The effect of the thermal infrared data on principal component analysis of multi-spectral remotely-sensed data

The physical interpretation of principal component analysis of multispectral imagery has been established for reflective multi-band remotely-sensed data. Attempts to include the thermal data have not yielded an improvement in classification results despite the fact that the thermal data carries additional unique discriminative information as it depends on the bulk properties of the ground composites as well. This issue is studied by using high resolution groundbased images and previously published data. The inclusion of the thermal data disturbs the original spectral composition of the eigenvectors set only if two conditions are met, which is frequently the case. The first is the existence of at least one eigenvalue whose magnitude is approximately unity. If such eigenvalue is indeed present, then the corresponding eigenimage must have a substantial correlation with the thermal image, and this forms the second condition. The second ranked principal component which is associated with the typical reflectance of vegetation frequently fulfills these conditions and hence is affected by the inclusion of the thermal data. Some methods to enhance the utility of joint analysis of reflective and thermal remotely sensed data are presented and discussed.