A Decision Model Based on an Optimized Choquet Integral: Multifactor Prediction and Intelligent Agriculture Application

Abstract

Smart farming aims at improving agriculture production by using artificial intelligence and smart devices and, in continuity, farming optimization aims at supporting autonomous decision-making to maximize crop yield. In this context, the question of predicting the future days of growth stages transition of a plant is still a challenge, as existing automated predictions are not accurate nor reliable enough to be effectively used in the farming process. We propose here an approach based on Choquet integral, performing an aggregation of multiple imperfect predictions into a more accurate and reliable one, considering the specific relevance of various prediction sources as well as interactions, synergies, or redundancies between them. To identify the numerous parameter values defining the Choquet-based decision model, we propose a generic approach of optimization based on observed history, ensuring a reduced sensitivity to parameters, thanks to a principle of less specificity. Our proposal defines so an evaluation function assigning to any potential solution a predictive capability, quantifying the conformance of its outputs to evidence, as well as an associated optimization process based on the satisfaction degrees regarding a set of stated inequalities. The case study concerns an implemented prototype that enables, for a given culture and several input sources, to help farmers, providing them with better predictions of the growth stages. We also analyze the reliability of the process, enabling the assignment of an objective probabilistic criteria to any provided prediction. The experimental results are very encouraging, the predicted day remaining stable despite presence of noise and local errors.