Prediction of Forest Fire Danger by Using BiGRU Deep Learning Model and Comparable Data Scaling Methods: A Case of SAO Hill Forest Plantation, Tanzania

Forest plantations are crucial in the daily lives of humans, playing an important role in producing raw materials for the wood industry, generating personal incomes, contributing to economies, attracting tourists, conserving biodiversity, and regulating the climate. Failure to accurately and timely predict forest fires can have devastating effects due to the destruction of forests by fire, resulting in loss of businesses and incomes, destruction of biodiversity, loss of tourist attractions, and shortage of wood raw materials. Fire Weather Index (FWI) is commonly used to indicate fire danger as it gives useful information on the impact of wind and fuel moisture on the behaviour and spread of fire. This study utilizes FWI by developing a Bidirectional Gated Recurrent Unit (BiGRU) Deep Learning model, which uses the previous 5 days FWI values as input to predict the output (next day FWI) at SAO Hill Forest Plantation located in Iringa region, Tanzania, using three commonly used data scaling methods: Min-Max, Standard, and Robust scalers. The 13-year SAO Hill Forest Plantation daily FWI dataset was pre-processed using a scaling (normalization) approach and split into training, validation and test sets before being used for training and testing the developed BiGRU Deep Learning model. The trained BiGRU Deep Learning model was then saved into .h5 format and integrated with a Gradio-based Web App to provide a user interface for officials at SAO Hill Forest Plantation to predict daily FWI. The evaluation findings reveal that the choice of data scaler has an impact on the daily FWI prediction performance of the developed BiGRU model, and Min-Max is the best performing and optimal data scaler with a Root Mean Squared Error (RMSE) score of 0.065 on test data, followed by Standard scaler with a test RMSE score of 0.157, followed by Robust scaler with a test RMSE score of 0.311. Major contributions of this study include a pre-processed 13-year FWI dataset for SAO Hill Forest Plantation ready for Artificial Intelligence (AI) research and development, a novel BiGRU model for predicting daily FWI at SAO Hill Forest Plantation, and a Web App integrated with the developed BiGRU model and Min-Max data scaler to help officials at SAO Hill Forest Plantation predict daily fire danger and take precautionary measures to prevent forest fire ignition, respond to forest fire if it happens, and contain its spread