IoT and AI: A Panacea for Climate Change-Resilient Smart Agriculture

Abstract

The application of Internet of Things (IoT) and Artificial Intelligence (AI) for disaster preparedness and sustainable agriculture has been a topic of great interest lately. In the last few years, extreme weather swings due to climate change caused by global warming have caught the farming community off guard, especially in the developing world. One of the key objectives of smart agriculture is optimal use of freshwater, which has become an increasingly scarce resource around the world. Reference Evapotranspiration (ETo), an estimation of total flux of water evaporating from a reference surface is an important parameter for irrigation management. IoT & AI-based location-specific estimation of ETo for crop water requirements augments the decision-making process. In this work, we utilize the Hargeaves and Samani (H-S) model and six regression algorithms for the estimation of ETo. We create a location-specific dataset with locally sensed IoT data from a flood warning system and remotely sensed meteorological data, spanning over 5 years. We train and test Linear Regression (LR), Multilayer Perceptron (MLP), Radial Basis Function (RBF), Support Vector Regression (SVR), Bagging and Random Forest (RF) algorithms on the locally curated dataset with 20 basic, extracted, and derived attributes. We gradually reduce number of attributes in the dataset from 20 to 3 and compare performance of the six algorithms using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Relative Absolute Error (RAE), Root Relative Squared Error (RRSE), Coefficient of Determination R2, Kendall Tau and Spearman Rho metrics. SVR shows superior performance with an MAE of 0.03 and an RMSE of 0.05, followed closely by MLP with an MAE of 0.04 and RMSE of 0.06 with a dataset of 12 attributes. The performance of Bagging and RF algorithms remains relatively unchanged with feature reduction whereas RBF shows slight improvement in performance when number of attributes is reduced to 3. Finally, we develop a novel ensemble hybrid model using the Stacked Generalization technique, which outperforms all individual models in prediction accuracy when using reduced-feature datasets. This work clearly delineates the performances of a diverse set of ML algorithms for feature-rich and feature-scarce scenarios and demonstrates the efficacy of our hybrid ensemble ML algorithm for estimating ETo under limited availability of data in resource-constrained environments.