This dissertation is comprised of three papers. The first paper describes in detail a planar dielectric probe design using finite element analysis to determine sensing range and efficiency. The probe is subsequently connected to a Keysight impedance analyzer to measure dielectric properties of raw cotton at controlled levels of moisture content, compressed densities, and source frequency sweeps. Sensitivity to compositional differences such as turnout (lint vs seed) and variety is also explored. The response to the different factors is shown graphically and further quantified statistically in the form of a predictive model for the complex permittivity (dielectric constant and loss tangent).

The second paper extends the dielectric probe used in the first paper to real-time harvesting on a round-module cotton harvester by leveraging a packaged sensor with embedded impedance measurement circuit and probe all in one mobile unit. A moisture prediction model based on permittivity is developed from lab-measured data and adjusted based on field data collected during cotton harvesting in Fall of 2014 for pickers and Spring of 2015 for strippers. Verification of the prediction accuracy is performed on field data collected during cotton harvesting in 2016. Sources of variability and sensitivity to confounding factors are investigated and quantified. Finally, plots of diurnal trends of predicted and actual moisture content are overlaid for several days of harvesting.

The third paper draws on the first two in applying capacitive-based moisture sensing to large-square bales of alfalfa. A lab characterization is performed on alfalfa over a wide range of moisture contents and densities using both the Keysight impedance analyzer and packaged sensor to measure permittivity. Field data (on-machine permittivity measurements of bales and corresponding ground truth moisture content) is subsequently collected during baling in 2015 and 2016 for alfalfa hay (<30%) and silage (>30%) and used for training and validation of prediction models. In following with the other two papers, sources of variability are discussed and sensitivity to factors quantified. Limitations in sensing range of the packaged sensor lead to multiple prediction models: a simple but limited model restricted to hay and another using modern fitting techniques (feature engineering and artificial neural network) for both hay and silage. Real-time filtering of the prediction signal is investigated using the simple model in light of what seems like mechanically induced oscillations, using a Kalman filter to isolate and remove them while minimizing delay. The real-time prediction signal is finally overlaid with actual moisture content found from core samples of the same bales.