Unsupervised extraction of greenhouses using WorldView-2 images

Greenhouses provide controlled areas to produce high-quality and wide variety of agricultural products. However, their monitoring and mapping are essential in many aspects such as yield estimation, sustainable crop production, residue management and environmental impact. Increased coverage area in recent years poses another necessity for advanced methods in addition to the (costly and time consuming) conventional techniques. Supervised classification methods have been currently used for greenhouse mapping based on well-known features (extracted from very-high spatial resolution images) and classifiers. Our contribution is to extract greenhouses in an unsupervised manner: we utilize spectral features and Gabor textural featuresextracted from WorldView-2 images, and then we cluster the combined features by an approximate spectral clustering method based on a local density based similarity. Our experimental results are very promising for effective and automated detection of greenhouse areas with very limited user information.