Abstract:
Rice is an important crop in the Philippines, and its production is one of the most
vulnerable to climate change. Various works studied the relationship between
those climate parameters and rice yield, highlighting their relevance to each other.
This led to the development of different models and architectures to analyze the
multivariate time series data. However, those are purely statistical. Moreover,
the Philippines have limited accessible tools for nowcasting rice production. This
study aims to develop a machine learning-based nowcasting tool (web application)
for rice production with climate parameters as the factors. The analysis
used Bukidnon’s quarterly climate and rice yield dataset from 2000 to 2019. The
ARDL model identifies the significant lags and variables for the hybrid model.
The study compared seven models: GPR, BVAR, XGBoost, RF, RNN, LSTM,
and CNN; additional models include the ARDL and the ARDL-BVAR. These are
evaluated using APE, MSE, RMSE, and MAPE. In the initial evaluation among
the machine learning models, BVAR has the lowest MAPE and was used in the
hybrid model, ARDL-BVAR. Using the significant variables as factors, the model
improved by gathering much lower error metrics. The best-performing models
were then adapted into the website application called AniCast, which displays
visualizations, nowcasts, and evaluation metrics in table form and can be downloaded
into a PDF file. This study addresses the gap by utilizing machine learning
models and provides a valuable, accessible tool for informed decision-making regarding
rice production management in the Philippines.
