Several image classification methods for large-area forest delineation from AVHRR data are found in the literature and these may be grouped into four main categories.
This is the simplest and most flexible approach, since the eye-brain system can make allowances for the inevitable large variations in feature appearance that occur with different viewing angles and differential bidirectional reflectance of various surfaces. At the same time local variations can be detected for improved feature discrimination. Visual interpretations made by trained people familiar with the relevant area and image data characteristics, can be accurate in many cases. However, the delineation of boundaries is also time-consuming, difficult and somewhat subjective. If the interpretations are made on a hard-copy image rather than on-screen image displays, then it often requires digitisation for input into a GIS. A further drawback of visual interpretation of hard-copy images is that it can be based on an image representation of, at most, three spectral bands at any one time, and the results can be quite dependent on the contrast-stretch or enhancement used. However, this method is easy to understand and teach, and it requires no specialised hardware. Examples of the successful use of this method are provided in Tucker et. al. (1984), Malingreau and Tucker (1990) and Malingreau et. al. (1989).
Single-image radiance or brightness temperature thresholds applied to channels 3 and/or 4 have been found to delineate quite well the forest/non-forest boundary, at least in areas where the transition is very sharp, e.g. Malingreau and Tucker 1988, Malingreau et. al.1989, Stone et. al. 1991; Woodwell et. al.1986, Amaral 1992, Cross 1990, Cross 1991). The technique is simple to apply and understand, but requires visual interaction with the data. Also, even in the cases cited, the thresholds used have differed both in time and space, and the method is not reliable in areas of diffuse or complex forest/non-forest interfaces.
This technique offers an improvement on the previous methods, because it is said to be more quantitative, objective and can be based on more than three channels at a time. Examples of its successful application are provided in Cross (1990), Stone et. al. (1991), Paivinen and Witt (1988) and Horning and Nelson (1993). Normally, a supervised classification requires recourse to quite detailed geo-located reference data that define the locations and extent of representative land cover units for classification training and accuracy assessment procedures. An unsupervised approach can be more automatic in the first stage, but requires significant human interaction and reference data to assign meaningful classes to the resultant spectral clusters. In some cases, however, supervised and unsupervised approaches have been found to yield quite similar results, as in the mapping of deforestation in Rondonia state in Brazil (Skole 1993).
Because NOAA AVHRR images cover such large areas, the quality of classification may vary across the image, or some parts of some images may be obscured by haze or cloud. In these cases, a successful approach by D’Souza et. al. (1995) was to carry out several image classifications for overlapping areas, and assign a reliability value at each location. The classification with the most reliable indicator was then retained in the final product.
This method as used by Achard et. al. (1993) involves the examination of the behaviour of a particular radiometric characteristic (e.g. vegetation index or surface temperature) through time. This is the most sophisticated and best method for making a seasonal/evergreen forest discrimination. However, it relies on a good set of high quality (cloud and haze-free) imagery throughout the season and is therefore normally only practical in tropical areas where there is a marked dry season, or when a large series of data is available. This technique usually involves the reduction of the AVHRR time series into weekly, 10-day or monthly maximum value composites (Holben, 1996), and until recently were performed on very low spatial (nominally 8km) sampled NOAA AVHRR data (e.g. Townshend et. al., 1987).
With the recent improvements in computing power, and in the global acquisition and processing of NOAA AVHRR data (see D’Souza et. al., 1995), 10-day maximum value composites have recently been developed at continental and global scales for the full spatial resolution data. Loveland et. al., (1991) showed that such monthly full spatial resolution NDVI composites could be used to characterise land cover types at continental scales and a resolution of nominally 1km. In their work NDVI monthly composites were used to develop a land cover database for the conterminous United States. Eight monthly NDVI composites were used in an automatic unsupervised classification scheme to produce 70 spectral temporal “seasonally-distinct” classes. These were then refined and used with expert knowledge and a wide variety of ancillary data to subjectively assign class labels to each pixel. The labels assigned were taken from the USGS land cover classification. This work is continuing under the auspices of the International Global Biosphere Programme (IGBP). See Belward and Loveland (1995) and section later.
Other work that uses multitemporal NDVI composites seeks to extract phyto-phenological measures or characteristics rather than categoric land cover classes from the data (see Lloyd 1990). Such measures included maximum photosynthetic activity, the length of growing season, and mean daily NDVI. DeFries et. al. (1995) also concluded that decision-tree classification of phyto-phenological measures proved to be more accurate than straight classification of monthly composited NDVI values alone. This method of classification is probably more reliably repeatable because it describes and characterises land cover types according to their radiometric behaviour through time, rather than the forcing a correspondence to a classification scheme developed on some other bases. However, it has not achieved widespread acceptance because the final classes do not correspond to the well-established land cover classifications.
