2.3 Methods
2.3.2 Separating impacts of urbanization from that of climate change on temporal
2.3.1.1 Characterizing effects of urbanization
The 23-year annual percentage of impervious surface area (%ISA) from ESA’s CCI land cover dataset enables us to derive the temporal trajectory of urbanization process at each 0.05˚ grid within each of the selected 196 large cities, including when and how much land conversions have occurred. The process of urbanization could be divided into three distinct periods including a pre-urbanization period when urban fractions remain low, an urban growth period during which urban fractions increase rapidly, and a post-urbanization period when the urban fractions
stabilize (Song et al. 2016). The temporal trajectory of %ISA can thus be modeled using a logistic function (Song et al. 2016) (equation 2.2):
∗ (2.2)
where describes the amplitude of urbanization, shows the rate of change, controls the phase of the change, indicates the year, and represents the %ISA value of pre-urbanization period (Figure 2.3). This approach is based on a common assumption that urbanization-related land cover change is irreversible at decadal scales (Gao et al. 2012; Mertes et al. 2015; Seto et al. 2011). It should be noted that this approach is only applicable to pixels with only one urban growth period to reach its stable urban status. The property of the logistic function, i.e. only one maximum value in its first derivative, can help us select the pixels that were suitable for the curve fitting. To identify these pixels, we calculated the slopes of each five-year moving window over the first derivative of the %ISA time series. We then identified the number of maximum values (i.e. peaks) from the slopes. Because random noises and uncertainties in the ESA’s CCI land cover data can introduce small local maxima in the slope time series, we further screened
out local maxima less than 10% of the global maximum to remove variations irrelevant to urbanization. The remaining number of maximum values was counted and only the pixels that have one peak (i.e. one urban growth period) were selected. Note that for urban areas already reached stable state, the first derivative of their %ISA time series will be likely to remain constant and have zero peaks. We then applied the logistic function to the selected pixels in the study area using an iterative Levenberg-Marquardt non-linear least-square regression (Levenberg 1944). The amplitude of urbanization was defined by the parameter . The two local maximum values of the changing rate of the curvature correspond, respectively, to the starting and ending years of urbanization (Figure 2.3). Similar definitions are widely used to derive transition dates of spring phenology (Zhang et al. 2003), i.e. onsets of green-up increase and maturity. The pre- urbanization period was defined as the time interval between year 1992 and the starting year of urbanization. The post-urbanization period was defined from the ending year of urbanization to year 2014. A threshold of 8% for amplitude of urbanization was used to excluded pixels that did not exhibit significant %ISA changes over the 23 years. This threshold assumed that the urban areas within each 0.05˚ grid experienced an average annual increase of approximately one pixel in the ESA CCI’s land cover product. Therefore, all the stable urban areas (i.e. amplitudes of their urbanization were below the threshold) are not included the analysis.
We verified our derived amplitude of urbanization using Landsat-derived urban land maps developed by Liu et al. (2018c). Those urban land maps were generated using an extensive collection of Landsat imagery (Liu et al. 2018c). The distinct advantage of this dataset (hereafter referred to as Landsat-based urban maps) is that it provides urban land maps back to 1990s at 30- meter spatial resolution. We first aggregated the 30-meter binary Landsat-based urban maps (i.e. 1 means urban and 0 means non-urban) into percentage of %ISA within the 0.05˚ urban grids (i.e. pixels that experienced one urban growth period) following the same way we
calculated %ISA based on the ESA’s CCI land cover dataset. Please note that although Landsat- based urban maps covered from year 1990 to year 2015 with a 5-year interval, there were large missing data in year 1990 (Liu et al. 2018c). Therefore, we excluded the Landsat-based urban maps in year 1990 and selected 0.05˚ grids with the starting year of urbanization no earlier than 1995 to calculate the Landsat-based %ISA. Given the 0.05˚ urban grids used in this study only experienced one urban growth period and have already reached stable period by year 2015, we
Figure 2. 3 Three typical examples of the temporal trajectory of the annual percentage of
impervious surface area (%ISA). Dotted line is the raw time series of %ISA. Solid line represents the fitted time series. (a) an example of urban growth from non-urban to low-density urban
(%ISA between 0.2 and 0.49) at 47.925˚N, 37.825˚E (Donetsk, Ukraine); (b) an example of urban growth from non-urban to moderate-density urban (%ISA between 0.49 and 0.8) at 40.625˚N, 109.975˚E (Baotou, China); (c) an example of urban growth from low-density urban to moderate- density urban at 45.725˚N, 126.525˚E (Harbin, China).
calculated the Landsat-based amplitude of urbanization as the differences of %ISA between year 1995 and year 2015. We used stratified random sampling to select 100 cities from the 8 climate zones (i.e. random selection based on the percentage of city number within each climate zone) and calculate the city-level zonal mean of amplitude of urbanization using Landsat-based %ISA. We then compared them against city-level zonal mean of our ESA’s CCI land cover derived results. In addition, we also randomly selected 0.05˚ grids and compared the temporal
consistency of %ISA time series derived from ESA’s CCI Land cover and Landsat-based urban maps to evaluate the quality and robustness of using the two datasets in understanding
urbanization dynamics.