Prediction of environment variables in precision agriculture using a sparse model as data fusion strategy

Abstract

Precision agriculture seeks to optimize production processes by monitoring and analyzing environmental variables. For example, establishing farming actions on the crop requires analyzing variables such as temperature, ambient humidity, soil moisture, solar irradiance, and Rainfall. Although these signals might contain valuable information, it is vital to mix up the monitored signals and analyze them as a whole to provide more accurate information than analyzing the signals separately. Unfortunately, monitoring all these variables results in high costs. Hence it is necessary to establish an appropriate method that allows the infer variables behavior without the direct measurement of all of them.This paper introduces a multi-sensor data fusion technique, based on a sparse representation, to find the most straightforward and complete linear equation to predict and understand a particular variable behavior based on other monitored environmental variables measurements. Moreover, this approach aims to provide an interpretable model that allows understanding how these variables are combined to achieve such results. The fusion strategy explained in this manuscript follows a four-step process that includes 1. data cleaning, 2. redundant variable detection, 3. dictionary generation, and 4. sparse regression. The algorithm requires a target variable and two highly correlated signals. It is essential to point out that the developed method has no restrictions to specific variables. Consequently, it is possible to replicate this method for the semiautomatic prediction of multiple critical environmental variables.As a case study, this work used the SML2010 data set of the UCI machine learning repository to predicted the humidity's derivative trend function with an error rate lower than 17% and a mean absolute error lower than 6%. The experiment results show that even though sparse model predictions might not be the most accurate compared to those of linear regression (LR), support vector machine (SVM), and extreme learning machine (ELM) since it is not a black-box model, it guarantees greater interpretability of the problem.