Annual Past-Present Land Cover Classification from Landsat using Deep Learning for Urban Agglomerations

Worameth Chinchuthakun, David Winderl, Alvin C.G. Varquez, Yukihiko Yamashita, Manabu Kanda

Overview

UrbanLC is a Python library for land cover classification (LCC) from Landsat Images. It features pretrained deep learning models, which are compatible with all Landsat sensors up-to-date: MSS, TM, and OLI-TIRS. The library further contains some utility functions for analyzing and visualizing land cover maps and tutorials for researchers and practitioners.

• Python library is available on GitHub repository

• Pretrained models are available on Hugging Face

Abstract

Historical land cover data is crucial for understanding urbanization dynamics, climate modeling, and monitoring water resources. Following recent advancements in deep learning for processing Landsat archive data, prior studies have released high-resolution historical land cover maps on a global scale. However, these works often present prediction results limited to specific periods of coverage, which hinders their utility in conducting time series analysis across different urban agglomerations. To address this issue, we propose deep-learning models for land cover classification from Landsat images at a 30-meter spatial resolution. Our models are specifically designed for urban areas and are trained to be compatible with the sensors used in the Landsat series from 1972 to the present. Experimental results demonstrate that our models are highly effective in predicting land cover maps in new cities, particularly in built-up land and water regions. Our research provides pretrained models for land cover classification, facilitating future studies in related fields.

Citation (Bibtex)