Detecting Vegetation Phenology in Various Forest Types Using Long-Term MODIS Vegetation Indices

摘要

Vegetation phenology is the timing of seasonal events, such as the onset and offset of green-up, that can be used to monitor the response of climate variations on short- and long-term periods. In particular, accurate detection of seasonal phenological events is an important variable in ecosystem simulation models and general circulation models based on the regional and global climate conditions. However, phenological filed observation is collected in limited areas and time periods. An alternative approach has been developed using satellite remote sensing data, it allows for the most efficient spatial-temporal observation. A remote sensing-based vegetation index (VI) has been used to monitor phenological and seasonal changes in vegetation development. The normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) are most widely used. Forest ecosystems account for 63.7% of land area in South Korea, covered by deciduous, coniferous, and mixed forest types. Seasonal change in coniferous forest is not easily observed compared to deciduous trees, the start of the growing season is different. Homogeneity of the target area is essential to accurate phenology based on the remote sensing vegetation index. However, the pixel size between the remote sensing pixel and filed survey point is often unmatched in many regions. An adapted method required to extract phenology from one pixel including various ecosystem types. Our objectives were: 1) Test three different methods for extracting phenology events, particularly at the start of growing season (SOS); 2) determine whether NDVI or EVI represent the phenology of forest ecosystems; 3) match a resized digital forest type map to remote sensing pixel size; and 4) find the represented pixel among the pixels of the field observation location and eight neighbors. Three extract methods were applied to the data fitted by a double logistic curve. In order to establish a method to extract SOS that can be generally derivative function rather than “trs” with EVI. Field SOS data were collected from each site, however, it cannot be represented the pixel to which observation points belong. Therefore we extracted nine pixels, including the exact observation point pixel (center pixel). The center pixels from deciduous trees sites were 45.2% of deciduous forest, 22.3% of coniferous forest, and 7.2% of mixed forest, with 67% of deciduous forest, 14.3% of coniferous forest, and 7.7% of mixed forest for the best pixels on average, and the mean coefficient of determination (R2) was 0.64. The best pixels of the field observed from the coniferous tree species were with 56.2% of deciduous forest, 25.4% of coniferous forest, and 4.9% of mixed forest (averaged R^{2=0.29). The study demonstrates that use of long-term satellite vegetation indices for detecting phenology are related to the specific forest types in terms of investigative matching between high resolution digital maps and low spatial resolution remote sensing images.}