ASABE Technical Library - Abstract

Regional Variability of Topographic Index and Machine Learning Model Applications for Prediction of Ephemeral Gullies

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  Soil Erosion Research Under a Changing Climate, January 8-13, 2023, Aguadilla, Puerto Rico, USA  .(doi:10.13031/soil.23037)
Authors:   Christian Conoscenti, Aleksey Y Sheshukov
Keywords:   Ephemeral gully, Erosion, Machine learning models, Topographic indices.

Abstract.

Classical and ephemeral gully erosion are the hillslope processes responsible for land degradation under a wide range of environmental conditions. Ephemeral gullies (EG) are small channels formed in a cultivated field that can be filled by tillage and reappear again in the same location due to a consequent runoff event and specific topographic features. Land susceptibility to gully erosion is crucial for understanding driving factors and developing effective erosion control practices. Topographic index (TI) models aggregate morphometric features of a landscape and use them to predict trajectories and initiation points of classical or ephemeral gullies. As an alternative to the index-based models, statistical machine learning approaches (MLA) have recently been gaining increased attention for identification of areas for gully susceptibility.

Application of index-based or statistical models is normally restricted to a single area of study, thus the objectives of the study were to explore the ability of TI and MLA models to predict EG trajectories and evaluate model transferability within and across two regions. Seven TI models and two MLA algorithms were applied to eight watersheds in Kansas, USA, and two watersheds in Sicily, Italy. TI models included Compound Topographic Index (CTI), Stream Power Index (SPI), Topographic Wetness Index (TWI), Stream Grid Order Index (GORD), and their three modifications using the convergence factor. MLA methods included Multivariate Adaptive Regression Splines (MARS) and Random Forests (RF) models. The predictive ability of the models was measured by using both cut-off independent (area under the receiver operating characteristic curve, AUC) and dependent statistics (Cohen's kappa index, sensitivity, specificity). The performance statistics of both topographic indexes and statistical models were cross-examined by finding an optimal threshold in one watershed and applying it to the other watersheds. The predictive ability of the models ranged from insufficient (<0.7) to outstanding (0.9 < AUC < 1.0). Overall, CTI showed the worst performance among the TI models, and GORD had the best ability to discriminate between gully and non-gully pixels with average AUC across all models and all watersheds of 0.962. In Kansas, the stream order (GORD) model was found to perform similar to machine learning approaches, whereas the modified stream power index (MSPI) model overperformed other topographic index models in Sicily. The boxplots of AUC in watershed 3 in Kansas are presented for eight models in Figure 1. In terms of the model transferability, all models showed adaptive variability of the applied thresholds in the other watersheds with values varying from slightly positive (CTI, SPI, MSPI) to close to 1 (GORD, MARS, RF). Again, the GORD model performed better than the other TI models and had the lowest variability with the performance similar to the MLA models when gully predictive models are transferred among all the watersheds. In summary, different landform characteristics in cultivated areas of the High Plains region in Kansas and in the steep hillslopes of Central Sicily caused models to have variable success rates when the indexes from one region were transferred to another region. The results also showed that well-calibrated topographic index models, especially the GORD model, can be viewed as a valid alternative to a data driven approach for ephemeral gully mapping, but transferability of the optimal thresholds can be applied with caution. This work was supported by the USDA-NIFA Hatch Project No. S1089.

(Download PDF)    (Export to EndNotes)