This project was completed as a part of the University Scholars Program! The University Scholars Program is a highly competitive research program at the University of Florida where 200 students are selected to work one-on-one with UF faculty on research projects. Learn more about USP and check out my profile here!
Citrus crops play a vital role in the economy by provide humanitarian benefits. Growers face many crop-management challenges, such as the Huanglongbing (HLB) disease and fertilization. Since traditional means of detecting diseases and nutrient content are expensive and labor-intensive, new machine learning (ML) techniques have been applied to assess data in a cheap, fast and reliable workflow. Two datasets were used to generate ML classification models for disease detection (HLB) and regression models to estimate tree nutrient contents (nitrogen (N), phosphorus (P), potassium (K)). The first dataset contains the canopy reflectance of healthy and diseased citrus trees captured with a ground-based spectroradiometer, including wavelengths ranging from 450nm to 1300nm. Classification models of logistic regression (LR), support vector machines (SVM), and linear discriminant analysis (LDA) were generated for the first dataset. The second dataset includes spectral data from tree canopies collected with a multispectral camera on an Unmanned Aerial Vehicle (UAV) in 5 different color bands. Ridge, lasso, and elastic net regression models were generated for this dataset. Evaluation metrics were used to compare the models, such as accuracy for classification problems and mean absolute percentage error (MAPE) for regression problems. The best classification model was LDA with a 98.04% accuracy. Depending on the nutrient, ridge and elastic net both yielded small errors. For N, the MAPE was 7.16% (ridge), and for P and K the MAPEs were 7.19% and 17.02% respectively (elastic net). Overall, these findings provide an approach for assisting growers and researchers in agriculture.
Classification and Regression Machine Learning Models
