Drift-Aware Machine Learning for Precision Agriculture: A Survey on Concept Drift Detection and Adaptive Learning in IoT Environmental Monitoring

The use of algorithms and internet of things sensor networks in precision agriculture continues to grow, However as conditions in the field change over time, so does the original model. This paper will focus on the evaluation of techniques for drift detection and adaptation in precision agriculture systems. In particular, we will examine hybrid approaches that combine various types of sensors in agriculture. We will analyze different techniques for detecting changes in models, including adaptive windowing methods, statistics based approaches (several N-dimensional extensions of Kolmogorov-Smirnov tests), and incremental updating methods (such as using online sequential extreme learning machines). One of the primary contributions of this paper is an integrated drift detection system that is referred to as the Hybrid Adaptive Window (ADWIN) system. This system uses N-Dimensional KS Testing along with Hoeffding's bounds in order to allow for real-time updates of models. As part of this survey, we will also share our analysis about issues of scaling for computation, long term benchmarking against new models and federated deployment challenges. Finally, we will identify several significant limitations associated with edge computing efficiency and privacy and recommend several significant opportunities for future work related to the creation of intelligent, distributed agrometeorological monitoring systems capable of providing timely measurements while conserving limited resources.