Comparing machine learning models for mixed data: Forecasting Arabica coffee yields n in the northern region of Thailand

This research aims: 1) to study the relationship between weather conditions and soil quality with Arabica coffee yield, and 2) to develop a prediction model for Arabica coffee production using decision tree, random forest, and deep learning neural networks with short and long-term memory units. The panel data, collecting from 10 provinces in the Northern region, including Chiang Rai, Lampang, Chiang Mai, Mae Hong Son, Nan, Tak, Phrae, Phayao, Uttaradit, and Phitsanulok, with 9 years from 2013-2021 each was used. The data includes annual Arabica coffee yield, monthly weather data, and soil quality data in 2018. The data was
divided into two sets: 70% for training to construct models and 30% for testing the models.

The findings reveal that the deep learning neural network model with long short-term memory units outperforms than decision tree and random forest with R², RMSE, and r values of 0.91, 9.83, and 0.96, respectively. The most affecting factors on Arabica coffee yield are moisture during the fruit development period, with optimal conditions being a humidity level above 60%. Secondly, rainfall between 1200-1500 millimeters per year during the fruit development stage, and lastly a temperature ranging of 18-28 degrees Celsius during pruning. They are crucial factor on Arabica coffee yield.