Short Term Soil Moisture Forecasts For Potato Crop Farming: A Machine Learning Approach

Agricultural decision-making is crucial for future yields. In the context of smart farming, grower combine information from sensors located close to their crops with agronomic models to help them to better understand their crops. Irrigation management is therefore based on extrapolation of data and/or agronomic model responses. This problem can be seen as a learning task for which machine learning techniques have proven their relevance in many and diverse applications. In this paper we place ourselves in the context of potato farming, a crop for which irrigation plays a crucial role. We model the problem of soil water potential prediction as a learning problem solved by supervised leaning algorithms. The problem appears to be difficult since there are several potential inputs, and several outputs to predict. Experiments are conducted on several scenarios with data acquired during 3 years. We demonstrate the possibility of applying feature selection method to automatically design models with features relevant to the problem at hand while having good performances. We have also demonstrated the relevance of the machine learning for this kind of problem, since the methods are able to correctly predict the next water potential values.