Evaluating Models for Urban LULC Classification by Integrating Earth Engine and Python Ecosystem

Urban climate change studies necessitate long-term monitoring to manage urban environments. The Bangkok Metropolitan Region (BMR) faces challenges like increased urban heat island effects, flooding, and air pollution. Long-term land use and land cover (LULC) change monitoring is crucial for analyzing urban development. Landsat series data is vital for tracing LULC changes over 40 years, but classification accuracy often suffers due to spectral overlap. This study introduces deep learning techniques to enhance classification accuracy. We compare three CNN architectures: ConvlD, Conv2D, and VGG16. ConvlD got the best accuracy of 94.86% with a Kappa (K) score of 0.890, showing its effectiveness in handling spectral information. Conv2D achieved an accuracy of 94.25% and a K score of 0.876 by leveraging spatial context. VGG16 showed slightly lower performance exhibiting an accuracy of 87.58% and a K score of 0.803. Our study evaluates these models for urban L ULC classification using Landsat data specific to BMR. By understanding each model's strengths and limitations, we aim to recommend the most suitable CNN architecture for accurate, scalable urban L ULC mapping, contributing to better urban environmental management.