Mapping sub-pixel variation in land cover at the global scale using NOAA AVHRR imagery

Remote sensing studies have tended to be conducted at local to regional scales. However, recently attention has increasingly been focused at regional to global scales. Remote sensing imagery with moderate and coarse spatial resolutions (for example, the National Oceanographic and Atmospheric Administration (NOAA) Advanced Visible High Resolution Radiometer (AVHRR) with a spatial resolution of (1.1 km by 1.1 km) are ideal for this since large contiguous regions are completely covered and data are provided synoptically. One problem which such imagery is that the intrinsic scale of spatial variation in much land cover (for example, New Forest area in the U.K.) is finer than the scale of sampling imposed by the image pixels. The result is that most pixels contain a mixture of land cover classes.Where the intrinsic scale of spatial variation at the ground is smaller than or equal to the scale of sampling imposed by the image pixels the objective should not be to assign a pixel to a single class (as with standard traditional spectral classification for example, maximum likelihood), but to several classes, where the sum of the proportions sum to one. There are, however, several alternatives. This thesis compares three possible techniques for classifying sub-pixel variation in land cover at the global scale: (i) artificial neural networks (ANN), (ii) mixture modelling (MM), and (iii) fuzzy c-means classification (FCM). The specific objective was to evaluate these techniques at the global scale, that is, for mapping sub-pixel variation in land cover over the entire globe for five text sites which involve drastically different land cover types.The major findings of the thesis are as follows: the ANN was consistently found to be more accurate than FCM classification (and MM) when trained with two-thirds of the available data. The ANN was found to be more accurate than (pilot project) and of similar accuracy to (main project) FCM (and more accurate than MM, pilot project only) when trained with an equivalent number of data. The ANN trained with data selected systematically from the available NOAA AVHRR imagery was of greater accuracy than the ANN trained with data representing the largest class proportions (obtained from the fine spatial resolution imagery). The FCM classifier trained with pixels selected from the NOAA AVHRR image using judgement was found to be sometimes more accurate and sometimes less accurate than the FCM classifier trained with the largest class proportions (obtained from the fine spatial resolution imagery).