Optimization of loss functions for predictive soil mapping

Abstract

Loss function is an integral part of any machine learning algorithm. Loss function is used to measure the current performance of machine learning model during training. Choosing a relevant loss function is therefore important. A better loss function can drive the model towards a good sub-optimal solution. Machine learning has been used to model relations between variables of complex systems, e.g. soil. Using a good machine learning algorithm, we can predict various soil properties like pH, soil electrical conductivity (SEC), etc. Prediction of these variables is important since this can help in deciding which food crops should be planted. SEC refers to the amount of soluble salt present in the soil. If the soil contains too much salt, then the vegetation of the area suffers. Also, the prediction of soil electrical conductivity is important because it is difficult to visit each site to estimate it. Using machine learning models to predict SEC, without any in-situ analysis, we can get an intuitive idea about the value of SEC of the soil. An attempt to predict SEC using neural network model is done in this study. To train the machine learning model, several loss functions were optimized that are generally used for prediction in machine learning—mean squared error loss (MSE or L2 loss), mean absolute error loss (MAE or L1 loss), Huber loss. All the loss functions—mean squared error loss, mean absolute error and Huber loss along with different optimizers like Adam and stochastic gradient descent (SGD)—were experimented in the model. Many techniques like dropout, batch normalization and adaptive learning rate were also attempted in order to improve the model. The error metric for evaluation of our predictions with the actual values was MSE. After a lot of iteration and optimizations, the best model estimated MSE of 0.029.