The ability to quantify soil moisture spatial variability and its temporal dynamics over entire fields through direct soil observations using remote sensing will improve early detection of water stress before crop physiological or economic damage has occurred, and it will contribute to the identification of zones within a field in which soil water is depleted faster than in other zones of a field.
The overarching objective of this research is to develop tools and methods for remotely estimating soil moisture variability in agricultural crop production. Index-based and machine learning methods were deployed for processing hyperspectral data collected from moisture-controlled samples.
In the first of five studies described in this dissertation, the feasibility of using “low-cost” index-based multispectral reflectance sensing for remotely delineating soil moisture content from direct soil and crop residue measurements using down-sampled spectral data were determined. The relative reflectance from soil and wheat stalk residue were measured using visible and near-infrared spectrometers. The optimal pair of wavelengths was chosen using a script to create an index for estimating soil and wheat stalk residue moisture levels. Wavelengths were selected to maximize the slope of the linear index function (i.e., sensitivity to moisture) and either maximize the coefficient of determination (R2) or minimize the root mean squared error (RMSE) of the index. Results showed that wavelengths centered near 1300 nm and 1500 nm, within the range of 400 to 1700 nm, produced the best index for individual samples; however, this index worked poorly on estimating stalk residue moisture.
In the second of five studies, 20 machine learning algorithms were applied to full spectral datasets for moisture prediction and comparing them to the index-based method from the previous objective. Cubic support vector machine (SVM) and ensemble bagged trees methods produced the highest composite prediction accuracies of 96% and 93% for silt-loam soil samples, and 86% and 93% for wheat stalk residue samples, respectively. Prediction accuracy using the index-based method was 86% for silt-loam soil and 30% for wheat stalk residue.
In the third study, a spectral measurement platform capable of being deployed on a UAS was developed for future use in quantifying and delineating moisture zones within agricultural landscapes. A series of portable spectrometers covering ultraviolet (UV), visible (VIS), and near-infrared (NIR) wavelengths were instrumented using a Raspberry Pi embedded computer that was programmed to interface with the UAS autopilot for autonomous reflectance data acquisition. A similar ground-based system was developed to keep track of ambient light during reflectance target measurement. The systems were tested under varying ambient light conditions during the 2017 Great American Eclipse.
In the fourth study, the data acquisition system from the third study was deployed for recognizing different targets in the grayscale range using machine learning methods and under ambient light conditions. In this study, a dynamic method was applied to update integration time on spectrometers to optimize sensitivity of the instruments. It was found that by adjusting the integration time on each spectrometer such that a maximum intensity across all wavelengths was reached, the targets could be recognized simply based on the reflectance measurements with no need of a separate ambient light measurement.
Finally, in the fifth study, the same data acquisition system and variable integration time method were used for estimating soil moisture under ambient light condition. Among 22 machine learning algorithms, linear and quadratic discriminant analysis achieved the maximum prediction accuracy.
A UAS-deployable hyperspectral data acquisition system containing three portable spectrometers and an embedded computer was developed to classify moisture content from spectral data. Partial least squares regression and machine learning algorithms were shown to be effective to generate predictive models for classifying soil moisture.
Identifer | oai:union.ndltd.org:uky.edu/oai:uknowledge.uky.edu:bae_etds-1059 |
Date | 01 January 2018 |
Creators | Hamidisepehr, Ali |
Publisher | UKnowledge |
Source Sets | University of Kentucky |
Detected Language | English |
Type | text |
Format | application/pdf |
Source | Theses and Dissertations--Biosystems and Agricultural Engineering |
Page generated in 0.0302 seconds