Development of a Deep Learning-Based Smart Irrigation System Using IoT for Automated Crop Water Optimization

Abstract

Conventional irrigation systems account for approximately 70% of global freshwater withdrawals yet routinely waste 25%–40% of applied water through static time-based scheduling that fails to account for dynamic soil and atmospheric variability. This study developed and field-validated a smart irrigation system integrating a multi-sensor IoT network with a hybrid convolutional neural network–long short-term memory (CNN- LSTM) deep learning architecture for automated, data-driven binary irrigation scheduling across three economically important crop types. Over an 18-month field trial (January 2023–June 2024), capacitive soil moisture, DHT22 temperature-humidity, and SR05 pyranometer sensors were deployed across 12 active sensing nodes in 0.5-hectare plots of maize (Zea mays), tomato (Solanum lycopersicum), and wheat (Triticum aestivum) in a semi-arid climate zone, generating 52,416 multivariate time-series observations. On the technical modelling front, benchmarked against artificial neural network, support vector machine, random forest, and standalone LSTM baselines, the CNN -LSTM achieved the highest binary classification accuracies of 94.7%, 93.2%, and 95.6% for maize, tomato, and wheat, respectively, with RMSE values below 0.045 across all crops. On the applied field-performance front, relative to conventional fixed-schedule control plots maintained across four replicates per crop, the system reduced seasonal water consumption by a statistically significant mean of 37.5% (95% CI: 35.1%–39.9%, p < 0.001) and improved end-of-season crop yield by a mean of 21.2% (95% CI: 19.7%– 22.7%, p < 0.001). These results demonstrate the practical feasibility of IoT -integrated deep learning irrigation within a semi-arid, sandy loam experimental context, with potential implications for precision water management in comparable agroclimatic settings pending multi-site external validation.