Subpixel Level Discrimination of Vegetable Crops in a Complex Landscape Environment

Undeterred anthropogenic activities in the forest ecosystem of India have been altering the spatial distribution and species composition of the forest ecosystem. Mitigation of human activities in deep and fragile ecosystems in global biodiversity hotspots such as the Western Ghats requires robust detection, mapping, and monitoring techniques. The cultivation of vegetable crops is an undeniable proxy for the existence of human activities in a dense forest landscape. As these crops are cultivated for personal consumption, the areal extent and spatial distribution are limited to the backyard, shifting, or terrain cultivation with areal coverage of only several square meters. Remote sensing has been extensively used for the mapping and composition classification of forests. Supervised classification is the method of choice for mapping various land cover categories including forest species. However, when there are very few or no reference pixels available in the imagery for training, the classification approach fails to label pixels. Sub-pixel level spectral mixture modeling has the potential to map sparsely distributed land covers even if there are very few or no reference pixels in the imagery. When it is not feasible to identify unambiguous reference pixels in the imagery, spectral mixture modeling, in principle, permits the infusion of independent reference spectra from a spectral library for sub-pixel lever mapping. We have explored the potential of detecting vegetable crops whose spatial distribution is very sparse and are in a complex forest landscape. For this, we adopted and implemented several sparsity-based spectral unmixing algorithms for the targeted detection and mapping of vegetable crops (banana and bitter gourd). We have implemented the methodology on airborne hyperspectral imagery acquired over the Mudumalai national forest region of the Western Ghats, India. The quality of detection suggests the possibility of mapping vegetable crops even if they occur only a fraction of a pixel or surrounded by forest canopies.