Convergence of IoT AI and natural language processing to support low-literacy rural farmers in early detection of crop diseases: Case study of maize in Tanzania

Abstract

Maize diseases are a significant threat to global yield and food security, especially for smallholder farmers in sub-Saharan countries where maize is a staple food and preferred crop due to its drought resistance. The main challenge in improving maize yields is the early detection of diseases before symptoms become visible on leaves or tissues. Currently, farmers rely on visual observation and the expertise of extension officers or plant pathologists, as do many existing technologies like Deep Neural Networks (DNN) that use images of visible symptoms to identify diseases. However, visual symptoms appear when a disease pathogen has already undergone several asymptomatic phasing events, such as inoculation, penetration, infection, incubation, reproduction, and survival. In that context, performing early detection during asymptomatic phases would be an ideal solution to stop the disease development cycle by optimally applying the needed amount of pesticide or removing infected plants to avoid higher disease intervention costs later. Acknowledging the low cost of emerging Internet of Things (IoT) sensors, nonvisual signs of plant diseases have been explored in this research work for early disease detection. IoT sensors were used for monitoring patterns of the following parameters over time total volatile organic compounds (VOCs), soil chemical properties, ultrasound, temperature, and relative humidity. The study generated a dataset from monitored parameters between healthy maize plants and infected maize plants in a controlled environment. A VOC-based deep learning model was developed and trained on time-series data from healthy maize, achieving 99.98% accuracy in identifying anomalies indicative of disease. When deployed, the model was able to detect anomalies 14 days before visual symptoms appeared, offering a significant improvement over traditional methods where symptoms are typically seen 21 days after infection. Integrated within a Flask web application, the system processes real-time sensor data stored in an SQLite database and deploys OpenAI's API for Natural Language Processing (NLP) to provide detailed analyses and recommendations. Leveraging Twilio's API, farmers receive immediate WhatsApp SMS alerts, facilitating timely interventions. Conversational chatbots based on generative AI enhance this system, offering significant benefits over human extension agents. These chatbots continuously learn and adapt to highly dynamic agricultural contexts, such as climate change, thus providing up-to-date and relevant advice. This robust integration of deep learning, IoT, and NLP technologies offers a practical solution for enhancing agricultural productivity in Tanzania.