Abstract

In addition to effectiveness, practicality and efficiency have been considered crucial when considering an automated irrigation system. Awareness of such requirements has only increased since freshwater resources are becoming scarce, particularly in many agricultural regions of Vietnam. A considerable amount of effort has been put into creating approaches to solving these problems, which can be classified into two main approaches: Supervised learning, and reinforcement learning. Ordinary supervised learning approaches tend to rely on input from farmers and experts' knowledge. However, such approaches may lead to inaccuracy due to human over-estimation or underestimation of the amount of water needed, thus leading to resource waste and ramping up production costs. In contrast, reinforcement learning methods have proven to be efficient given their ability to hastily adapt to new changes or trends in the environment. But despite the benefits, its need for a reliable simulation system and commitment of time for running through trial-error steps has rendered it impractical for real-world uses. Moreover, deployment of such methods encounters resource-wise and architecture-wise setbacks. This paper proposed a simple mixture of said approaches that attempt to adapt the environment to a desired state. This paper also presented an overview of the environment settings and the system architecture in which the proposed method will be deployed in a way that the method can interact with the states of the environment. Our approach is also deployable on machines with limited computing power, does not require pre-configurations in a simulated environment, and the need for human intervention is minimal. The performance evaluation of the proposed method is also presented and shows remarkable improvement of the method over a set of data gathered from the environment.