SMART IRRIGATION STRATEGY RECOMMENDED SYSTEM USING ML

Abstract

Water used for irrigation forms the largest share of freshwater consumption in agriculture, and much of it is wasted because fields are watered on a fixed schedule rather than according to the actual needs of the crop. Over-irrigation leaches nutrients, raises pumping cost, and can damage roots, while under-irrigation reduces yield during critical growth stages. This paper presents a Smart Irrigation Strategy Recommendation System that uses soil and weather sensing together with machine learning to recommend when to irrigate, how much water to apply, and which irrigation method suits the field and crop. The sensing part of the system consists of field nodes built around an ESP32 microcontroller that read a capacitive soil moisture sensor at two depths, a soil temperature probe, and an air temperature and humidity sensor. Each node samples these values at regular intervals, averages several readings to reduce noise, and sends them over Wi-Fi or a LoRa link to a gateway. The nodes are powered by a small solar panel and battery, and they spend most of their time in deep sleep to extend battery life. Weather forecast data for the location, including expected rainfall, temperature, and wind speed, is obtained from a public service and merged with the sensor readings. The machine learning part of the system operates in two stages. In the first stage, a regression model predicts the soil moisture level expected over the next twenty-four hours using current sensor readings, reference evapotranspiration computed from weather data, crop type, and growth stage. Linear Regression, Random Forest, Gradient Boosting, and Support Vector Regression were compared for this task using RMSE and MAE. In the second stage, a classification model recommends one of several irrigation actions, namely no irrigation, light irrigation, normal irrigation, or deferral because rain is expected. Decision Tree, Random Forest, K-Nearest Neighbours, and XGBoost classifiers were compared using accuracy, precision, recall, and F1-score. The recommendations are presented through a web dashboard that shows live soil moisture for each field zone, the predicted moisture curve, the recommended action with the estimated water volume and pump running time, and a record of past irrigation events. The system can also switch a relay-controlled pump or solenoid valve automatically when the farmer chooses automatic mode, while manual confirmation remains the default. In trials using collected and public agronomic data, the Random Forest models gave the best results, and the recommended schedules used noticeably less water than a fixed daily schedule while keeping moisture within the desired range. The system is intended for small and medium farms, kitchen gardens, polyhouses, and institutional campuses where water availability is limited and labour for manual checking is scarce. It combines inexpensive electronics, low-power wireless communication, and interpretable machine learning in a single practical tool. Future work includes crop-specific models trained on local field trials, integration of satellite vegetation indices, and support for drip fertigation scheduling.