) observed in low SES of dense forest, vegetation cover and settlement built-up areas. Pearson correlation matrix revealed that four landscape indices category i.e. edge metrics (p < 0.01), aggregation metrics (p < 0.001), shape metrics (p < 0.01-0.001) and diversity metrics (p < 0.01) signified the influence of landscape patterns on IC and SES. Accordingly, RUSLE, SDR and landscape matrices reveals that maximum sediment export rate associated with high connective delivery outlet and high SES in laterite, double crop and bare land due to simple landscape and greater homogeneity, whilst minimum export rate related with low connectivity and low SES in dense forest, vegetation cover and settlement built up area causes of fragmented landscape and spatial heterogeneity. Finally, findings could immense useful for formulating the optimizing measures of SWC in the watershed.