Rapid Data Analytics of Remote Sensing Data for Phenotyping and Precision Agriculture Applications

Abstract

Plant breeding programs require rapid, reliable, and scalable methods to evaluate crop performance across hundreds and thousands of field plots. Traditionally, breeders collect phenotypic information through manual field scouting, visual scoring, handheld sensors, and post-season yield measurements. Although these methods are useful, they are labor-intensive, time-consuming, subjective, and difficult to apply at a large scale. In recent years, Unmanned Aerial Vehicles (UAVs) equipped with multispectral sensors have become important tools for high-throughput phenotyping because they can collect high-resolution spatial and spectral information over large breeding trials within a short time. However, collecting UAV imagery is only one part of the workflow. The major challenge begins after image acquisition, where researchers must process large orthomosaics, identify individual plot locations, create plot boundaries, assign unique plot identifiers, extract vegetation indices, merge spatial outputs with field data, perform statistical analysis, build predictive models, and visualize results. These steps are often fragmented across different software tools and require repeated manual work, GIS knowledge, and programming skills. Therefore, this dissertation focused on developing automated, reproducible, and user-friendly workflows for rapid data analytics of UAV-based remote sensing data in field-based phenotyping and precision agriculture applications. The first study developed an automated pipeline for plot boundary delineation, data extraction, and mapping using planter-based RTK-GNSS data and UAV multispectral imagery. Instead of depending only on visible crop rows or image-based segmentation, the developed workflow used high-accuracy planter GPS points to generate plot boundaries based on actual planting geometry. The pipeline cleaned and processed RTK-GNSS data, grouped GPS points by planter rows and plot positions, converted points to centerlines, generated parallel lines based on planter spacing, created buffered polygon boundaries, assigned unique plot identifiers, and extracted plot-level vegetation index values using zonal statistics. The pipeline produced shapefiles, CSV files, vegetation-index maps, and plot-level data tables for downstream analysis. The results showed that the automated plot boundaries aligned well with UAV imagery and remained useful across different crop growth stages, including situations with canopy overlap, sparse vegetation, lodging, and drifted plots. Compared with manual GIS-based plot drawing and zonal-statistics extraction, the automated workflow reduced time, labor, and user involvement by approximately 90%. The second study developed RY Sense, an automated yield prediction workflow for soybean breeding programs using UAV-derived remote sensing data. This study integrated RTK-GNSS-derived plot boundaries, multispectral vegetation indices, spectral bands, canopy coverage, maturity information, and Gray-Level Co-occurrence Matrix texture features for soybean yield prediction. Five machine learning models, including Support Vector Regression, Random Forest, XGBoost, AdaBoost, and K-Nearest Neighbors, were evaluated across multiple soybean growth stages. The results showed that yield prediction performance varied across growth stages, feature groups, and algorithms. The R6 and R7 stages were the most informative for soybean yield prediction, and Support Vector Regression achieved the highest prediction accuracy, with R² reaching approximately 0.76. The combination of vegetation indices and texture features improved prediction performance, showing that both spectral and structural canopy information are important for modeling yield variability. The third study developed AutoPlotlytics, a modular graphical user interface for spatial, tabular, statistical, and predictive analysis in field-based breeding and precision agriculture. AutoPlotlytics was developed as a Python- and Streamlit-based proof-of-concept platform to reduce software switching and make UAV-based data processing more accessible for users with limited coding or GIS experience. The platform included modules for image and raster processing, vector processing and boundary creation, zonal analysis, exploratory data analysis, statistics, modeling and prediction, and data simulation. A proof-of-concept workflow demonstrated that the platform could support project setup, image enhancement, region-of-interest creation, grid generation, vegetation-index calculation, zonal-statistics extraction, data cleaning, dataset merging, exploratory analysis, and predictive modeling within the same interface. Overall, this dissertation demonstrated that UAV-based high-throughput phenotyping can become more efficient, reproducible, and decision-oriented when RTK-GNSS-based plot boundary creation, automated data extraction, machine-learning prediction, and user-friendly software interfaces are connected. The developed workflows provide a foundation for future cloud-native, database-connected, collaborative platforms that can support spatial-temporal phenotyping, genomic and phenotypic databases, automated reporting, management-zone creation, and advanced breeding and precision agriculture decision support.