Imaging Science Thesis
Defense
On the Optimization of
Sensors for the Remote Sensing of Crop
Yield
Nate Burglewski
Imaging
Science
Rochester Institute of
Technology &nb
sp;
Register for
Zoom Link Here
Abstract:
Remote sensing has long
been a part of precision agriculture research, with products related to
spatially-explicit, within-field management of yield, disease, and
harvest scheduling, to name a few application areas that are of interest
to industry. These spaceborne, airborne, and in-situ sensors are used to
monitor crop health throughout the growing season, inform management
decisions for farmers, and dictate logistical or planning operations.
Crop yield forecasting, specifically, comes in many forms, but recently
machine learning algorithms have gained prominence, at the cost of more
traditional parametric regression models. While much work has been
accomplished towards grain yield estimation using remote sensing data,
silage yield forecasting has been less investigated. Furthermore, to date
no sensor has been designed to optimize the spatial-spectral trade space
that is relevant to yield estimation. We thus sought to identify the
ideal spectral and spatial sampling parameters for a sensor to be used
for this purpose. We first set out to identify a yield estimation
algorithm which was favorable for both grain and silage yield, finding
that support vector regression performed best at both field-level and
regional yield estimates for corn grain. We then leveraged three separate
spectral band selection algorithms to quantitatively assess the most
important spectral content found in our high spatial resolution, high
spectral dimensional imaging spectroscopy data. These data were resampled
both spectrally and spatially to determine the interplay of spatial
resolution and spectral content on corn grain and silage yield estimation
results. Corn grain yield estimation accuracies ranged from 7.02% mean
absolute percentage error (MAPE) (root mean square error [RMSE] = 0.332
Mg/ha) at 0.06 m GSD to 11.45% MAPE (RMSE = 0.435 Mg/ha) at 30 m GSD
using data collected at the same five band multispectral sampling
configuration during the R4 (dough) growth stage. We found that this
favorability extended to silage yield estimation at the same growth
stage, with MAPE ranging from 2.22-2.84% MAPE (0.62-0.68 Mg/ha) from
0.06-30 m GSD.
We thus were able to
identify the spectral features which were ideal for this purpose using
high resolution imaging spectroscopy data, while also investigating the
changes in estimation accuracies as a function of spectral content and
spatial resolution. Ultimately, the most favorable spectral configuration
was multispectral, containing blue (450 nm), green (550 nm), red (670),
and three near-infrared spectral bands (770, 830, and 920 nm). This
arguably is an ideal outcome, as sampling spectral content along these
bandpasses greatly reduces dimensionality resulting from imaging
spectroscopy data, while preserving the most important features for a
machine learning regression algorithm to exploit. This spectral
configuration was found to give the most accurate regression results,
specifically centered at 8 -16 m GSDs for corn silage yield, even though
there were changes in accuracy across a wide range of spatial
resolutions. We attributed this result to matching the semivariance range
of the yield data, which in turn generated favorable modeling conditions
for the support vector regression estimation. For corn grain, the average
semivariance range was greater than our largest sampling distance, so in
general the higher spatial resolutions below 1 m GSD were most accurate.
These findings inform sensor engineers of the optimal spectral sampling
configuration for an idealized sensor for use in corn grain or silage
yield estimation at any spatial scale. Future work should address the
underlying yield drivers behind these selections to further isolate the
spectral content correlated to crop health and status metrics like
nitrogen content, heat stress, water stress, and structural parameters.
These are all correlated with yield and thus important to stakeholders,
leading to potential mitigation via management techniques to improve
yield projections.
Intended Audience: All are
Welcome!
To request an
interpreter, please visit myaccess.rit.edu