Accelerated soil water erosion in agricultural hilly areas impacts negatively on crop yield, surface water resources and territorial infrastructures [3]. In the context of an increasing anomaly of rainfall patterns and water scarcity due to climate change, reducing soil erosion by sustainable management practices is a priority for Europe [4]. Soil erosion is influenced by several factors, both natural and human-driven. In the last decades, empirical models have been developed and largely used to assess soil erosion, which is the first step in soil conservation. Machine learning techniques offer new possibilities to face the quantification of soil erosion, based on agro-meteorological data available from existing databases and field monitoring. The objective of this investigation was to estimate single event soil loss and runoff using a machine learning approach. The inference was developed using 20-years data collected on an hydraulically bounded vineyard plot in Piedmont, North Italy. At first, rainfall erosivity for each event was derived from hourly precipitation records, using well calibrated conventional model and a simple feed-forward neural network. Then, the nonlinear relationship between hydrological variables (soil loss and runoff) and rainfall characteristics (rainfall amount, rainfall duration, maximum intensity, derived rainfall erosivity) was tested using a decision-tree-based ensemble machine learning algorithm, namely XGBoost [1]. The ensemble and the training/test set splitting were designed to deal with the difficulties of managing small dataset. Results are encouraging: the out-of-sample R2 between the simulated and the observed soil loss values reached 0.33, while the runoff of unseen events was estimated with a mean absolute error of 14 mm. Furthermore, we studied the explicability of the model using the SHAP method [2]. In this way, we identified the most influential features in predicting each dependent variable. Our findings could contribute to the development of machine learning applications in the context of soil erosion related problems.
Predicting soil loss and runoff in vineyards using a gradient boosting framework
Marcella Biddoccu;Eugenio Cavallo
2022
Abstract
Accelerated soil water erosion in agricultural hilly areas impacts negatively on crop yield, surface water resources and territorial infrastructures [3]. In the context of an increasing anomaly of rainfall patterns and water scarcity due to climate change, reducing soil erosion by sustainable management practices is a priority for Europe [4]. Soil erosion is influenced by several factors, both natural and human-driven. In the last decades, empirical models have been developed and largely used to assess soil erosion, which is the first step in soil conservation. Machine learning techniques offer new possibilities to face the quantification of soil erosion, based on agro-meteorological data available from existing databases and field monitoring. The objective of this investigation was to estimate single event soil loss and runoff using a machine learning approach. The inference was developed using 20-years data collected on an hydraulically bounded vineyard plot in Piedmont, North Italy. At first, rainfall erosivity for each event was derived from hourly precipitation records, using well calibrated conventional model and a simple feed-forward neural network. Then, the nonlinear relationship between hydrological variables (soil loss and runoff) and rainfall characteristics (rainfall amount, rainfall duration, maximum intensity, derived rainfall erosivity) was tested using a decision-tree-based ensemble machine learning algorithm, namely XGBoost [1]. The ensemble and the training/test set splitting were designed to deal with the difficulties of managing small dataset. Results are encouraging: the out-of-sample R2 between the simulated and the observed soil loss values reached 0.33, while the runoff of unseen events was estimated with a mean absolute error of 14 mm. Furthermore, we studied the explicability of the model using the SHAP method [2]. In this way, we identified the most influential features in predicting each dependent variable. Our findings could contribute to the development of machine learning applications in the context of soil erosion related problems.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
