Relating the Land-Cover Composition of Mixed Pixels to Artificial Neural Network Classification Output

ses, the classification procedures generally used to produce a ~ ~ t i f i ~ i ~ l neural networks are attractive for use in the classiland-cover map are "hard" techniques which force allocation fication of land cover from remotely sensed data. In common to One class. Moreover, depending on the nature of the mixwith other classification approaches, artificial neural netture and its composite spectral response, the allocated class works are used typically to derive a "hard" classification, need not even be one of the pixel's component classes with each case (e.g., pixel) allocated to a single class. How1987). ever, this may not always be appropriate, especially i f mixed The proportion of mixed pixels generally increases with pixels are abundant in the data set. This paper investigates a coarsening of the spatial resolution of the sensing system the potential to derive information on the land-cover compo(Townshend and Justice, 1981; Crap~er , 1984). Consequently, sition of mixed pixels from an artificial neural network clasthe effects of the mixed pixel problem may be felt most sification. The approach was based on relating the activation strongly when resolevel of artificial neural network output units, which indicate lution data sets. Unfortunate1~, many regi0na1 global the strength of class membership, to land-cover composition, scales studies are often constrained to the use of relatively Two case studies are discussed which illustrate that the coarse spatial resolution sensor data. At regional to global vation level of the artificial neural outputs themscales, however, existing land-cover data sets are known to selves were not strongly related to pixel composition. be of poor quality, and remote sensing is the only feasible However, re-scaling the activation levels, to remove the bias approach for land-cover mapping (Townshend et ~ l . 1 lggl; towards very high and low strengths of class membership imDeFries and Townshend, 1994). For instance, maps of tropiposed by the unit activation function, produced measures cal vegetation are required to assess the role of land-cover that were strongly related to the land-cover composition of change, particularly deforestation$ On the global mixed pixels. In both case studies, significant correlations (Wisniewski and Sampson, 1993). The available land-cover (a11 r > 0.8) between the re-scaled activation level of an outdata sets, however, vary considerably and, consequently, esput unit and the percentage cover of the class associated timates of phenomena such as deforestation vary markedly with the unit were obtained. (Grainger, 1993; Curran and Foody, 1994), limiting our understanding of the carbon cycle. Although remote sensing Introduction has considerable potential for mapping tropical land cover, Remotely sensed data have been used to map land cover at a the only practical sensing Use is the AVHRR range of spatial and temporal scales. The accuracy and value which has a relatively coarse spatial resolution, 1.1 km at of the derived land-cover maps are dependent on a range of best. The large proportion of mixed pixels in AVHRR data can factors related to the data sets and methods used. ~ h ~ ~ , for lead to significant errors in the estimation of forest extent example, the accuracy of maps derived from conventional and its change over time et ~ l .2 lggl; SkO1e and supervised image classification techniques is a function of Tucker* lgg3; Curran and FOOdyj lgg4). factors related to the training, allocation, and testing stages of "hard" image techniques the classification (e.g., Swain, 1978; Thomas et al., 1987). may therefore provide a poor representation of the distribuconventional image classification techniques assume that tion of land cover and be a poor base for the estimation of all the pixels within the image are pure, that is, that they the areal extent of land-cover classes. In some applications it represent an area of homogeneous cover of a single landis therefore desirable to unmix pixels into their component cover class. This assumption is often untenable with pixels Parts. A range mixture have been develof mixed land-cover composition abundant in an image. oped for this task (e.g., Clark and Canas, 1993; Holben and These land-cover class mixes may arise horn the gradual inShimabukuro, 1993; Settle and Drake, 1993). Of these, linear tergradation of continuous land-cover classes (Csaplovics, mixture models are the most widely used. These, however, 1992; Foody et al., 1992) or, perhaps more commonly, as a be as the consequence of the relationship between the sensor's spatial ten including normally distributed data as we11 as linear resolution and the fabric of the landscape (Irons et a]., 1985; mixing, are often untenable. One simple alternative approach Campbell, 1987). Irrespective of their origin, mixed pixels which is widely available, and may be appropriate when the are a major problem in land-cover mapping applications. For aim is to map land cover, is to "soften" the classification example, while a mixed pixel must contain at least two clasPhotogrammetric Engineering & Remote Sensing, Department of Geography, University of Wales Swansea, SingVol. 62, No. 5, May 1996, pp. 491-499. leton Park, Swansea, SA2 8PP, United Kingdom. Presently with the Telford Institute of Environmental Sys0099-1112/96/6205-491$3.00/0 terns, Department of Geography, University of Salford, Sal