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

Highlights

• A single boll weight prediction model for cotton was established.

• Using FCN to extract cotton bolls from images can improve the accuracy of prediction.

• The model can make large-scale predictions of cotton single boll weight.

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.

Introduction

China has the largest cotton production area worldwide, and cotton planting area accounts for 30% of the total cultivated land area. The total cotton production area in China was 6,841,593 ha in 2019 (National bureau of statistics, 2019). Xinjiang is the main production area in China, with the corresponding cotton production area and output accounting for more than 50% and approximately 70% of the national cotton production area and total cotton production, respectively (Yu et al., 2019). The single boll weight is the weight of the seed cotton in a single cotton boll, and it exhibits complex characteristics. Moreover, this weight is positively correlated with the burr weight, number of chambers per boll, seed weight, number of seeds, and fiber weight, and negatively correlated with the number of bolls per plant and lint percentage (Karademir et al. 2009). The clear correlation between the single boll weight and yield is has been verified experimentally (Batool et al. 2010). Nevertheless, if the cotton boll is extremely large, the shell is expected to be thick, leading to an extended growth stage and decreased rates of cotton boll filling and boll opening. Generally, in the process of selection and breeding of cotton in China, the suitable weight of a single boll is 5.0–6.5 g (Yu et al., 2016). The efficient prediction of the single boll weight in a large area can help improve the efficiency of cotton variety breeding, thereby improving the cotton yield and quality.

With the rapid development of remote sensing technology, it is being widely used in agricultural management and agricultural monitoring (He and Mostovoy, 2019, Liu et al., 2019, Wei et al., 2019, Weissteiner et al., 2019). As a type of low altitude remote sensing system, UAVs are being widely used owing to their convenience, flexibility, low cost and capacity for high-resolution imaging (Xu et al., 2019, Campos et al., 2019, Yan et al., 2019). Several researchers have attempted to study cotton by using remote sensing systems based on a UAV platform. For example, a UAV installed with a multispectral camera was used to collect data and calculate the NDVI to predict the final yield of cotton. However, single-parameter modeling is not sufficiently accurate (R² = 0.47) (Huang et al., 2013), and the introduction of additional dependent variables to the prediction model can help improve the prediction accuracy. By installing a high-resolution digital camera in the UAV to collect RGB images and introducing parameters such as color characteristics and vegetation coverage into the model, a higher prediction accuracy (R² = 0.97) can be obtained (Zhang et al., 2019a, Zhang et al., 2019b). Moreover, a LiDAR sensor can be installed on the UAV to collect quantitative information regarding the crop canopy and generate a digital surface model. These aspects can facilitate the quantitative modeling of biomass (Liu et al., 2019, Simpson et al., 2016, Wang et al., 2019). Moreover, the canopy height and surface model of the crop canopy can be obtained by using an aerial triangulation algorithm in combination with a high-resolution digital camera. Moreover, because of the positive correlation between plant height and biomass, the canopy digital surface model can also be used to realize the inversion of the plant height (Lati et al., 2013). The information provided by various information sources can be extended by using multisource information fusion technology to fuse the remote sensing data pertaining to different time phases. This aspect can considerably improve the effectiveness of feature extraction and classification (Zhang et al., 2019a, Zhang et al., 2019b), and it plays a very important role in the study of farmland classification, vegetation coverage, land desertification and crop growth, which change over time (Huang et al., 2017, Yi et al., 2019, Yu et al., 2019).

The cotton boll in the remote sensing images is extracted through machine vision, which is of great help in studying the rate of cotton boll opening, defoliant effects and to direct intelligent agricultural machinery (Li et al.,2015). In the past decade, researchers have used machine vision technology to extract cotton boll development data. For example, the cotton boll and its background have been segmented by using random forest classification (Zhang et al.,2011) and an object-oriented method based on near-infrared imaging (Li et al.,2016). The rapid improvement in the operation ability of computers has facilitated the advancement of deep learning technology. At present, convolutional neural networks represent the branch of deep learning with the most research and the best development, and they have been applied to yield excellent results (Rawat and Wang 2017). The fully convolutional network (FCN) used in this study was improved using convolutional neural networks (CNN). The convolutional layer was used to replace the last fully connected layer of the CNN. Subsequently, the deconvolutional layer was used to upsample the featured heat map to restore its size to that of the input image to classify the image at the pixel level and retain the spatial information in the input image (Long et al., 2015, Shelhamer et al., 2017). Some researchers attempted to realize cotton extraction using fully convolutional networks (FCN), which is the foundation of the present study (Li et al., 2017).

In this paper, multisource information fusion technology was used to fuse UAV remote sensing data from different time phases at the pixel level. The fully convolutional neural network was used to achieve pixel-level semantic segmentation of the remote sensing images, and the cotton segments in the images were extracted to eliminate the influence of the soil pixels on the model accuracy. The boll opening area in the images was extracted, and analyzed the correlation among the single boll weight of the upper, middle and lower layers of the cotton plant with VDVI and other parameters. VDVI (Du and Noguchi, 2017, Wang et al., 2015a) is a vegetation index used in the case of visible light remote sensing images that can distinguish vegetation from nonvegetation objects without using the near infrared (NIR) domain. Moreover, the correlation coefficients between different parameters and the single boll weight were determined. The objective was to establish a method that could enable the large scale prediction of single boll weight, in order to provide a novel opportunity for cotton yield estimation and variety selection.