Research Of Soil Erosion Prediction Modeling Based On Artificial Intelligence In The Central Sichuan Hilly Region

Soil erosion has become one of the most severe ecological environment problems in our country, so the work of predicting it is important for monitoring water and soil loss and evaluating the effect of the soil conservation measures taken. The establishment of soil erosion models is a foundation for implementing the predicting work. At present there are a variety of soil erosion prediction models, such as empirical models, physical models, distributed models, and conceptual models, all of these models require large amounts of a long series of historical data, however, there are shortages of soil and water conservation experimental observation data, due to the experimental observation started late in China. So it has been hotspots and difficulties currently in finding soil erosion accurate prediction models based on shorter series of data.This research took the observational data of the six runoff plots at Suining Soil and Water Conservation Experiment Station, which located in the central Sichuan Hill Region, from 1991 to 2000 to establish two kinds of Matlab-based artificial intelligence Soil Erosion models, which are BP neural network and SVM, then the two models were used for predicting the soil erosion, using five factors, nine factors, ten factors and specific tillage method as input respectively. We explored the advantage, potential, their optimized ways of the BP neural network and SVM modeling predicting.In this research, we quantified the P value of 11 kinds of tillage method such as slope farming, cross ridge in the study area, then used it as one input factor for the BP neural network and SVM predictive modeling, and obtain very good effect.1 For the study area, there are highly significant correlation among rainfall, rainfall duration and soil erosion, and significant correlation between the vegetation coverage and soil erosion. However, the rainfall intensity and other input factors have no significant correlation with the soil erosion.After calculating the value of the soil and water conservation measures factor (P), this research identified the order of soil conservation effect among different tillage methods, which is:cross-slope ridge and digging ditchs in the middle>cross-slope plus ridge with crosspiece>contour strip ridge>idling the land after deep plowing>cross-slope ridge(without crosspiece)>transverse ridge slope angle of 20°with the contour lines> transverse ridge slope angle of 20°with the contour lines>longitudinal ridge>idling the land>transverse ridge slope angle of 30°with the contour lines>longitudinal farming and digging ditchs in the middle. However, the order of their corresponding soil and water conservation measures factor (P values) is in the opposite direction。The P values of the same kinds of tillage methods are distinct when their slope are different, and the P values of slope 15°are 1.46-2.03 times of the P values of slope 10°.2 The predicting results will achieve best when the normalized values range from 0.05 to 0.95.For a single-hidden-layer BP network, by far the number of the hidden layer neurons can not be determined by a fixed formula. We find that when the number of the hidden layer neurons is within 6N (N is the sum of the of input and output factors number) the simulating results is better or the best. The number of training samples and the coefficient of determination of the input factors on the output factors have greater impact on the results when modelling, and the more training samples and greater coefficient of determination, the better the modeling results are.The SVM is more sensitive to the number of training samples and the coefficient of determination than BP network. SVM is better than BP network for the soil erosion prediction under the same conditions, especially when the training samples are large and determination coefficient of input factors on the output factor is high.3 In the central Sichuan Hilly Region, antecedent rainfall factor, soil and water conservation measures factor (P) have a greater effect on soil erosion, therefore, both of them should be quantified and used as input factors of modelling in order to obtain good results; However, for a certain kind of soil and water conservation measures in a single model, using factors like topography, rainfall, antecedent rainfall, vegetation cover as input factors can achieve a satisfactory prediction results when modeling.4 For the slope soil erosion prediction of the central Sichuan Hilly Region, we can get satisfactory results by establishing artificial intelligence-based models such as BP neural network and SVM in predicting soil erosion as long as the training samples are enough and the input factors are appropriately selected.