Remote sensing is a common tool in agriculture for crop classification and monitoring. Globally available high resolution imagery makes it possible to delineate individual field boundaries, which can serve as a foundational data set for secondary agricultural analysis. For the most accurate results, the methods employed in field classification and delineation are often fine-tuned to the agricultural conditions within the local geographic context of interest. This fine-tuning, however, can make it difficult to implement the same models in other locations to the same degree of accuracy. While these locally-tuned examples provide valuable insight into developing crop classification systems, a classification model that is applicable at larger spatial scales (e.g. national, global) requires a different approach. This paper proposes a geographically scalable workflow that integrates unsupervised machine learning, local knowledge, and emerging deep learning techniques to enable accurate, flexible field delineation. This approach achieved 88.1% overall accuracy in classifying agricultural fields for the state of Minnesota without relying on any external training data. This workflow can serve as a prototype for globally scalable field mapping.
University of Minnesota M.A. thesis. August 2020. Major: Geography. Advisor: Eric Shook. 1 computer file (PDF); iv, 104 pages.
Every Field in Minnesota: Building a Geographically Scalable Satellite Imagery Analytics System for Mapping Crop Fields.
Retrieved from the University of Minnesota Digital Conservancy,
Content distributed via the University of Minnesota's Digital Conservancy may be subject to additional license and use restrictions applied by the depositor.