Calibration to maximize temporal radiometric repeatability of airborne hyperspectral imaging data.

Many studies provide insight into calibration of airborne remote sensing data but very few specifically address the issue of temporal radiometric repeatability. In this study, we acquired airborne hyperspectral optical sensing data from experimental objects (white Teflon and colored panels) during 52 flight missions on three separate days. Data sets were subjected to four radiometric calibration methods: no radiometric calibration (radiance data), empirical line method calibration based on white calibration boards (ELM calibration), and two atmospheric radiative transfer model calibrations: 1) radiometric calibration with irradiance data acquired with a drone-mounted down-welling sensor (ARTM), and 2) modeled sun parameters and weather variables in combination with irradiance data from drone-mounted down-welling sensor (ARTM+). Spectral bands from 900-970 nm were found to be associated with disproportionally lower temporal radiometric repeatability than spectral bands from 416-900 nm. ELM calibration was found to be highly sensitive to time of flight missions (which is directly linked to sun parameters and weather conditions). Both ARTM calibrations outperformed ELM calibration, especially ARTM2+. Importantly, ARTM+ calibration markedly attenuated loss of radiometric repeatability in spectral bands beyond 900 nm and therefore improved possible contributions of these spectral bands to classification functions. We conclude that a minimum of 5% radiometric error (radiometric repeatability<95%), and probably considerably more error, should be expected when airborne remote sensing data are acquired at multiple time points across days. Consequently, objects being classified should be in classes that are at least 5% different in terms of average optical traits for classification functions to perform with high degree of accuracy and consistency. This study provides strong support for the claim that airborne remote sensing studies should include repeated data acquisitions from same objects at multiple time points. Such temporal replication is essential for classification functions to capture variation and stochastic noise caused by imaging equipment, and abiotic and environmental variables.

Data-driven vermiculite distribution modelling for UAV-based precision pest management.

In recent decades, unmanned aerial vehicles (UAVs) have gained considerable popularity in the agricultural sector, in which UAV-based actuation is used to spray pesticides and release biological control agents. A key challenge in such UAV-based actuation is to account for wind speed and UAV flight parameters to maximize precision-delivery of pesticides and biological control agents. This paper describes a data-driven framework to predict density distribution patterns of vermiculite dispensed from a hovering UAV as a function of UAV's movement state, wind condition, and dispenser setting. The model, derived by our proposed learning algorithm, is able to accurately predict the vermiculite distribution pattern evaluated in terms of both training and test data. Our framework and algorithm can be easily translated to other precision pest management problems with different UAVs and dispensers and for difference pesticides and crops. Moreover, our model, due to its simple analytical form, can be incorporated into the design of a controller that can optimize autonomous UAV delivery of desired amount of predatory mites to multiple target locations.