Using Artificial Neural Network for Evaluating the Relationship between Weather Parameters and Soil Moisture

• Main Authors: Hie Missa, 米沙
• Other Authors: Wang Yu-Min
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/47qr77

碩士 === 國立屏東科技大學 === 土壤與水工程國際碩士學位學程 === 105 === As a major state variable of the land surface-atmosphere system and agriculture, the estimation of soil moisture (SM) using weather parameters and the knowledge of their interaction is of great importance for scientists and practitioners in many related fields. Accordingly, the major objective of this study was to analyze the variability of SM throughout a loamy soil, at 10 cm (SM10), 20 cm(SM20), and 30 cm (SM30) depth and its relationship with climate variables. This study assessed the capability of a Multi-Layer Perceptron (MLP) neural network model in comparison to a Multi Linear Regression (MLR) used as baseline, for estimating SM with weather parameters comprising maximum air temperature (Tmax,), minimum air temperature (Tmin), average air temperature (Tavg) , dew point temperature(DewPt), average relative humidity(RHavg), rainfall(Rain), mean wind speed(WSmean), maximum wind speed(WSmax), average wind speed (WSavg), wind direction (WDir), Sunshine(Sun), solar radiation(Rad), and atmospheric pressure (APress) as predictors, and examined their influence on SM predictive accuracy using a sensitivity analysis to identify the key input weather variables which can efficiently predict SM. After assessing the linear correlation between all variables, the above-mentioned parameters were first used for SM simulation, then a sequential feature selection (SFS) algorithm was applied to identify the relevant ones that contribute effectively to SM variation. The results indicated that the linear correlation between some weather parameters was very strong but, remained very weak with SM variables (correlation coefficients less than 0.40) indicating the non-linearity of soil-atmosphere interaction. It was also found that MLP outperforms MLR in both simulation scenarios: (1) using all the 13 initial and then (2) the relevant variables as predictors. Simulations showed that a MLP (9-25-1), using 9 relevant variables obtained from the sequential feature selection and ranked as Rain, Tavg, DewPt, RHavg, APress, WSavg, WSmax, WDir, and Rad out of the 13 initial input variables, can estimate SM efficiently and accurately with decreasing R values (0.913; 0.873; 0.848) greater than 0.70 (from previous studies) and increasing MSE (0.009 cm3/cm3; 0.017 cm3/cm3; 0.028 cm3/cm3) during the test stage for SM10; SM20, and SM30 respectively. The Research also outlined a procedure to evaluate the effects (relative importance) of each predictor on SM based on the resulted connection Weight of MLP (9-25-1) model which yielded Rainfall to be the most influential parameter and revealed that 8 variables including Rain, Tavg, DewPt, RHavg, APress, WSavg, WSmax, and WDir, statistically, explain 96% of the variance in the dataset based on their relative importance and can constitute key SM predictors in limited climatic data condition. Furthermore, it was found that by eliminating wind variables from the SFS result, a MLP (6-10-1) estimates SM with a quite good performance having R values of 0.802; 0.780; 0.760 and MSE of 0.020 cm3/cm3; 0.028 cm3/cm3; 0.034 cm3/cm3, both for SM10; SM20, and SM30 respectively. Such an analysis performed on the ANN-SM model developed was able to explain the reasons for the ANN’s potential and robustness in estimating SM effectively from weather parameters and can serve for understanding the complex process of soil-atmosphere interaction.