Document Type : Research Paper
Authors
1 Soil Science Department , Agriculture Faculty, Shahid Chamran University of Ahvaz, Ahvaz, Iran.
2 Department of Soil Science, Faculty of Agriculture, Shahid Chamran University of Ahvaz, Ahvaz, Iran.
Abstract
Introduction: The processes of soil erosion and sediment transport along rivers are the main causes of some socio-economic and environmental problems, such as a reduction in water quality, storage capacity of dams, destruction of aquatic habitats, failure of hydroelectric power plants, and soil degradation. Therefore, understanding the sedimentation status of watersheds is crucial for the effective management of soil and water resources. However, due to the lack of technical and human resources, continuous recording of sediment data is not possible in most sediment measuring stations, and sediment data are recorded only for a few days. In such a situation, a model that can estimate the amount of sediment load using auxiliary variables such as stream discharge and rainfall becomes crucial. Today, it is believed that techniques based on artificial intelligence have a much greater ability to uncover hidden relationships between variables than classical methods and are thus very useful and effective in modeling natural processes.
Materials and methods: In this study, various machine learning techniques, including Artificial Neural Network (ANN), Adaptive Fuzzy-Neural Inference System (ANFIS), and Random Forest (RF), were used for sediment load modeling and sediment forecast for days without measurements. To achieve the research objectives, long-term meteorological and hydrometric data ranging from 2000 to 2020 were collected from related organizations and pre-processed before entering the model. The input variables for the models included 24-hour rainfall, flow rate, normalized difference vegetation index, maximum and minimum temperature, and daily suspended sediment as the dependent variable. Prior to modeling, the entire dataset was divided into two parts, training and testing, in a 70:30 ratio. Relationship modeling was performed using the training data, and model validation was conducted using the test dataset. The efficiency of the models was evaluated using two indicators, the coefficient of explanation (R2) and the root mean square error (RMSE). Additionally, morphometric parameters such as form factor (FF), drainage density (DF) coefficient, and relief ratio (RR) were utilized in modeling.
Results and discussion: The hydrological analysis of the basin revealed that the highest annual amount of rainfall and erosivity index were recorded at the Sheyvand station in the east of the basin, while the lowest values were observed at the Ramhormoz station. The highest average monthly flow rate of 5.8 cubic meters per second was obtained at the Manjeniq station in April, and at the Mashin station, the highest average monthly flow rate of 8.8 cubic meters per second was recorded in December and January. Morphometrically, the studied basin belonged to the class of elongated basins, sloping basins in terms of relief, and the medium class in terms of drainage density. Analysis of the time series of NDVI index showed that the highest vegetation cover occurred in March, while the lowest values were recorded in September and October. The annual trend of the vegetation index indicated an overall improvement in vegetation cover in the region from 2000 to 2020, with the NDVI value increasing from 0.15 to 0.22.
Among the different machine learning techniques studied, the Artificial Neural Network (ANN) model had the highest coefficient of explanation (R2=0.87) and the lowest RMSE for both sediment measuring stations in the region, making it the best model. The optimal inputs for the neural network model at Mashin station were daily average flow adjusted by the basin shape factor, daily rainfall, last day's rainfall, daily minimum temperature and daily maximum temperature. For the Manjeniq station, the optimal inputs were daily average flow, daily rainfall, last day's rainfall, cumulative rainfall for the past two days, and cumulative rainfall for the past three days. The NDVI index was removed from the model due to its low significance. The Random Forest (RF) model ranked second, and the Adaptive Fuzzy-Neural Inference System (ANFIS) model ranked third, with weak performance, especially for the Mashin station, where out-of-range errors occurred.
Temporal analysis of sediment values showed that the highest sediment production occurred in December and January for Mashin station and in April for Manjeniq station. The highest production of sediment occurred in 2006 and 2002, and the trend of changes from 2011 to 2018 showed a decline, attributed to consecutive droughts and lack of rainfall. The annual average sediment production calculated using the values estimated with the neural network model was 88017 tons, equivalent to 1 ton per hectare per year.
Conclusion: Overall, this research demonstrated that machine learning methods, especially the neural network model, are highly effective for modeling and predicting sediment on a daily scale. These methods can compensate for the lack of sediment measuring facilities and equipment in most existing hydrometric stations in the country and eliminate the need for continuous sediment data and other water quality parameters.
Keywords
- Sedimentation
- artificial neural network
- adaptive fuzzy-neural inference system
- random forest
- understimation
Main Subjects
)