Multiclass Classification on Soybean and Weed Species Using a Customized Greenhouse Robotic and Hyperspectral Combination System

Abstract

HighlightsA novel combination of robotic and hyperspectral camera systems was customized to collect data in a greenhouse environment.Five popular weed species and crop (soybean) data were collected and analyzed in this research.Seven different types of hyperspectral data preprocessing methods were tested to obtain the best performance of the model. Abstract. Soybean production is greatly affected by different types of weeds such as horseweed, kochia, ragweed, redroot pigweed, and waterhemp in the midwestern region of the United States. Identification of the soybean plants and the weeds is crucial to controlling the weeds in precision agriculture. The objective of this study was to classify soybean plants and five weed species where a hyperspectral imaging camera with a spectral range of 400 to 1000 nm was used to acquire the images. To acquire the HSI images, a customized robotic hyperspectral data collection scanning platform was developed and used in the greenhouse. A total of 983 hyperspectral data cubes were captured from the greenhouse environment (n = 252, soybean; n = 731, weeds). Spectral information was extracted from the collected images and then a classification model was developed by applying partial least squares discriminant analysis (PLS-DA). To construct the calibration and validation data set, the images were each divided into 70% and 30% ratios for model training and testing, respectively. Seven types of data preprocessing techniques, including mean normalization, maximum normalization, range normalization, multiplicative scatter correction (MSC), standard normal variate (SNV), Savitzky–Golay first derivatives, and Savitzky–Golay second derivatives, were explored individually and accepted as the best preprocessing method based on the highest performance in calibration and validation results. The results showed maximum validation model performance was found at 86.2% by applying the Savitzky–Golay second derivatives preprocessing method for multiclass and 89.4% for binary class. The most important wavelength information was evaluated from the beta coefficient developed using the same preprocessing method. Finally, chemical images were generated using a best-performer model to identify the soybean plants from weeds. The generated images showed a significant difference in chemical composition between soybean and weed plants at 443 nm, 633 nm, 743 nm, and 968 nm. The correlation between these peaks and the chemical components of the plants involves a-carotenoid, chlorophyll, and moisture, respectively. This study shows promising results for the application of HSI in weed control systems for soybeans and relevant weed identification in precision agriculture applications. Keywords: Greenhouse robotic system, Hyperspectral imaging, Precision agriculture, Weed identification.