Continuous subpixel monitoring of urban impervious surface using Landsat time series

A novel method called Continuous Subpixel Monitoring (CSM) was developed to map and monitor urban impervious surface change continuously at the subpixel level. Time series model of each pixel was first estimated based on clear Landsat observations between 2000 and 2014, and any land surface change was detected by the Continuous Change Detection and Classification (CCDC) algorithm. These coefficients and Root Mean Square Errors (RMSEs) of the estimated time series models were then employed as the inputs of random forest regressor. A few experiments with different combinations of variables and bands were explored to better construct random forest regression models. We successfully applied this algorithm to map subpixel urban impervious surface area (ISA%) and characterize its dynamics in Broome County, New York. Several conclusions can be drawn from the results and analyses. First, the integration of subpixel mapping technique and time series analysis in CSM can yield a relatively satisfactory ISA% result at one point in time. With higher precision and smaller bias, its mapping accuracy is better than that of National Land Cover Database (NLCD) percent developed imperviousness product, without using extensive auxiliary data, such as nighttime light image and transportation network. Second, the ISA% change of any time interval can be easily derived and detected by CSM with relatively high accuracy, which have the potential to generate sub-annual ISA% change products. Furthermore, this approach can detect not only urban expansion/intensification (ISA% gain), but also different patterns of urban transitions overtime, such as urban demolition/redevelopment to vegetation (ISA% loss), and surface modifications (no mechanical change). Finally, CSM works well in one of the cloudiest regions in the United States. This algorithm could provide a new direction to map and monitor percent urban impervious surface change in a reliable and efficient way, which also has the potential to apply to other land cover types (e.g., tree, shrub, and barren lands) at the subpixel level.