Exploring Built-Up Indices and Machine Learning Regressions for Multi-Temporal Building Density Monitoring Based on Landsat Series.

Uncontrolled built-up area expansion and building densification could bring some detrimental problems in social and economic aspects such as social inequality, urban heat islands, and disturbance in urban environments. This study monitored multi-decadal building density (1991-2019) in the Yogyakarta urban area, Indonesia consisting of two stages, i.e., built-up area classification and building density estimation, therefore, both built-up expansion and the densification were quantified. Multi sensors of the Landsat series including Landsat 5, 7, and 8 were utilized with some prior corrections to harmonize the reflectance values. A support vector machine (SVM) classifier was used to distinguish between built-up and non built-up areas. Regression algorithms, i.e., linear regression (LR), support vector regression (SVR), and random forest regression (RFR) were explored to obtain the best model to estimate building density using the inputs of built-up indices: Urban Index (UI), Normalized Difference Built-up Index (NDBI), Index-based Built-up Index (IBI), and NIR-based built-up index based on the red (VrNIR-BI) and green band (VgNIR-BI). The best models were revealed by SVR with the inputs of UI-NDBI-IBI and LR with a single predictor of UI, for Landsat 8 (2013-2019) and Landsat 5/7 (1991-2009), respectively, using separate training samples. We found that machine learning regressions (SVM and RF) could perform best when the sample size is abundant, whereas LR could predict better for a limited sample size if a linear positive relationship was identified between the predictor(s) and building density. We conclude that expansion in the study area occurred first, followed by rapid building development in the subsequent years leading to an increase in building density.

Tropical peat subsidence rates are related to decadal LULC changes: Insights from InSAR analysis.

Peatlands in Indonesia are subject to subsidence in recent years, resulting in significant soil organic carbon loss. Their degradation is responsible for several environmental issues; however, understanding the causes of peatland subsidence is of prime concern for implementing mitigation measures. Here, we employed time-series Small BAseline Subset (SBAS) Interferometric Synthetic Aperture Radar (InSAR) using ALOS PALSAR-2 images to assess the relationship between subsidence rates and land use/land cover (LULC) change (including drainage periods) derived from decadal Landsat data (1972-2019). Overall, the study area subsided with a mean rate of -2.646 ± 1.839 cm/year in 2018-2019. The subsidence rates slowed over time, with significant subsidence decreases in peatlands after being drained for 9 years. We found that the long-time persistence of vegetated areas leads to subsidence deceleration. The relatively lower subsidence rates are in areas that changed to rubber/mixed plantations. Further, the potential of subsidence prediction was assessed using Random Forest (RF) regression based on LULC change, distance from peat edge, and elevation. With an R2 of 0.532 (RMSE = 0.594 cm/year), this machine learning method potentially enlarges the spatial coverage of InSAR method for the higher frequency SAR data (such as Sentinel-1) that mainly have limited coverage due to decorrelation in vegetated areas. According to feature importance in the RF model, the contribution of LULC change (including drainage period) to the subsidence model is comparable with distance from peat edge and elevation. Other uncertainties are from unexplained factors related to drainage and peat condition, which need to be accounted for as well. This work shows the significance of decadal LULC change analysis to supplement InSAR measurement in tropical peatland subsidence monitoring programs.