Agriculture is a significant contributor to global water consumption, making the optimization of irrigation practices essential for sustainability. CropX, an agritech startup, seeks to automate irrigation event detection using large-scale Volumetric Water Content (VWC) data from IoT sensors. Ensuring scalability and accuracy is vital for decision-making within CropX’s farm management system.

This paper describes the development of scalable machine learning (ML) models to automate irrigation detection in large, unbalanced datasets. We discuss the architectural patterns and software design decisions that enable these models to be effectively deployed within an AI-driven system, emphasizing MLOps and CI/CD practices. Multiple ML models were tested, including statistical methods (ARIMA, Kalman Filter), ano-maly detection techniques (Isolation Forest), and ensemble approaches. These were evaluated using performance metrics such as F1 score, precision, and recall. Additionally, we highlight the role of human-in-the-loop strategies in refining model predictions, showcasing the interaction between agronomists and AI-driven recommendations.

Our contribution includes an analysis of the software architecture used for deploying ML models, focusing on microservices, data pipelines, and scalable cloud-based solutions. We illustrate how this system integrates with existing farm management platforms and discuss its implications for future AI-based agent systems in precision agriculture.

The ensemble models achieved superior performance, significantly improving the F1 score compared to individual models. Their integration into CropX’s infrastructure enhanced irrigation event detection while optimizing resource usage. A robust software architecture supports continuous integration and model evolution, ensuring system scalability.