By Edmond Muller
ABSTRACT
The development of more advanced satellite instruments demands for new processing strategies. Conventional processing methods do not yield the result that one would expect from the newest generation of satellite and airborne instruments, such as very high resolution imagery (VHR) or hyperspectal imagery (HSI). This document presents a soil iron classification (Fetotal) that was carried out by means of hyperspectral DAIS 7915 data. Two conventional pixel based classification, Maximum Likelihood and Linear Unmixing, were compared with an object-based classification called the Aggregation Mosaic Theory (AMT). The AMT uses Minimum Area and Relative Border to Neighbour (fuzzy) membership functions to (post) classify land cover into land cover mosaics. The classification results of both MLC and LUM were post classified according to the AMT concept, and compared with a similar post classified object based classification. The hypothesis was that the object based classifier would yield a higher overall accuracy. Furthermore it was expected that a post classification to a higher scale level could improve the classification accuracy and would yield crisper maps, which besides are more tailored to the needs of end-users.
In total four classifications were carried out: MLC, LUM and two object-based classifications with different parameter settings and band combinations. For the accuracy assessments, overall accuracies were calculated by means of error matrixes. Also kappa values and Z-scores were used for evaluation and mutual comparison between the classification results.
The overall classification accuracies on elementary object level were disappointingly low for all four classifications. The Kappa values were even more disappointing, with values close to zero and even negative value for the MLC. One of the object-based classifications indeed outperformed both pixel based classifications -though not convincingly- with an overall accuracy of 29.2% and a kappa value of 11%. The second object based classification -carried out with a different band combination- performed equally well as the LUM-classifier with an overall accuracy of around 21%. The MLC classifier performed most disappointing of all four classifications, and yielded an overall accuracy of only 12.5% and a negative kappa value.
Post classification improved the overall accuracy in three of the four cases, ranging from +4.2% to +16.7% improvement. However, for one of the object-based classifications, the overall accuracy did not improve by post classification at all. With respect to crispness of the map product, both pixel based classifications initially contained much more noise on elementary level than the two object based classifications. As a result these maps were consequently less attractive and less tailored to possible end-users. In all cases however, post classification removed this noise to a considerable degree while preserving the elementary map heterogeneity and improving overall accuracies.
It has not convincingly been proven that object based classifications yield better classification results for soil iron mapping in heterogeneous areas than pixel based methods. Nevertheless, they are certainly not inferior to pixel based methods and it is interesting to continue with object-based classifications in soil mineralogy research. Additional research should however be carried out concerning the control of the segmentation process when using object based approaches.