The Department of Agriculture Extension (DOAE) in Thailand has taken a step forward by implementing Artificial Intelligence technologies to aid farmers in the country. The Personalized Data system, developed by DOAE, will help millions of Thai farmers receive tailored suggestions in a cost-effective and timely manner.
The project’s goal was to bring AI advancements to assist farmers in making better decisions, including what crops to grow and where to sell them. The main objective was to provide simple, precise, and succinct information to farmers, so it needed to find an appropriate tradeoff between computing time and available computation time. The chosen algorithms had to provide good results robustly in a real-world environment.
The DOAE collects data from seven large databases, which have been gathering information from a real-world environment over several decades. There are fifteen plants that are enlisted as Thailand’s prime crops, but we only focused on rice as it is the main crop in the country. The research shows how to deliver the required agricultural extension capabilities with minimized training time. There was analyzed three key factors – price, cost, and yields of crops – to make predictions simple but meaningful to farmers.
It was discovered that an artificial neural network with Multilayer Perceptron (MLP) and Random Forest (RF) models could effectively predict the yield, cost, and price of crops. Adjusting parameters such as the learning rate and the number of hidden nodes affects the accuracy of crop yield predictions. Smaller data sets required fewer hidden layers in model optimization. MLP models consistently produced more accurate yield predictions than RF models. Although RF is considered a less accurate method compared to MLP, it works fine in most cases.
MAPE (Mean Absolute Percentage Error) is used to measure both RF and MLP. MLP models produced accurate predictions, with an accuracy rate of around 5% in most cases. Similarly, RF was used to provide suggestions when time constraints were tight, and the accuracy was still high, with an MAPE of around 5%. In some difficult scenarios where data is not adequate, the results were still good, and MAPE was around 20%.
More advanced learning techniques can be analyzed and installed in the system. The involved factors should also be taken into account because more external data sources will be connected to the system.