A Comparison of Statistical and Fuzzy Approaches for Cascade Classification of Multitemporal Remote Sensing images

This work empirically compares a probabilistic and a fuzzy technique for cascade multitemporal classification based respectively on Hidden Markov Models and on Fuzzy Markov Chains. The analysis follows the OBIA paradigm and is conducted upon two bi- temporal data sets. The performance of variants of each approach relative to maximum likelihood mono-temporal counterparts is assessed. The experiments simulate distinct operational conditions. In particular, the robustness of each method against outliers in the training sets is investigated. Further, the study assesses the improvement resulting from incorporating the prior-knowledge concerning the admissible class-transitions in the target site, within the underlying time frame, into the classification model. Differently from multitemporal methods designed to detect changes having occurred in an area over time, cascade-classification approaches explore the correlation contained in the temporal data sets in order to classify one or more images in the multitemporal sequence. Few cascade methods have been proposed so far, including statistical as well as fuzzy approaches. Whereas statistical methods benefit from a well-established theoretical foundation, fuzzy techniques can be more naturally incorporated into Object-Based Image Analysis (OBIA) interpretation models, which can rely on fuzzy rules for knowledge representation. To our knowledge no extensive comparison between these two approaches has been reported in the literature so far. This work addresses this issue and aims at experimentally comparing a probabilistic and a fuzzy approach for cascade multitemporal classification under distinct operational conditions. The probabilistic method is based on Hidden Markov Models (HMM), whereas the fuzzy approach relies on Fuzzy Markov Chains (FMC).