Unsupervised extraction of greenhouses using approximate spectral clustering ensemble

Monitoring and mapping greenhouses are important for yield estimation, sustainable crop production, residue management and environmental impact. Conventional approaches based on in situ surveys, which are costly and time consuming, are being replaced by supervised classification of commonly used features extracted from very-high spatial resolution images. Alternatively, we extract (both plastic and glass) greenhouses from Worldview-2 images in an unsupervised manner by approximate spectral clustering ensemble using hybrid geodesic similarity criterion. Our proposed approach is promising for automated detection of greenhouse areas with limited user information and outperforms earlier unsupervised extraction methods for greenhouses.