A Tactile Method for Rice Plant Recognition Based on Machine Learning

Abstract

Accurate and real-time recognition of rice plants is the premise underlying the implementation of precise weed control. However, achieving desired results in paddy fields using the traditional visual method is difficult due to the occlusion of rice leaves and the interference of weeds. The objective of this study was to develop a novel rice plant recognition sensor based on a tactile method which acquires tactile information through physical touch. The tactile sensor would be mounted on the paddy field weeder to provide identification information for the actuator. First, a flexible gasbag filled with air was developed, where vibration features produced by tactile and sliding feedback were acquired when this apparatus touched rice plants or weeds, allowing the subtle vibration data with identification features to be reflected through the voltage value of an air-pressured sensor mounted inside the gasbag. Second, voltage data were preprocessed by three algorithms to optimize recognition features, including dimensional feature, dimensionless feature, and fractal dimension. The three types of features were used to train and test a neural network classifier. To maximize classification accuracy, an optimum set of features (b (variance), f (kurtosis), h (waveform factor), l (box dimension), and m (Hurst exponent)) were selected using a genetic algorithm. Finally, the feature-optimized classifier was trained, and the actual performances of the sensor at different contact positions were tested. Experimental results showed that the recognition rates of the end, middle, and root of the sensor were 90.67%, 98%, and 96% respectively. A tactile-based method with intelligence could produce high accuracy for rice plant recognition, as demonstrated in this study.

Accurate recognition of rice plants based on visual and tactile sensing

“…1(a) shows an image of the data acquisition process. The middle of the device was selected to contact the plants based on the results of a study by Chen et al 8 The output strength of the PVDF sensor and the graphic varied with the degree of flexure of the flexible strip when the device touched a plant. The image acquired by the camera (Hikvision, Hangzhou, China) and the output signal of the PVDF sensor acquired by an Arduino Board (Longzhan Information Technology Co. Ltd, Shanghai, China) were transmitted to a computer via a wireless transmission module.…”

“…1(a) shows an image of the data acquisition process. The middle of the device was selected to contact the plants based on the results of a study by Chen et al 8 . The output strength of the PVDF sensor and the graphic varied with the degree of flexure of the flexible strip when the device touched a plant.…”

“…[1][2][3][4][5][6][7] However, unlike dry-field environments, the rice-growing environment is very specific, with complex backgrounds, aquatic plants and visual disturbance conditions. 8,9 Current mainstream visual perception methods do not meet the recognition requirements for rice plants, making it difficult to apply and promote mature intelligent technologies and equipment in paddy field environments. 10 Therefore, accurate and reliable recognition of rice plants in paddy field environments is of great importance.…”

“…Accurate crop recognition is a prerequisite for intelligent agricultural machine operations, such as crop navigation, automatic row alignment of weeding components and precise drug application 1‐7 . However, unlike dry‐field environments, the rice‐growing environment is very specific, with complex backgrounds, aquatic plants and visual disturbance conditions 8,9 . Current mainstream visual perception methods do not meet the recognition requirements for rice plants, making it difficult to apply and promote mature intelligent technologies and equipment in paddy field environments 10 .…”

“…Therefore, Chen et al created an airbag-type recognition perceptron based on data features and obtained multi-touch data of the recognition course through an air pressure sensor inside the bag for machine learning classification; the accuracy of riceweed recognition reached over 90% in the laboratory environment. 8 In addition, in order to reduce weed interference, many scholars have assessed weed density. Chen et al used a flexible gasbag with a surface filled with special microstructures to assess weed density with an average accuracy of 91.7%.…”