Improving Urban Forest Expansion Detection with LandTrendr and Machine Learning

Abstract

Annual urban forest expansion dynamics are crucial for assessing the benefits and potential issues associated with vegetation accumulation over time. LandTrendr (Landsat-Based Detection of Trends in Disturbance and Recovery) can efficiently detect the dynamics of interannual land cover change, but it has difficulty distinguishing urban forest expansion from urban surface rapid conversions, as changes are usually filtered by magnitude-of-change thresholds. To accurately detect annual urban forest expansion dynamics, we developed an improved method using random forest-supervised classification to filter urban forests. We further enhanced the performance of the improved method by incorporating trend features between segments. Additionally, we tested two threshold-based filtering baseline methods. These methods were tested with various spectral and parameter combinations in Beijing’s Central District and the 1st Greenbelt from 1994 to 2022. The improved method with trend features achieved the highest average accuracy of 89.35%, representing a 25% improvement over baseline methods. Post-change trend features aided in accurate identification, while quantitative features rather than extremum features were more important in filtering. The improved method with trend features tested in Beijing’s 2nd Greenbelt also showed an accuracy of 88.27%, confirming its stability. SWIR2 and a higher maximum segment number are efficient for filtering by providing the most detailed dynamics. Accurate annual expansion dynamic mapping offers insights into change rates and precise expansion years, providing a new perspective for urban forest research and management.