Nature-based metaheuristic optimisation techniques for band selection in drone-based hyperspectral images for plant-level crop classification

Hyperspectral imagery-based remote sensing has emerged as a major contributing factor for precision agriculture. Drone-based hyperspectral imagery provides high resolution data for plant-level analysis of agriculture fields, making it possible to discriminate crops with subtle differences in their spectral signatures. However, the high dimensionality and correlated bands in the hyperspectral imagery make data processing challenging in an operational environment. Feature selection plays a vital role in reducing the dimensionality of hyperspectral data without losing spectral integrity. Though evolutionary computing-based optimal feature selection has been applied for hyperspectral data at a broader level, their applications are not explored for crop classification in drone-based hyperspectral imagery, especially at plant-level. This work has evaluated several nature-inspired metaheuristic optimisation algorithms to classify drone hyperspectral imagery at different flying altitudes for crop classification at plant level. We acquired ultra-high spatial resolution hyperspectral imagery from a drone over research farms of the University of Agricultural Sciences, GKVK, Bengaluru, India. From the perspective of studying drone-based hyperspectral imaging for plant-level mapping and the interference of background introduced by varying flying height, this work is the first of its kind to study the impact of the optimisation techniques for hyperspectral band reduction on drone data acquired at different heights. The models are trained on a dataset captured at the height of 30 m and further tested on images acquired over 25 m and 40 m. The outcomes suggest the requirement of only selected bands from certain wavelengths for achieving classification accuracy as compared with the original set of bands. The results also emphasise the effective application of evolutionary computing-based optimisation techniques for hyperspectral band reduction, irrespective of the height of the data capture.