Principal component analysis on multispectral data “Landsat-TM” and interpretation of vegetation cover in the Tejeda and Almijara mountain ranges

Abstract

The simultaneous consideration of all reflectance bands during image processing of multiespectral, remotely-sensed data, usually involves a very large volume of data to handle. However, this increase in data volume is not followed by a similar increment in the volume of additional information that is accounted for, which is due to the existence of a very high correlation between reflectance bands. In this paper we report the results of applying a multivariate statistical tool (Principal Component Analysis, PCA) to the first four reflectance bands of a LANDSAT «Thematic Mapper» image of Sierra de Tejeda and Sierra de Almijara (Southern Spain). The aim of this analysis was to achieve a reduction in the volume of data to handle during image processing while avoiding a significant loss of information. The first two principal components resulting from the analysis accounted for more than 99% of the total variance in the original data set. The image obtained through the first principal component transform could be interpreted as a weighed-sum image of all reflectance bands (thus similar to a black and white panchromatic photograph with an extension into the infrared wavelength). This image showed optimum feasibility for the interpretation of geological, hydrological and topographic features of the study area. The second principal component transform was able to extract the «green» signal of the vegetation canopy from the complex set of multiespectral data. Relief features (i.e., areas of shadow in north-facing slopes versus well-illuminated, south-facing slopes) did not appear in the resulting image (these features had already been accounted for by the first principal component), which is an advantage when studying vegetation cover in areas with steep relief. A strong possitive correlation was found between the second PCA image and those resulting from the application of standard spectral vegetation indexes (RVI and NDVI). The initial assesment of the images has shown that severe deaforestation is undergoing in the area. About 30% of the study area was occupied by bare soils. This figure went up to 85% if areas with scattered vegetation were also included. Multivariate analysis has confirmed as a valuable tool for the interpretation of multiespectral satellite data from the study area. Eventually, the obtention of a principal component transform based on the selective sampling of appropiate subsets of pixels from the area would provide of algorithms that applied to the entire set in multitemporal images could, for instance, be used to monitor long-term and large-scale deaforestation processes in the region.