Hybrid Time-Series Forecasting and Anomaly Detection for Early Deforestation Monitoring in Indonesia

Abstract

Deforestation remains a major environmental issue in Indonesia, affecting biodiversity and increasing carbon emissions. This study analyzes forest dynamics by combining time-series forecasting and anomaly detection approaches. Annual data from 2000 to 2025 were examined using ARIMA, Prophet, exponential smoothing, as well as linear and polynomial regression models. Performance was evaluated using MAE and RMSE, where Prophet achieved the lowest error values, indicating a stronger ability to capture non-linear trends in forest area changes. To provide deeper insight, anomaly detection using Isolation Forest identified several years with unusual patterns, particularly 2014 and 2015. In addition, change point detection with the PELT method revealed structural shifts in 2004 and 2014, highlighting 2014 as a key transition period in forest dynamics. The findings indicate that forest area changes tend to follow a gradual pattern, but certain periods show notable irregularities. The integration of forecasting, anomaly detection, and change point analysis enables a more comprehensive understanding of long-term trends as well as critical changes, supporting the development of data-driven deforestation monitoring systems.