High-spectral-resolution hyperspectral data are acquired by sensors that gather images from hundreds of narrow and contiguous bands of the electromagnetic spectrum. These data offer unique opportunities for characterization and precise land surface recognition in urban areas. So far, few studies have been conducted with these data to automatically detect and estimate areas of photovoltaic panels, which currently constitute an important part of renewable energy systems in urban areas of developed countries. In this paper, two hyperspectral-unmixing-based methods are proposed to detect and to estimate surfaces of photovoltaic panels. These approaches, related to linear spectral unmixing (LSU) techniques, are based on new nonnegative matrix factorization (NMF) algorithms that exploit known panel spectra, which makes them partial NMF methods. The first approach, called Grd-Part-NMF, is a gradient-based method, whereas the second one, called Multi-Part-NMF, uses multiplicative update rules. To evaluate the performance of these approaches, experiments are conducted on realistic synthetic and real airborne hyperspectral data acquired over an urban region. For the synthetic data, obtained results show that the proposed methods yield much better overall performance than NMF-unmixing-based methods from the literature. For the real data, the obtained detection and area estimation results are first confirmed by using very high-spatial-resolution ortho-images of the same regions. These results are also compared with those obtained by standard NMF-unmixing-based methods and by a one-class-classification-based approach. This comparison shows that the proposed approaches are superior to those considered from the literature.

Keywords: Object-Based Image Analysis, Data Mining, Classification rule Induction, Genetic Algorithm The success of the object-based image analysis (OBIA) paradigm can be attributed to the fact that regions obtained by means of segmentation process are depicted with a variety of spectral, shape, texture and context characteristics. These representative objectsattributes can be assigned to different land-cover/land-use types by means of two options. The first is to use supervised classifiers such as K-nearest neighbors (KNN) and Support Vector Machine (SVM), the second is to create classification rules. Supervised classifiers perform very well and have generally higher accuracies. However one of their drawbacks is that they provide no explicit knowledge in understandable and interpretable forms. The building of the rule set is generally based on the domain expert knowledge when dealing with a small number of classes and a small number of attributes, but having a dozens of continuously valued attributes attached to each image object makes it a tedious task and experts quickly get overwhelmed and become totally helpless. This is where data mining techniques for knowledge discovering help to understand the hidden relationships between classes and their attached attributes. The aim of this paper is to highlight the benefits of using knowledge discovery and data-mining tools, especially rule induction algorithms for useful and accurate information extraction from high spatial resolution remotely sensed imagery.