A bi-directional strategy to detect land use function change using time-series Landsat imagery on Google Earth Engine: A case study of Huangshui River Basin in China

Constructed land, cropland, and ecological land are undergoing intense competition in many rapidly-developing regions. One of the major reasons to cause frequent land use (LU) conversions is the policy dynamics. The detection of such conversions is thus a prerequisite to understanding urban dynamics and how policies shape landscapes. This paper presents a bi-directional strategy to detect the LU change of the Huangshui River Basin of China from 1987 to 2018 using time-series Landsat imagery. We first initialized classification and optimization of remote sensing images using the Random Forest algorithm; We then detected bi-directional spatio-temporal changes based on the distribution probability of land-cover types. Our results reveal complicated dynamics underlying the net increase in urban and built-up land (UB) and the net decrease in cropland. In this area, due to the implementation of ecological compensation projects such as ecological migration and mine restoration, we found that on average 5.52 km² of UB was converted into ecological land (forest, grassland and shrubland) every year, even though UB has expanded 3.6 times in the last 30 years with multiple conversions for cropland and ecological land. Meanwhile, 60% of lost cropland was converted to shrubland and grassland, and 40% was converted to UB. The accuracy of LU classification increases by 6.03% from 88.17%, and kappa coefficient increases by 2.41% from 85.16, compared to the existing initial results and uni-directional detection method. This study highlights the importance of the use of an effective remote sensing-based strategy for monitoring high-frequency LU changes in watershed areas with complicated human-nature interactions.