It has also been suggested that the multi-temporal examination surface temperature together with a spectral index may increase the discrimination of regional land cover classes (e.g. Running et. al., 1994). Generally, surface temperature is observed to be inversely proportional to the amount of vegetation canopy due to a variety of factors including latent heat transfer through evapotranspiration and the lower heat capacity and thermal inertia of vegetation compared with that of soil (Nemani et. al., 1993). Multitemporal NDVI and surface temperature have been used together to produce land cover classifications of the conterminous U.S. (Nemani and Running, 1997), a forest/non-forest classification of Europe (Roy, 1997) and land cover change maps of Africa at 8km resolution (Lambin and Ehrlich, 1996)
Regression analyses and mixture-modelling approaches may assign several classes in different proportions to a single pixel. The results of regression analysis or mixture modelling are usually a set of images showing the concentrations of a different class component (called endmembers) in each image. Although they do not provide any more detailed about the location of forest/non-forest boundaries and sub-pixel forest locations, the overall accuracy of total forest cover estimation is often improved. The mixture modelling approaches assume some form of scene reflectance model (implicit or explicit) that describes the mixing relationship between the different scene components. Iverson et. al. (1989) used an implicit model based on forest/non-forest estimates made from a sample of overlapping Landsat TM and NOAA AVHRR images. The percentage forest cover estimates made from TM classifications were related to the red and near-infrared AVHRR pixel values by linear multiple regression. Then the resulting regression coefficients were used to estimate the percentage forest area over the rest of the AVHRR imagery. This approach has been used to make regional forest cover maps of the U.S. (e.g. Zhu and Evans, 1992; Iverson et. al., 1994; Ripple, 1994), of central Africa (DeFries et. al., 1997) and of the conterminous U.S. (Zhu and Evans, 1994). Regression analysis has also proven to be effective for biomass estimation of coniferous forests, but less successful where the target area includes conifers and broad leaves trees (Hame et. al., 1997).
Cross et. al. (1991) produced tropical forest proportion images by linear mixture modelling four channels of an AVHRR sub-image. The endmembers were derived by examination of the principal components analysis of the AVHRR data. The resultant forest and non-forest proportion images were compared with supervised classifications of co-registered Landsat TM images. Cross et. al. (1992) concluded that 60–70% of an AVHRR pixel must be forested before it is recognised as such. Holben and Simabukuro (1993) presented a method to extract the reflected component of the mid-infrared AVHRR channel and used it with the red and near-infrared AVHRR channels to generate vegetation, soil and shade proportion images using a linear mixture model.
The attraction of mixture modelling and regression analysis approaches with NOAA AVHRR data is that they appear to present a more realistic case of what exists on the ground. Over a 1km area, there is likely to be a mix of vegetation types, and the assignment of just one class (e.g. forest) to the whole 1km area seems unrealistic. The assignment of a proportion (e.g. 70% forest) is more satisfying. However, the application of pure mixture modelling techniques is not straightforward. The initial derivation of endmembers is crucial to the accuracy of the mixture modelling approach, but it is complex and not well understood. It is generally recognised that for successful mixture modelling much more spectral information is required where a complex mix of land cover types exist.
Regression analysis has been used to create an AVHRR-based forest probability map of the Pan-European area (Hame et. al., 1999). However, such techniques have a requirement for a large amount of independent, reliable and geo-coded reference data, and a set of very clean noise-free imagery.
If image classification is reliable and repeatable, then the change in land cover can be identified between classifications made at different times. However, in nearly all of the reported forest classification exercises using NOAA AVHRR, the objective has been to achieve the classification. Very few studies have looked at or made repeated classifications at different time periods. One notable exception is the work of Cross (1991), in which he tested image difference, image division, image regression and edge difference techniques and applied them to AVHRR images of Rondonia and Mato Grosso from 1984, 1988 and 1990. Large scale changes such as the establishment of new “fazendas” and other ranching and “fish-bone” and other forest clearances for agriculture were clearly determinable from the examination of the individual channels on single-date AVHRR imagery.
Because of the lack of image classification repeatability tests in the literature, it is very difficult to assess the utility of NOAA AVHRR for forest extent or condition monitoring or change detection. However, it appears that where changes are abrupt, marked and larger than 1km square (e.g. denudations of dense forest and conversion to agriculture or bare soil) they are easily discernible from individual, good quality single-date full-resolution AVHRR images. However, land conversions which are more gradual (e.g. degradation or reduction of forest cover percentage) or of a smaller extent would not be easily discernible, even with recourse to multi-temporal image examination. For monitoring of such areas, some suggestions have been made as to the utility of AVHRR for circumstantial or qualitative evidence gathering (e.g. GOFC 1999) to identify deforestation “hotspots”, either independently or in conjunction with a multi-level approach using NOAA AVHRR together with higher spatial resolution imagery and ground survey for measurement of the nature or amount of the observed change (see section later and, for example, Jeanjean and Achard 1997).
