Risk-averse optimization of crop inputs using a deep ensemble of convolutional neural networks

Abstract Machine Learning algorithms have emerged in precision agriculture as a promising approach for increasing productivity. However, the diffusion of this technology is still limited by the lack of clear applicability for crop input management and by the farmer’s perception of risk. In this work, we tackle both problems by incorporating uncertainty quantification into our previously proposed Convolutional Neural Network (CNN) model for yield prediction, and by proposing a risk-averse optimization algorithm on top of it. We redesign our CNN architecture under the Deep Ensemble framework, so the predictive model outputs a probability distribution instead of a single value. Then, a gradient-based optimization algorithm uses this model to find the maps of crop inputs that maximize the expected net revenue while satisfying risk constraints. We show that the new model not only provides uncertainty quantification but also increases the predicted performance of its former version. Experiments with the optimization algorithm show an increase up to 6.4% in the expected net revenue when compared with the status quo management, and provide a flexible setup to match different levels of risk aversion.