Young presented an analysis of vegetative cover change in China using global-scale Advanced Very High Resolution Radiometer (AVHRR) data and the IDRISI Geographic Information System (GIS). Both the data set and the GIS system were chosen for their general availability and accessibility to anyone who might wish to work with them. IDRISI is a simple, inexpensive, raster-based GIS which can be used on most PCs. AVHRR data is very available, inexpensive, easy to handle, has low computer memory requirements (years of data covering the whole globe can be analyzed on a lap top), and potentially contains a great deal of information.
The AVHRR data used in Young's analysis (1982-92) comes from sensors on three satellites, NOAA 7 (1982-85), NOAA 9 (1985-88), and NOAA 11 (1988-92). AVHRR provides data on five channels, only two of which are used in this work, the visible red and the near infrared. The AVHRR data scales are:
AVHRR's NDVI is based on the relationship between near infrared light and visible red light (ir-r)/(ir+r). Vegetation absorbs highly in red light and reflects highly in infrared light. AVHRR's NDVI values range from -1 to +1, where water, snow, and ice are generally less than 0, rocks and bare soil around 0, and vegetation ranges from 0.1 to 0.6. For vegetated areas, the denser the vegetation the higher the NDVI.
The AVHRR sensor collects global data daily, however, about half the Earth is covered with clouds at any given time. A method known as Maximum Value Composites (MVC) has been developed where 7 days of data are overlaid taking the maximum NDVI value. In this way clouds are removed from the images. In some cases, 7 days aren't enough, so 10-day or monthly maximum value composites are used. Another potential problem with the AVHRR data used in this study is that there are three different satellites carrying the AVHRR sensor. This creates differences in NDVI between the three satellites. In addition, each of the satellites slows down as it circles the Earth creating later equatorial crossings over time. Calibration of these GVI data sets would reduce these sensor-related problems. Additional sensor problems were acknowledged by Young, but the emphasis of his talk was not on data formation.
Two AVHRR data sets are used in this work, the NGDC (U.S. National Geophysical Data Center) Monthly Generalized GVI from NOAA 9 Weekly GVI data (1985-1988), and the NOAA Weekly GVI data via UNEP/GRID-Geneva (United Nations Environment Programme /Global Resource Information Database) from NOAA 7, 9, 11 (1982-92).
Data preparation was the most time-consuming element of producing the vegetation analysis. For the NGDC GVI data, NGDC checked registration, data were visually inspected for noise which when found was reduced, data were composited monthly in 10-minute grids, biased toward median values (well-processed, but not calibrated). For the NOAA Weekly GVI data, the first generation data (NOAA 7) were reprojected into Plate Caree to make it the same as data from NOAA 9 and 11, data were composed weekly in 10-minute grids, biased toward maximum values (data not fully processed, and uncalibrated). Young aggregated the data into monthly composites (13 per year, based on four weeks per month), reviewed data for noise, removed noise and reaggregated data (data clear of noise, but not calibrated).
The technique used in this study was Time Series Analysis (TSA), a form of Principle Components Analysis (PCA) where the only variation is time. It is understood that there are some changes caused by the need for atmospheric and sensor correction. PCA under takes a linear transformation of a set of image bands to create a new set of images that are uncorrelated and are ordered in terms of the amount of variance explained in the original data. TSA is area weighted, therefore the early components show change which has a great magnitude over a larger area, while later components either show low magnitude change over large areas or large magnitude change over small areas.
The products of TSA are known as components which are made up of two aspects:
1) Loadings (correlations of components showing temporal change). The loadings are presented in graphic form and show: i) annual vegetation patterns as reflected in annual patterns of rising and falling NDVI; ii) long-term trends as indicated by an overall increase or decrease of NDVI over long periods of time; and iii) extreme events which are seen as outliers from the general patterns.
2) Spatial Image (images show spatial change in the time series). The image shows the spatial distribution of vegetation change as described in the loading graphs.
