Gully erosion susceptibility mapping using four machine learning methods in Luzinzi watershed, eastern Democratic Republic of Congo
Soil erosion by gullying causes severe soil degradation, which in turn leads to severe socio-economic and environmental damages in tropical and subtropical regions. To mitigate these negative effects and guarantee sustainable management of natural resources, gullies must be prevented. Gully management strategies start by devising adequate assessment tools and identification of driving factors and control measures. To achieve this, machine learning methods are essential tools to assist in the identification of driving factors to implement site-specific control measures. This study aimed at assessing the effectiveness of four machine learning methods (Random Forest (RF), Maximum of Entropy (MaxEnt), Artificial Neural Network (ANN), and Boosted Regression Tree (BRT)) to identify gully's driving factors, and predict gully erosion susceptibility in the Luzinzi watershed, in Walungu territory, eastern Democratic Republic of the Congo (DRC). In this study, gullies were first identified through multiple field surveys and then digitized using a very high-resolution image (CNI/airbus) from Google Earth. Overall, 270 gullies were identified, of which 70% (189) were randomly selected to train the four machine learning methods using topographical, hydrological, and environmental factors hypothesized to be gully-related conditioning factors. The remaining 30% (81 gullies) were used for testing studied methods using the threshold-independent area under the receiver operating characteristic (AUROC) and the true skill statistic (TSS) as performance measures. The results showed that RF and MaxEnt algorithms outperformed other methods; performance assessment results showed that the RF model with AUROC = 0.82 (82%) and MaxEnt (0.804: 80.4%) had higher prediction accuracies than BRT: 0.69 (69%) and ANN: 0.55 (55%). TSS results indicated that RF and MaxEnt are best methods in predicting gully susceptibility in Luzinzi watershed. On the other hand, the conditioning factors such as Digital Elevation Model (DEM), Normalized Difference Water Index (NDWI), Normalized Difference Vegetation Index (NDVI), slope, distance to roads, distance to rivers, and Stream Power Index (SPI) played key roles in the gully occurrence. Given the significance of these factors in gullies' occurrence, as shown in this study, policy-makers must adopt strategies that consider these factors to lower the risk of gully occurrence and related consequences at the watershed scale in eastern DRC.