Gully erosion is a primary driver of sediment load in freshwater catchment systems. This is particularly the case for the Murrumbidgee catchment, where gullies are the greatest contributor of instream sediment. In large catchments, an understanding of gully location and intensity can be hard to determine, making it hard to generate accurate sediment models and develop plans for erosion management. The Upper Murrumbidgee catchment exemplifies this, with large areas of gully erosion, substanially instream sediment deposition, and drastically reduced natural flow.
This study aims to address this issue by building on previous work done in the catchment by combining high-reslution remote sensing with machine learning. Using QGIS software, the catchment was divided into a 1km grid. Gullies were manually digitised over a high resolution imagery layer across 2% of the catchment. The gully density (total length of gullies per km2) was calculated for each measured cell. Gully density for the remainder of the catchment was then predicted using a Random Forest (RF) model using variety of publically available spatial data such as soil texture, land use, lithology, weather, and topographical data as independent variables. The RF model produced moderate predictive performance (R2 = 0.4 – 0.5) highlighting the complexity of fluvial erosion systems. Next, an XGBoost classification model was generated to predict the presence or absence of gullies in each 1km2 cell, producing strong performance (AUC = 0.8 – 0.9). However, restricting density predictions to areas classified as containing gullies did not improve overall model performance. These findings demonstrate the potential of machine learning for regional-scale gully mapping while also highlighting the challenges of predicting gully erosion intensity. The resulting gully distribution maps provide an important input for sediment budget modelling and support targeted erosion management within the Upper Murrumbidgee catchment.