Estimation of soil organic matter content by modeling with artificial neural networks

Soil organic matter has direct relationship with soil fertility and quality. However, its estimation in laboratory generates chemical residues which can contaminate the environment, and more ecological methods to determine the soil organic matter present high costs to the laboratories. This study aimed to evaluate the accuracy of artificial neural networks (ANNs) in estimating soil organic matter content from soil chemical attributes and to indicate whether network complexity affects estimation accuracy. A database was created containing 8556 samples, and 75% of the data were used for calibration and 25% for validation of the models. The variables used were: pH, potassium, phosphorus, calcium, magnesium and potential acidity. However, potassium and phosphorus were removed from the input variables. The ANNs were from the Multilayer Perceptron class, with two hidden layers, each of which had number of neurons ranging from 4 to 20. The 15 ANNs with lowest root mean square error (RMSE) were randomly presented by the program Statistica7 (R), and 6 of them were chosen for accuracy assessment. The fits were tested by analysis of variance (F test) and accuracy was assessed based on the coefficient of determination (R-2), RMSE, mean error (ME), index of agreement (d) and confidence coefficient (c). The ANNs showed high accuracy to estimate soil organic matter in the phases of both calibration (R-2 = 0.92; RMSE = 1.82 g kg(-1)) and validation (R-2 = 0.76; RMSE = 1.98 g kg(-1)). Less complex networks can be trained and show the same accuracy as more complex networks.