부산대학교 도서관

Yield assessment in broadacre crops is often base on time-consuming and labor-intensive approximations. However, the emergence of unmanned aerial systems (UAS) has allowed for rapid and cost-effective data acquisition. We evaluated red beet plant counts using multispectral UAS data via computer vision and regression analysis. Flight data were captured twice during summer 2019. Our preprocessing steps included i) vegetation detection, ii) feature generation, and iii) feature selection. Partial least squares regression was used as a statistical predictor. Results showed that plant count could be predicted with an acceptable coefficient of determination ($R^{2}=0.76$ for calibration; $R^{2}=0.54$ for cross-validation) and a low root-mean-square-error ($\text{RMSE} =12.27$ plants/plot for calibration, $\text{RMSE} =17.45$ plants/plot for cross-validation). These results are promising, since the error margin, relative to the average density (175 plants/plot), was below 10%. Future efforts should include different geographical locations, higher resolution imagery, and more advanced approaches such as deep learning algorithms with potential for improved accuracy and precision.