Exploring Forest Cover Dynamics in Bangladesh: A Machine Learning and Google Earth Engine Approach

Abstract

Aims: This study aims to explore forest cover dynamics in Bangladesh by applying machine learning techniques to satellite-derived data, and to evaluate the performance of different algorithms in detecting and forecasting land cover changes. Study Design: A quantitative, remote sensing-based analytical study was conducted using multi-temporal satellite imagery and machine learning models to assess land cover transitions over time. Place and Duration of Study: The study was conducted across Bangladesh, covering the period from 2012 to 2023, using data processed through the Google Earth Engine platform. Methodology: Satellite-derived indices and classified land cover datasets were analyzed to detect changes in forest and other land cover types using Landsat 8 (OLI) (30m), and Landsat 9 (OLI-2) (30m), MODIS MCD12Q1 (500m) sensors. Multiple machine learning algorithms, including Long Short-Term Memory (LSTM) to capture temporal dependencies; Random Forest, Decision Trees, XGBoost, and Support Vector Machine (SVM), were implemented and compared based on their classification accuracy. Time-series analysis was applied to evaluate temporal patterns and improve prediction performance. Results: Among the evaluated models, LSTM demonstrated the highest accuracy at 94%, followed by XGBoost (93%), Decision Trees (87%), and Random Forest (70%), while SVM showed the lowest performance at 67%. Land cover analysis revealed a substantial increase in waterbodies by 996.3 km² and built-up areas by 1,054.5 km², indicating changes driven by hydrological variation and rapid urbanization. In contrast, forest cover declined significantly by 1,999.8 km², along with a reduction in overall vegetation by 1,958 km², highlighting ongoing deforestation and ecosystem degradation. Conclusion: The findings highlight significant transformations in Bangladesh’s land cover, particularly the alarming loss of forest resources. The superior performance of LSTM underscores its effectiveness in capturing temporal dependencies for accurate forest cover change detection and forecasting. This study emphasizes the urgent need for sustainable land management strategies and provides valuable insights into the complex interactions between human activities and forest ecosystems.