To analyze broad vegetational changes in China, Young first attempted to determine the typical annual changes that occur. Five years of data for each month are averaged to get a typical year. As noted above, the components produced range from large over all changes to smaller changes. The first component produces a loading which shows no distinctive change in time and every month is highly correlated with the spatial image. The spatial image shows the NDVI for China as it would be regardless of season. That is, the tropical regions of southern China have a high NDVI while the deserts of Xinjiang have a low NDVI. This indicates that the major change in vegetation for China is not temporal, but rather spatial. That is, there is a greater difference in NDVI between tropical China and the deserts of China than the difference in NDVI between summer and winter for those regions. Component 1 turns out to be an integration of NDVI over the year. Some non-vegetation effects such as cloud and fog contamination, aerosols reducing light, and irrigation (water gives low NDVI) show up in this component. Component 2 shows the second-greatest variation, that between summer and winter. Here the deciduous forests of the Northeast show up very clearly. Component 3 shows the third-greatest variation, that between spring and fall which turns out to reflect various cropping patterns. Component 4 also shows some major cropping patterns in Eastern China. Later components were not as easily distinguished as the early components and represent more local changes. This methodology captures different annual vegetation patterns in China.
Young next looked at interannual variations in three time series (1986-88; 1982-87; and 1987-92). Before showing results in China he showed results for Africa using the same data base for the years 1986-88. The African studies have shown that TSA can lump sensor degradation into specific components so this degradation doesn't effect the results shown in the other components. Work with calibrated data for Africa has confirmed this conclusion. The African studies showed that TSA can show annual vegetation change, sensor degradation effects, and the effects of El Niño-Southern Oscillation (ENSO) events on African vegetation.
Looking at the long-term TSAs for China, the first four components were very similar to those found in the "typical year" analysis. The major difference was that for the long-term data sets, the loadings showed variations in NDVI between years and the images showed the specific areas where NDVI was changing on an interannual basis. Some of the later components showed some local high magnitude changes which occurred in specific years. One clear example was the great forest fire in Northeast China in 1987.
Component 6 on the two long-term TSAs (1982-87; 1987-92) show the problems associated with the sensor degradation and the change of satellites. The NOAA satellites carrying the AVHRR sensors slow down as they circle the earth resulting in later equatorial crossings as time passes. The effect of this later and later passing is that the light reflected from the earth has to travel through more atmosphere as time passes. As the light travels through more atmosphere, the visible red light is attenuated more quickly than the infrared light. As time passes, this makes NDVI increase, giving the appearance that vegetation is increasing. This is particularly evident in low vegetation areas such as deserts. An interesting aspect of the image for component 6 is that certain areas of forest (not all forest areas) are shown as being highly uncorrelated with the general trend, which would indicate that they are loosing NDVI, perhaps because of deforestation.
Young then ran a TSA using component 1 (or the annually integrated NDVI) for each year from 1982 to 1992. As with all of the other TSAs the first component showed the integrated NDVI over the period. The second component showed the change of sensors and the degradation of the sensors. The third component possibly showed the effects from various ENSO events during the 11-year period. The most prominent effect of this analysis was that in component 2 many forest areas showed up decreasing in NDVI. This is not surprising as many reports have indicated widespread deforestation throughout China's forests in the 1980s. The areas of deforestation were confirmed by profiling the specific areas and following the abrupt changes in NDVI which indicates deforestation, as opposed to the profiles of desert regions which show a very smooth linear progression of in creasing NDVI caused by the sensor degradation. Of interest here is that the sensor degradation problem may in fact be helping highlight the areas of deforestation which might otherwise be difficult to see. The great forest fire of 1987 in northern Heilongjiang also helped to verify the forest change as seen in component 2.
Young concluded that future work will use calibrated data with TSA to see if this method can effectively isolate sensor problems in the data base, and to confirm the areas of deforestation. Future work will also be done with the Institute of Remote Sensing Application, Chinese Academy of Sciences, to verify the regions of deforestation with Landsat data. Future cooperative work with the Chinese Academy of Sciences will also look at the ability of AVHRR data to isolate climate effects (such as those of ENSO events).