Abstract
Despite the importance of delineating spatial modelling of gully headcuts (GHs) in erosion-prone environments, assessments of the factors that control the occurrence of headcuts is lacking. To fill this gap in the research, we identified 129 GHS field surveys. These 129 cases were randomly divided into two groups: 90 GHs (70%) for model training and 39 GHs (30%) for model validation. Subsequently, new unmanned aerial vehicle (UAV) imagery is used to develop spatial modelling to predict the location of GHs at sites prone to soil erosion in Golestan Province, Iran. Mapping GHs enables evaluation of 4 machine-learning techniques (or ensembles) – best-first decision tree (BFTree), bagging best-first decision tree (Bag-BFTree), random-subspace best-first decision tree (RS-BFTree), and rotation-forest best-first decision tree (RF-BFTree) – for modelling GHs. We use the information-gain ratio method to analyze the relationships between GHS and 22 GH conditioning factors. The 4 ensemble outputs are validated using a receiver operating characteristic (ROC) curve. The areas under the curves (AUCs) for prediction rates of the ensemble methods applied to the training group are BFTree – 88.3%, Bag-BFTree – 92.7%, RS-BFTree – 95.7%, and RF-BFTree – 93.2%. The AUCs for the model-validation group cases, however, are 84.9%, 94.1%, 97.4%, and 9.18%, respectively. Therefore, RS-BFTree is, statistically, the most effective ensemble method for accurate modelling of GHs. Variable-importance analyses using information-gain ratio indicate that out of 22 GH-influential factors, land use, slope degree, and slope-length are of more importance in developing of GH occurrence. Finally, to address the need for detailed observations and highly accurate erosion data in the field, UAV image-acquisition technologies are demonstrated.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call