A Modern Way to Establish Smart Irrigation Using Dense Learning Model in IoT Environment

Abstract

Lack of water is a major challenge in agriculture to guarantee food supply stability. Smallholder farming communities (SFCs) often hesitate to embrace digital solutions due to high costs, implementation complexities, and faulty sensors. To address this, we developed a low-cost (~USD 260), open-source "Intelligent Irrigation in a Box" IoT framework. The system uses an ESP32 microcontroller with LoRaWAN wireless telemetry, capacitive soil moisture sensors (inserted at 5 cm and 20 cm depths), and DHT22 air temperature/humidity sensors. To ensure sensor reliability, we present a DenseDarkNet-53 deep residual network that identifies time-series abnormalities and reconstructs corrupt sensor readings. The dataset comprises 2,000 hourly time-series samples validated by agronomy experts into balanced normal and anomalous classes. Additionally, a CNN-based model predicts future short-term trends to maintain irrigation schedules. Experimental results yield evaluation accuracies of 95.9% for air temperature, 93.5% for soil moisture, and 97.6% for air humidity, providing a reliable data foundation for smallholder irrigation decisions.