Establishing a model to predict the single boll weight of cotton in northern Xinjiang by using high resolution UAV remote sensing data

Abstract

Single boll weight is the main factor of cotton yield and a key index used to evaluate the quality of cotton. Predicting the single boll weight in a large area is important for variety selection and yield improvement. A model was established to predict the single boll weight by using the multitemporal high-resolution visible light remote sensing data obtained from UAV. Specifically, remote sensing data were collected for 29 fields in the Changji, Shihezi and Shawan areas in northern Xinjiang during the blooming period and boll opening stage. Five circular areas with a radius of 1 m were selected from each field as the ground investigation area for collection of the cotton boll samples. Fully convolutional networks (FCN) was used to recognize and extract bolls at the boll opening stage in the remote sensing images as a dependent variable of the model. Correlation analysis was carried out by combining VDVI (Visible-band difference vegetation index) at the flowering and boll setting stages, VDVI at boll opening stages, VDVI at boll opening areas (Extracted by FCN) and RGB mean values, then use the least squares linear regression and BP neural networks to model the upper, middle, lower cotton layers and average single boll weight in investigation area. Subsequently, K-fold cross-validation was performed to evaluate the results. The results showed that the results of the least squares linear regression (R2 = 0.8162) and BP neural networks (R2 = 0.8170) were nearly equivalent. The percentage of boll opening in the area and VDVI at the flowering and boll setting stages were highly correlated with upper single boll weight. This study proposes a method to realize the large scale prediction of single boll weight, which provided a new idea for cotton yield prediction and breeding screening.