A support vector machine and image processing based approach for counting open cotton bolls and estimating lint yield from UAV imagery

Cotton boll count is an important phenotypic trait that aids in a better understanding of the genetic and physiological mechanisms of cotton growth. Several computer vision technologies are available for cotton boll segmentation. However, estimating the number of cotton bolls in a segmented cluster of cotton bolls is a challenging task due to the complex shapes of cotton bolls. This study proposed a combination of spectral-spatial and supervised machine learning based methods for cotton boll candidate recognition and counting from high resolution RGB images obtained from unmanned aerial vehicles (UAVs). An algorithm consisting of machine vision, band-mean filter, Otsu thresholding, red/blue band ratio filter, and geometrical characteristics-based error removal techniques, was employed to detect open cotton boll pixels under several environmental settings. In addition, a support vector machine (SVM) based encoding method was developed using geometric features of cotton boll candidates to predict the number of cotton bolls from the segmented cotton boll candidates. This algorithm was implemented over three experiment sites with three cotton varieties, two tillage practices, seven cover crop treatments, two irrigation regimes (irrigated and rainfed), 26 irrigation levels, and two sensors (DJI FC6310 RGB and MicaSense Rededge) capturing images at two spatial resolutions (0.75 cm and 1.07 cm) over two growing seasons (2019 and 2021). These different experimental settings allowed the proposed approaches to be validated against a variety of complex backgrounds. A visual inspection of 1000 randomly selected pixels revealed that the proposed cotton boll candidate recognition approach was highly effective in segmenting cotton bolls and background pixels, with high classification accuracy (> 95%) and a low number of falsely classified pixels (precision > 0.96; recall > 0.93). A high correlation between ground truth observations and predicted cotton boll count indicated that the use of geometric features of segmented candidates as predictors in association with the SVM model demonstrated a good performance in estimating boll count from recognized cotton boll candidates. Furthermore, linear regression analyses revealed that both boll count and candidate area are potential predictors of lint yield, with boll count being a better predictor than candidate area. Overall, the study demonstrated that machine vision/learning techniques can be potentially used on UAV images to count the number of cotton bolls and predict lint yield over large acreages with reasonable accuracy.