By Dave DeFusco
When Mwansa Phiri began studying artificial intelligence at the Katz School, he didn鈥檛 expect that his coursework would lead him back to a problem unfolding thousands of miles away in Africa: how to help farmers grow enough food in the face of drought, flooding and tightening regulations on water and fertilizer use.
Phiri, a student in the Katz School鈥檚 M.S. in Artificial Intelligence, is developing a project called Smart Farming: A Machine Learning Approach to Crop Growth Prediction. At its heart, the project aims to ensure food security across Africa by giving farmers better information.
He was inspired by the work of Zambian agritech entrepreneur Nchimunya Munyama, 鈥攂orn from the challenges his grandfather faced and supported by mentorship and innovation programs鈥攚as initially trying to solve similar farming problems and motivated Phiri to choose this project.
鈥淪ince we鈥檙e in the United States, we don鈥檛 really get to hear what鈥檚 going on back home,鈥 said Phiri. 鈥淣chimunya came to visit us in the States and told us how difficult it is for farmers to know what to grow. They rely on generational knowledge鈥攚hat their parents always planted鈥攂ut climate conditions are changing.鈥
Artificial intelligence allows computers to learn patterns from data and make predictions. Machine learning, a branch of AI, trains computer models to recognize relationships, such as how soil moisture, rainfall and fertilizer levels affect crop growth. Phiri realized that this technology could help farmers 鈥渢hink outside the box,鈥 as he put it, by recommending crops that match current soil and weather conditions rather than tradition alone.
The work was not done alone. Phiri collaborated closely with AI students Jelidah Nayingwa and Esperance Tuyishime, who helped train, test and refine the machine learning models. 鈥淲e approached it as a team,鈥 said Phiri. 鈥淛elidah, Esperance and I used this project to help figure out how to finalize the model in a way that would actually work in real farming conditions.鈥
Their system combines small, affordable Internet of Things (IoT) devices with machine learning models. The IoT devices, equipped with sensors that measure soil moisture, temperature and humidity, are placed in fields. They send data through a Wi-Fi module to a cloud-based platform for analysis. From there, machine learning models predict three key outcomes: which crops are best suited to a field, when to plant them and how much yield to expect.
鈥淚t helps them with utilization,鈥 said Phiri, referring to new restrictions on fertilizer use in some African countries. 鈥淔armers weren鈥檛 given training on exact amounts. They just had a standard practice鈥攖hrow everything on the ground and hope it grows. With the system, you can monitor how much fertilizer or water is actually needed and track what worked well before. That way, you don鈥檛 waste resources.鈥
To train the system, the team worked with multiple agricultural datasets containing information on soil pH, rainfall, irrigation, fertilizer use and crop types. One major challenge was regional variation.
鈥淲hen some datasets wrote 鈥榤aize,鈥 I assumed it was the standard maize we have back home,鈥 he said. 鈥淏ut there are different variations. Some data came from Kenya, and the crops performed differently than we expected.鈥
To solve this, they standardized the data鈥攕ometimes treating similar crops as entirely separate plants鈥攖o ensure the models learned accurate patterns. They also engineered new features, such as calculating rainfall per day rather than total rainfall, to better capture how weather affects growth.
The project addresses two kinds of predictions. First, classification: determining whether a crop is suitable for a particular field. Second, regression: estimating how much yield a farmer can expect. After testing several machine learning models, including Random Forest, Support Vector Machines and Neural Networks, Random Forest performed best for crop suitability. When he tried using fewer data features, accuracy dropped sharply.
鈥淚t just showed that we needed more data,鈥 said Phiri. 鈥淚f you try to do it with less data, you might give results that people wouldn鈥檛 be happy with. We wanted to avoid telling farmers a crop would work and then having it fail.鈥
Accessibility is central to the project鈥檚 mission. While the system includes a mobile app with dashboards and predictive charts, Phiri鈥檚 team also built a text-based feature for farmers who use basic phones. 鈥淭he IoT device can send a summary by text message,鈥 he said. 鈥淭hat way, farmers don鈥檛 need smartphones or training on complicated interfaces.鈥
Looking ahead, Phiri hopes to integrate satellite imagery and drone data to monitor plant health using vegetation indexes. That may require more advanced deep learning models. 鈥淲e would redesign a new model at a larger scale,鈥 he said.
For Honggang Wang, chair of the Graduate Department of Computer Science and Engineering, the research demonstrates how AI can address urgent global challenges.
鈥淭his project shows the power of artificial intelligence when applied to real-world problems,鈥 said Wang. 鈥淏y combining IoT sensing, data analytics and machine learning, Mwansa鈥檚 work has the potential to make agriculture more resilient, sustainable and profitable, especially in regions where food security is fragile.鈥
The early results are promising鈥攊mproved yield prediction accuracy and better resource optimization, but moving from research to reality will require pilot deployments, investors and policy support. For Phiri, the motivation remains personal.
鈥淎griculture production is essential for food security,鈥 he said. 鈥淚f we can give farmers better tools to make decisions, we can help make farming smarter and help ensure there鈥檚 enough food for everyone.鈥
