Introduction

Continental and global land cover classifications from satellite data have largely been derived from annual time series of the Normalized Difference Vegetation Index (NDVI) as a measure of phenological variability throughout the year. This approach raises the question of whether there are remotely sensed variables in addition to NDVI that improve the accuracies of land cover clasifications.
Studies to address this issue conclude that multitemporal metrics derived from reflectances in individual AVHRR bands in addition to NDVI improve separability between cover types at a global scale. Other studies demonstrate that NDVI is sensitive to short-term climatic characteristics and that classifications based on a time series of the surface temperature to NDVI ratio produce more stable results over an eight year period for the African continent. Still other researchers suggest the use of metrics that describe an annual NDVI temporal profile, such as length of growing season and rate of greenup rather than the NDVI values directly and a rule-based approach for determining cover type.
We test a number of metrics derived from the six bands of the PAL data using MSS training data to determine the metrics to be used as input variables to the classifier. Based on these results, we use the metrics in a decision tree classifier to obtain a global land cover classification product.

Classification Algorithm

Decision tree theory has been applied previously to remotely sensed data. Decision trees predict class membership by recursively partitioning a data set into more homogeneous subsets. Different variables and splits are then used to split the subsets into further subsets. The grown tree can be selectively pruned based on decision rules to produce more stable predictions of class membership. For example, lower level splits which contribute little to the clasification accuracy can be merged into parent nodes.
The classification tree has a number of advantages over traditional classification methods. First, it is not based on any assumptions of normality within training area statistics as is a maximum likelihood approach. Thus it is better suited to those situations where a single cover type is represented by more than one set of remote sensing characteristics. Second, the tree can reveal non-linear and hierarchical relationships between the input variables and use these to predict class membership. Third, from a practical point of view, it is immediately apparent which variables contribute to the discrimination between classes. This can be useful for defining subsequent suitable inputs for landcover characterizations.

Methods and Results

The next issue concerns the input variables to be used in the decision tree classifier. Several types of metrics have been suggested to describe vegetation phenology, and are described in DeFries et al. (at press).
To determine which suite of metrics to use for deriving global land cover classification products, we tested five sets of metrics with the decision tree classifier using MSS training data. These metrics were derived such that they are independent of the timing of phenological events and thereby applicable at a global scale. For example, a metric characterizing the month when green up begins would not be globally applicable because of phasing of seasons in different parts of the world, nor would any individual monthly NDVI value. The sets of metrics are given in Table 3. The basis for each set is as follows:
  1. "surface temperature/NDVI metrics" are to characterize the temporal profile of the surface temperature/NDVI ratio. Surface temperature is calculated from channels 4 and 5.
  2. "seasonal NDVI-derived metrics" are calculated from 10-day NDVI values.
  3. "simple metrics" are calculated from annual maximum, minimum, mean, and amplitude (difference between maximum and minimum) for AVHRR bands 1,2,3,4, and 5 and NDVI.
  4. use of simple metrics in a "logic tree" where we derived a series of hierarchical trees to subdivide the cover types. For example, the first tree divides all the training pixels into woody and non-woody vegetation by aggregating the labeled cover types into these two major categories. Subsequently, we then grow a tree for all those pixels designated as "woody" to further subdivide them into "needleleaf" or "broadleaf." This procedure is carried out so that there is a series of trees for each cover type. Input variables to the trees were the 24 "simple metrics."
  5. combined use of all metrics in 1), 2), and 3) to determine whether some of these metrics in combination would better discriminate some cover types than the use of any one of these suites of metrics alone.
A sixth test was carried out with equal numbers of pixels in each cover type to avoid a bias in the tree toward higher accuracies for those cover types with a larger number of pixels at the expense of cover types with a fewer number of pixels. The number of pixels was equalized by replicating training pixels until each cover type included the same number as that cover type with the maximum number of training pixels. This test was carried out using the "simple metrics," the suite of metrics which produced the highest accuracies among the five tested (Figure 5).

Conclusions

On the basis of the results of these tests (Figure 5) we conclude that use of the "simple metrics" alone (test 3) is to be preferred. These metrics give substantially better results than either the Ts/NDVI (test 1) or "seasonal metrics" (test 2) alone. Moreover, applying the "logic tree" approach to the simple metrics gives similar overall accuracy but rather lower mean class accuracy and involves considerable greater computation. Combining the three sets of metrics from tests 1, 2, and 3 (test 5) gives no better results than for the "simple metrics" alone.
The test using an equal number of training pixels for each cover type (test 6) yielded a slightly higher mean accuracy for all classes, suggesting that the tree might be slightly biased toward those cover types with a large number of training pixels. From these results, we decided to use the "simple metrics" with an equal number of training pixels in each cover type to generate the final global land cover classification products.

  • To return to the homepage, click here.