Empirical analysis of machine learning-based moisture sensing platforms for agricultural applications: A statistical perspective

Abstract

Modelling of accurate detection & estimation soil moisture sensors requires integration of various signal processing, filtering, segmentation, and pattern analysis methods. Sensing of moisture is generally performed via use of resistive, or capacitive materials, which change their parametric characteristics w.r.t. changes in moisture levels. These sensors are further classified depending upon capabilities of measurements, which include, volumetric sensors, soil water tensor sensors, electromagnetic sensors, time domain reflectometry (TDR) sensors, Neutron probe sensors, tensiometer-based sensors, etc. Each of these sensors are connected to a series of processing blocks, which assist in improving their measurement performance. This performance includes parameters like, accuracy of measurement, cost of deployment, measurement delay, average measurement error, etc. This wide variation in measurement performance increases ambiguity of sensor selection for a particular soil type. Due to this, researchers & soil engineers are required to test & validate performance of different moisture sensors for their application scenario, which increases time & cost needed for model deployment. To overcome this limitation, and reduce ambiguity in selection of optimum moisture sensing interfaces, this text reviews various state-of-the-art models proposed by researchers for performing this task. This review discusses various nuances, advantages, limitations & future research scopes for existing moisture sensing interfaces and evaluates them in terms of statistical parameters like accuracy of detection, sensing & measurement delay, cost of deployment, deployment complexity, scalability, & type of usage applications. This text also compares the reviewed models in terms of these parameters, which will assist researchers & soil engineers to identify most optimum models for their deployments. Based on this research, it was observed that machine learning models are highly recommended for error reduction during moisture analysis. Machine learning prediction models that utilize Neural Networks (NNs) outperform other models in terms of error performance, and must be deployed for high-accuracy & low-cost moisture sensing applications. Based on similar observations, this text also recommends fusion of different sensing interfaces for improving accuracy, while optimizing cost & complexity of deployment. These recommendations are also based on context of the application for which the sensing interface is being deployed. These recommendations must be used to further improve overall sensing performance under multiple deployment scenarios.