A machine learning approach to monitoring and forecasting spatio-temporal dynamics of land cover in Cox's Bazar district, Bangladesh from 2001 to 2019

Abstract

This study uses machine learning techniques to monitor the areal extent of land use land cover (LULC) classes in Cox's Bazar district, a well-known tourist destination in Bangladesh from 2001 to 2019. The main objective of this study is to quantify the changes in LULC classes during this period and forecasting the areal extents classes for 2019 and 2025. Satellite imagery of 2001, 2005, and 2010 from Landsat 4-5 TM and, 2015, and 2019 from Landsat 8 OLI/TIRS were used to classify the study area into four LULC classes (water bodies, urban areas, vegetation, and bare land) using Random Forest (RF) classifier. These classes were used to forecast areal extents of land cover using Cellular Automata Simulation (CAS) for 2025. The eleven LULC classes from MODIS land cover type products (IGBP scheme) were also forecasted using RF regressor to compare prediction accuracy. This study observed an increase in vegetation cover and urban settlements by 26.65% and 2.52%, respectively, with decreased water bodies and bare lands by 17.25% and 11.91%. The overall accuracy of the classification is between 89.04% to 95.24%; CAS forecast accuracy is 93.2% and RF regressor forecast accuracy ranges from 87% to 99.8%. Grid search with cross-validations was performed to boost the accuracy of both models. The results are consistent with overall global trends of LULC change.