Component 6 loadings graph showing a steadily increasing trend for which forested areas are negatively associated. 

Context in source publication

Context 1

... (they all measure NDVI). As a consequence, the components produced reduce redundancy by describing patterns in NDVI that recur over time -- hence the value of this procedure as a time series analysis technique. The key to using PCA as a time series analysis technique is to examine both the temporal and the spatial characteristics of each component. For example, Figure 2 illustrates the first two components of an analysis of the entire continent of Africa for the period 1986-90. The temporal characteristics of these components are described in a loading chart -- a graph of the correlation between the component image and each of the original input images. Thus the Y axis indicates the degree of correlation while the X axis indicates time, from the beginning to the end of the series. As can be seen in this chart, the first component correlates strongly and essentially identically with each of the 60 input images. Thus the first component represents the characteristic NDVI over the continent, regardless of the season and other variations that may be present. In contrast, the second component shows a strong sinusoidal pattern, correlating positively with the northern hemisphere summer months and negatively with the winter months. Thus we can see that the biggest deviation from the characteristic NDVI shown in the first component is that which arises from the apparent movement of the sun -- the winter/summer seasons. This logic carries through for all remaining components -- i.e., each is a residual (left over) pattern after the effects of all previous components have been removed. Thus, for example, the image of the eighth component in Figure 3 represents a residual pattern that lies seven levels deep in the data -- a concept that gives real meaning to the term data mining . Interestingly, this deep residual pattern can be shown to represent very clearly the effects of the El Niño / Southern Oscillation (ENSO) phenomenon that brings drought to Southern Africa. To substantiate this, the Southern Oscillation Index (SOI), an index to the nature of this phenomenon derived from atmospheric pressure readings in the Pacific Ocean, is plotted as a second element on the loading chart. The relationship is clear, as is the location of the drought in Southern Africa that took place during the El Niño of 1986/87 and the wetter conditions that followed during the La Niña of 1988/89. An important feature of this time series analysis procedure is that the results are very much scale dependent, both in space and time. At a continental scale, local effects such as changes in land cover lack the coherence over space and time that would cause them to come out in the early component patterns. In essence, they are noise to the main signal of events such as the seasons and large-scale continental influences such as El Niño. However, if the same analysis is conducted for a more limited region, it is logical to expect that more localized phenomena may have the weight they require to come out in the early components. Thus when exploring the potential applications of the NDVI imagery with our Malawian counterparts, we examined analyses both at the regional and national scales. Regionally, the procedure has clear importance for understanding the dynamics of the ENSO phenomenon and its implications for food security. Malawi is a predominantly agrarian society that is strongly affected by the El Niño drought phenomenon. However, when we came to examine the potential of these data at a national level, we had no clear sense of what to expect, other than that we should see more localized effects such as land cover change. What became immediately clear is that the procedure holds enormous potential for the monitoring of forest cover change. For this more local exploration, a subset of the 7.6 km NDVI archive was extracted for the seven year period from October 1987 to September 1994 (the agricultural year runs from October to September). In addition, the subset was restricted to Malawi only, with pixels outside the national boundary being assigned a constant value of 0 (thus effectively removing these cells from consideration). A Standardized Principal Components was then run on the 84 subsetted images using the IDRISI GIS and Image Processing software system. As with the continental scale analysis, the first component of the Malawi-specific analysis illustrated characteristic NDVI, with the next four components representing seasonal changes and anomalous effects such as recurrent cloud effects and inconsistencies in the calibration of the AVHRR sensor system. The first component with a non-artifactual interannual element was Component 6 (Figure 4). Unlike the continental scale analysis, where the first inter-annual component relates to El Niño, Component 6 in the Malawian study is not related to ENSO. Rather, it highlights a number of isolated but very strong negative anomalies that are associated with a steadily increasing trend over time (Figure 5). Comparison of these locations to map data showed that the negative anomalies were predominantly associated with forest reserves, game reserves and national parks. These boundaries are delineated in Figure 4. The data suggested then that there was a progressive decrease in NDVI in the reserves. To confirm this, two further analyses were undertaken. First, temporal profiles were constructed for five of the major reserves, measuring the average NDVI within the reserve over the 84 months in the series. These are plotted in Figure 6 along with the mean trend. The second analysis consisted of a more detailed examination of the Phirilongwe forest reserve -- one of the smaller anomalies to be highlighted by the analysis. Figure 7 shows a Landsat MSS false color composite image for 1981 and a Landsat TM image for 1991 resampled to the same resolution. In false color multi-spectral images such as this, vegetation appears red. The change between the dates is strongly evident. Malawi is a country in which significant deforestation has taken place over the past 30 years. However, the Department of Forestry is significantly hindered by a lack of funds to do forest inventorying and monitoring throughout the country. Current methods entail detailed inventorying of forested areas using a variety of manual techniques including on-site surveying, and when further funds permit, aerial surveying. But these techniques have not been very effective for routine monitoring. Aerial surveying is done on an ad hoc basis, but on average every ten years. These methods are significantly more expensive and not very efficient when constant monitoring is essential. Furthermore, since they rely on the visual delineation of forest stands, they are unable to detect changes in the density of forest cover. Although the time series analysis conducted here was based on very coarse imagery, the fine temporal detail has allowed it to pick up a very subtle effect: loss of forest cover in the reserves not by clear cut, but rather, by progressive thinning due to poaching by local residents. Over 90% of Malawi's energy needs are met by wood and the illegal sale of public wood products has become an important cash crop in most rural areas. Our colleagues in the Forestry Department were aware that such poaching was taking place, but were surprised at the extent. In fact, using the temporal profiles for the five reserves in Figure 6, the estimated rate of decline is between 2-3 percent per year -- a rate that they have indicated is consistent with the losses experienced by clear cutting over the previous 30 years. The value of this analysis, however, does not stop with the simple awareness that this level of poaching is going on. Clearly there are differences from one reserve to another. Why do some reserves experience lower rates of loss (some in fact would appear to be stable)? Is this the result of differences in management and enforcement? The analysis can very effectively target more detailed investigations, as well as provide the basis for further monitoring to gauge the effect of different intervention strategies. As a result of this exploration, we have been able to develop a simple and cost-effective procedure for the monitoring of a very important resource in Malawi -- the national forest reserves. The imagery is free to the Government of Malawi and requires only a simple microcomputer and suitable software for the analysis to be undertaken. Similar explorations are being undertaken with other agencies using a variety of inexpensive image data, including METEOSAT Cold Cloud Duration (CCD) data for precipitation mapping, and 1.1 km multispectral AVHRR imagery for land cover monitoring. However, a more general finding from this work is that the information that data can yield is not solely a function of the data alone, but also of the analytical procedure employed. It also suggests that we need to be more liberal in our thinking about the concept of resolution. The 7.6 km NDVI data, such as those used in this study, are available free of charge to anyone world wide ( see the sidebar ). They are very coarse in the spatial domain. However, they are extraordinarily fine in the temporal domain. This study has shown that subtle, but important, effects, both global and local, can effectively be monitored by using this temporal information. As a consequence, current attempts to develop and maintain inexpensive time series archives should be encouraged. J. Ronald Eastman J. Ronald Eastman is Director of the Clark Labs for Cartographic Technology and Geographic Analysis and Professor of Geography at Clark University, USA. He is also a Senior Special Fellow of the United Nations Institute for Training and Research (UNITAR). However, he is perhaps best known as the author the IDRISI GIS and Image Processing software system. He can be reached at: IDRISI@VAX.CLARKU.EDU. James Toledano James Toledano is a Senior Research Associate at the Clark Labs ...