Approximate spectral clustering for unsupervised agriculture monitoring

Unsupervised clustering methods produce land cover/use identification for monitoring agricultural resources with remote sensing, with no requirement of labeled training samples. Traditional methods, which are derived from some parametric models, are often insufficient for accurate identification. In contrast, approximate spectral clustering, a recently popular manifold learning algorithm depending on graph-cut optimization, extracts classes with various characteristics using a similarity criterion describing the data properties. We show in this paper that approximate spectral clustering, with advanced hybrid similarity criteria merging different information types, can achieve high accuracies for land cover classification to monitor agricultural resources in an unsupervised manner.