Sub-pixel Discrimination of Soil and Crop in Drone-based Hyperspectral Imagery

Adapting technological interventions for plant-level nutrition optimization and pest-disease control requires detection and spatial aggregation of agricultural plant material—leaves, canopy gaps, and soil exposure. The potential of within-plant canopy crop-soil discrimination with sensitivity to canopy gaps and structure is not yet reported. This research aims to discriminate crop-soil within-plant canopies for plant-level discrimination using high-resolution hyperspectral imagery. A drone-mounted frame-based hyperspectral imaging sensor in the VNIR area is used for acquiring hyperspectral images of vegetable crops over an experimental farm in Bangalore, India. Further, reference in situ spectral reflectance measurements at plant-level of background soil, and crops are recorded from hyperspectral spectroradiometer and are used for generating a spectral library. Spectral signatures from the library are used as endmembers to simulate within-plant canopy level discrimination of plant and soil. Abundances of crop and soil are retrieved using various linear, non-linear, and sparse-based spectral unmixing approaches. The results are validated using ground truth fractional measurements carried out over diverse plant development and lighting conditions. The comparison of estimated and measured abundances indicates that plant-level soil and discrimination is possible with about 97% accuracy.