Practical methods for the advancement of precision conservation via land cover classification and conformal prediction
Date
2022-12
Authors
Labuzzetta, Charles Jacob
Major Professor
Advisor
Zhu, Zhengyuan
Zhou, Yuyu
Berg, Emily
Kaleita, Amy
Miguez, Fernando
Niemi, Jarad
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Technological advancement has enabled the conversion of a vast majority of Iowa's original prairies and wetlands to agricultural use over the last 200 years. This landscape of industrial agriculture contributes to widespread environmental issues, including wind and soil erosion, poor soil health, and nutrient pollution in waterways. A promising approach to mitigating the environmental impacts of agriculture is the targeted adoption of best management practices (BMPs) such as grassed waterways, terraces, pond dams, and wascobs, among others, as outlined in the Iowa Nutrient Reduction Strategy. However, conservation efforts must keep pace with industrial-focused precision agriculture to ensure environmental goals become a priority rather than remaining subordinate to yield optimization. Innovation is necessary in the field of precision conservation to better identify locations to install conservation practices at the field-level in a targeted manner and to transform markets to provide farmers with stronger incentives to adopt such practices. A first step is to improve the record of already existing conservation practices across Iowa. In this dissertation, several practical methods to improve the Iowa BMP Mapping Project database are presented. First, we describe a deep learning approach to detecting BMPs in remote sensing imagery via image segmentation. Segmenting these features is a difficult task involving a high amount of uncertainty that cannot be adequately represented by standard classification outputs alone. Coincidentally, an accurate method for representing confidence in image segmentation problems does not exist in the literature. To solve this problem this dissertation contributes a novel method of conformal prediction which takes a calibration set subsample to produce approximately valid estimates of confidence even under covariate shift. Finally, we demonstrate that this method can be applied to image segmentation to produce valid estimates of confidence for land cover classification problems. These contributions can be used in tandem to improve upon the Iowa BMP Mapping Project database. More broadly this work can help agriculture and natural resource stakeholders identify key locations where conservation practices still need to be installed to better mitigate the negative environmental impacts of industrial agriculture in Iowa. Additionally, the advancements in conformal prediction in this dissertation are generally applicable to a variety of classification and image segmentation problems outside the fields of land cover classification and precision conservation.
Series Number
Journal Issue
Is Version Of
Versions
Series
Academic or Administrative Unit
Type
dissertation