A simplified training data collection method for sequential remote sensing image classification

This study examined urban expansion of selected cities (Suzhou, Wuxi, and Changzhou) in Yangtze River Delta, China using time-series satellite remote sensing data from 1986 to 2014. A total of 16 cloud-free Landsat satellite images for the study area were downloaded and three-class (impervious cover, vegetated area, and water) land cover map products were developed using advanced remote sensing image classification techniques. Image classification training data points were selected from 1986 and 2014 images, focusing on common/no-changed areas within the 28-year time period. The locations of such training area were then used to extract training data points from each individual satellite image and fed into a random forest classifier to support land cover classification. Image classification accuracy assessment was conduct for years of 2010 and 2014. Overall classification accuracies were above 85% and accuracies for impervious surface class were above 75%. Our method generated more consistent training data and required less image interpretation time than traditional image-by-image approach. From 1986 to 2014, total impervious surface area increased from 375km
²
to 1073km
²
, corresponding to a very high annual urban growth rate of 7%. By examining all 16 land cover maps from 1986 to 2014, we observed increased level of landscape fragmentation. Furthermore, there was a linearly increasing trend of total impervious surface. A simple linear trend model predicted 1423.8 km
²
(∼ 66% of land area) of impervious surface area in 2030.