2 METHOD AND MATERIALS

2.1 Study area

Zhejiang Province, one of the most socioeconomically developed provinces in China, is chosen to be our study area for the following reasons. Located at the southeastern coast of China, Zhejiang is located at the southern fringe of the Yangtze River Delta (Figure 1). Lying between 27°12´-31°30´N and 119°42´-122°06´E, it has a subtropical monsoon climate with abundant rainfall (1,100 to 2,000 mm on average annually). With a land area of 10,550,600 ha, Zhejiang is a mountainous province characterized mostly by mountains (~70%), agricultural land (~20%), and water body (~10%), with flat areas mainly in the northeast and mountainous areas in the southwest (Figure 1). As of the end of 2019, the total population was 58.5 million, of which 70% lived in cities, where the forest coverage was 61.15% of the total land area (citation from Zhejiang Statistical Yr book). Zhejiang province, with a GDP of 954 billion USD and GDP per capita of 16,474 USD in 2019, ranks the fourth and the fifth in China’s 31 provinces, respectively. The geographical diversity and rapid economic development have led to fundamental changes in land cover and land use. Recently, deforestation associated with urbanization has become a serious problem, which renders Zhejiang Province an ideal case study area for conducting research on policy related deforestation.
Figure 1. Location and Elevation of Zhejiang Province, China

2.2 Data sources and processing

We utilized primarily GlobeLand30 to detect and analyze land use changes in Zhejiang Province from 2000 to 2020. GlobeLand30 is the first open-source, fine-scale global land cover database based on remote sensing models (J. Chen et al., 2015). This dataset is the only product worldwide on land cover with a 30 meters resolution for the years of 2000, 2010 and 2020. The GlobeLand30 dataset is derived from over 20,000 Landsat and HJ-1 (China Environment and Disaster Reduction Satellite) imageries using machine-learning models in combination with pixel-level and object-based processing procedures. In particular, the dataset of 2020 also incorporates the 16-meter resolution GF-1 (China High Resolution Satellite) multispectral images. The principle of image selection in the dataset is to select multispectral images of the vegetation growing season within ± 2 years of the baseline year in which the data were generated and updated, provided that the images are cloud-free or with least cloud. For areas that are difficult to acquire data, the timing of image acquisition can be adjusted to ensure the integrity of the overall coverage. The classification scheme includes 10 land cover types, which are agricultural land, forest land, grassland, shrub land, wetland, water body, tundra, artificial surface, bare land, and perennial snow and ice, with no mosaic pixels (J. Chen et al., 2015).
Based on a third-party evaluation, the overall accuracy of classification based on GlobeLand30 (for 2010) is 83.50% and the Kappa coefficient is 0.78. This result is from validation effort based on over 150,000 points in 80 tiles of 853 in total. On the other hand, the overall accuracy of classification based on GlobeLand30 (for 2020) is 85.72% and the Kappa coefficient is 0.82 according to our validation results based on over 230,000 points from the whole datasets using landscape index sampling model (Jun, Ban, & Li, 2014; Liang et al., 2015).
To better analyze the results of land use change in Zhejiang Province, we also used data from Zhejiang Land Statistics Yearbook. Specifically, we analyzed official statistics on land use change in Zhejiang Province since 2000. To more comprehensively capture forest loss in Zhejiang Province, we also selected the European Space Agency (ESA) - Climate Change Initiative (CCI) as another data source. With a medium-resolution (300m) resolution, the dataset has a global coverage from 2000 to 2018, which classified pixels (using a machine-learning algorithm) into over 22 land cover categories (for instance, mosaic natural vegetation of tree, shrub, herbaceous vegetation) (Bontemps et al., 2013). The accuracy of the map is reported to be 71.5% (Defourny et al., 2017). In our study, we reclassified the products according to the IPCC classification criteria and extracted the forest class for further analysis; here for alignment with our land classification typology, all ESA-CCI land cover types with trees and mosaic trees and shrubs were reclassified as forestland.
To further assess the annual change regarding forest loss, we used the latest version (Version 1.7 Update) of Hansen’s forest cover dataset, which is available online on the Google Earth Engine (GEE) website and the Global Forest Watch website. The most updated dataset contains the layers of 2000 tree canopy cover and 2001-2019 forest loss, providing the information regarding the year of forest loss. The product has a 30m spatial resolution and is synthesized by processing 654,178 Landsat 7 ETM+ images in high quality (Hansen et al., 2013). The dataset defines trees as “all vegetation above 5m in height” and forest loss as “the mortality or removal of all tree covers in a 30m by 30m pixel” (Hansen et al., 2013). Hansen’s previous update of forest gain was in 2012 and thus may be biased from forest growth in reality. The overall accuracy, assessed by the Food and Agriculture Organization (FAO) statistics using both LiDAR surveys and other satellite data, has been shown to be over 99% (Hansen et al., 2013). In order to minimize any data error and improve classification accuracy, we have combined remote sensing datasets with field surveys. The validation results turned to be fairly good: we have achieved an overall accuracy of 80% based on Hanen’s data in Zhejiang Province (Xiong et al., 2020). In addition, we believe that the next generation of Hansen products (e.g., Version 2.0) may provide more information on actual forest growth and loss (Zeng et al., 2018). In this assessment, by referring to the Global Forest Watch website, we set 30% as the threshold of defining the canopy cover for all the following analyses of forest loss.
The administrative division data of Zhejiang provincial is derived from Global Administrative Areas (GADM). These datasets have their own strengths in showing the spatial and temporal patterns of land change in Zhejiang Province (Table 1), and the combined use of the results of these data analyses is beneficial in exploring the spatial and temporal characteristics of land use transformation.
Table 1. Different land data sources