Hyperspectral Imaging from Ground Based Mobile Platforms and Applications in Precision Agriculture

This thesis focuses on the use of line scanning hyperspectral sensors on mobile ground based platforms and applying them to agricultural applications. First this work deals with the geometric and radiometric calibration and correction of acquired hyperspectral data. When operating at low altitudes, changing lighting conditions are common and inevitable, complicating the retrieval of a surface's reflectance, which is solely a function of its physical structure and chemical composition. Therefore, this thesis contributes the evaluation of an approach to compensate for changes in illumination and obtain reflectance that is less labour intensive than traditional empirical methods. Convenient field protocols are produced that only require a representative set of illumination and reflectance spectral samples. In addition, a method for determining a line scanning camera's rigid 6 degree of freedom (DOF) offset and uncertainty with respect to a navigation system is developed, enabling accurate georegistration and sensor fusion. The thesis then applies the data captured from the platform to two different agricultural applications. The first is a self-supervised weed detection framework that allows training of a per-pixel classifier using hyperspectral data without manual labelling. The experiments support the effectiveness of the framework, rivalling classifiers trained on hand labelled training data. Then the thesis demonstrates the mapping of mango maturity using hyperspectral data on an orchard wide scale using efficient image scanning techniques, which is a world first result. A novel classification, regression and mapping pipeline is proposed to generate per tree mango maturity averages. The results confirm that maturity prediction in mango orchards is possible in natural daylight using a hyperspectral camera, despite complex micro-illumination-climates under the canopy