Fast Classification of Large Germinated Fields Via High-Resolution UAV Imagery

Abstract

Crop breeding consists of the process of editing crop genetic profile for increasing many crop qualities. In order to achieve optimal results, crop breeders have to plant thousands of plants and keep a track of their growth almost daily. This process requires increased man-hour inspection over large fields, which results in poor accuracy due to human fatigue and a time-inefficient strategy. In this letter, two machine vision approaches were compared for classifying three crop germination classes (good, average, and bad). A naive approach using a classical segmentation and an unsupervised learning approach using k-means segmentation were compared within a high-resolution unmanned aerial vehicles imagery dataset. Experimental results demonstrate the classification of germinated patches up to 0.05 $\text{m}^{2}/\text{patch}$ of resolution with a minimum F1-score of $\text{76}\%$ and $\text{80}\%$ , and AUC of $\text{95}\%$ and $\text{91}\%$ for high and low spatial image resolutions, respectively.