1. Introduction
Rice is one of the major global food crops and feeds over 65% of Chinese people [
1]. One of the basic questions impeding the growth of crops concerns the competition of rice plants from weeds in farmland. Weeds in rice fields compete with rice for water, nutrients, and sunlight, resulting in a detrimental impact on rice yield and quality if not properly controlled [
2].
Different operations have been attempted to control weeds, with chemical and mechanical weeding being widely used in rice fields nowadays. Conventional chemical weeding sprays herbicides uniformly to cover the total field, regardless of the presence of weeds or not, resulting in high herbicide costs. Overuse of herbicides in agriculture causes catastrophic environmental pollution problems, especially in China [
3]. Another widely adopted weeding method is mechanical weeding, which is much more efficient but generally unsatisfactory in terms of weed control performance, causing differences in the bending of rice rows, leading to contact between weeding hoes and rice plants and potentially causing rice plant damage [
4,
5,
6].
In this case, precise identification of rice plants is conducive to control weed growth, because it provides necessary information for subsequent decision-making and implementation procedures. Crop safety can be enhanced by adjusting the working path of weeding hoes or reducing the amount of herbicide applied through targeted spraying.
Many different sensing methods were attempted, within the critical period for weed control, an overall classification accuracy of 87 ± 5.57% was achieved for >5% vegetation coverage in a wheat field by field spectroscopy [
7]. In the laboratory, under controlled illumination conditions, the hyperspectral imaging system provided high-quality images and high-accuracy differentiation of glyphosate-resistant (GR) weeds from glyphosate-sensitive (GS) weeds with accuracies from 75% to 95% [
8]. In addition, visible and near infrared (Vis-NIR) spectroscopy [
9], fluorescence [
10], and distance sensing techniques (light detection and ranging-LiDAR and ultrasonic sensing) were also used for crop detection [
11,
12]. However, these methods are a non-real-time process and have poor anti-interference abilities in complex field environments.
Optical imaging using the machine vision technique is a very promising tool for precision farming and was investigated extensively for crop detection [
13,
14]. Using semi-supervised machine learning to recognize crops and weeds, Søgaard and Olsen [
15] calculated the crop’s center of gravity in the horizontal direction to extract the navigation line. The critical procedure for precise rice detection is digital image processing, through which rice can be segmented and extracted from the acquired images. The recognition rate of the vision method for rice and weeds is only 85% [
16], which is lower than the research method proposed in this paper. Image processing performance is dominantly influenced by complex paddy backgrounds, such as cyanobacteria or green algae, variable lighting conditions in the field, occlusion or overlapping of rice and weed leaves, different growth stages of plants, etc. [
17,
18,
19].
In order to improve the anti-interference ability, several recent studies concerning crop classification by machine learning were designed. Cheng and Matson [
20] used multiple machine learning algorithms such as decision tree, support vector machine (SVM) and neural network in rice and weed discrimination using images downloaded from the internet, achieving a best result of 98.2% precision. Hung et al. [
21] classified three weed species using sparse autoencoders with precision scores of 72.2%, 92.9%, and 94.3% for each species. However, these methods using machine learning were of weaker performance during occlusion or overlapping of rice and weed leaves. Therefore, machine vision is more accurate in dry field crop recognition, but many interference factors exist when identifying rice plants or weeds in paddy field environments, therefore, the recognition process is not ideal.
Tactile methods provide an excellent solution to this problem because they do not depend on sunlight, background, and overlapping leaves. Wide varieties of tactile sensing technologies have been attempted, including optical [
22], resistive [
23], capacitive [
24], piezoelectric [
25], magnetic [
26], and surface acoustic waves [
27], among others, allowing the recognition of an object by gaining information such as contact shape, surface texture, roughness, and slippage detection [
28,
29,
30,
31,
32,
33], enabling tactile sensing to be of huge potential in areas of industry. Conversely, crop recognition based on tactile methods are used less in agricultural fields, unless there are significant characteristic differences between crops, such as physiological height or bending resistance. In the previous study, our laboratory team developed a rice recognition method based on a tactile method by using a bending sensor [
34]. According to the difference of mechanical threshold between rice and weeds, the method could realize rice recognition in different water layer thickness and different rice varieties. However, rice recognition was highly dependent on the accuracy of mechanical threshold setting, which varied with the transplanting days, so the recognition rate was not stable. Xu Liming et al. [
35] designed an auto-obstacle avoidance mechanism based on tactile perception for intra-row mechanical weeding between grape plants, with contact pressure detected by this machine when the contact rod was blocked by a grapevine. The automatic obstacle avoidance mechanism started to work when the contact pressure reached the threshold set by the control system. According to the differences in height and force between corn and weeds in the middle ploughing period, Jia Honglei et al. [
36] designed a flexible shaft type tactile sensor to identify and locate corn crops by setting a reasonable contact position and contact force threshold. However, these tactile methods are unfit for identifying rice plants or weeds because the differences in height and contact force between rice plants and weeds are not as obvious as the stems of grapevines and corn. In addition, the mechanical recognition threshold of rice plants for different rice growing periods were constantly changing and were shown to not be significant. Further, the tactile feedback signal is complex and multidimensional, and cannot be directly used to obtain recognition information by setting an artificial fixed threshold.
Therefore, a rice plant recognition sensor which depended on pressure changes from a flexible gasbag when touching with rice plant or weeds based on machine tactile methods was proposed in this paper. To improve the recognition rate and anti-interference ability of the tactile sensor, it was used in the period when the differences in structure and mechanics between weeds and rice were significant (16–21 days after rice plants transplanting). During operation, the sensor moved along with the machine and its height was adjusted to ensure that the gasbag touched the rice stem and the canopy of weed. The tactile signals were obtained by sensor for deep mining. After conversion of low-level raw data into high-level information (feature extraction), the machine recognition of rice plant in weeds was realized by a back propagation (BP) neural network. In this paper, the main purpose of this study was to identify rice plants among weeds using a tactile method. One key technology involved the acquisition of subtle data with identifying features for rice plants and weeds. Another was the effective classification of the extracted sensing data.
Our contributions are mainly as follows.
A novel sensing method for identifying rice plants and weeds was proposed to address the poor effect of the vision recognition method in rice field, which was different from the previous tactile perception method based on artificial threshold recognition. This study was based on the features of tactile perception data of rice plants and weeds.
A flexible tactile sensor was designed. The gasbag structure of the cantilever beam type showed good adaptability and barometric sensitivity, which was conducive to obtaining differences in structure and mechanics between rice plants and weeds and provided a basis for the depth mining of tactile identification data of rice plants and weeds.
A classification method of rice plants and weeds was proposed, including feature extraction with dimension, dimensionless, and fractal dimension, feature selection with genetic algorithm, and feature classification with neural network. To some extent, this improved the accuracy of identification of rice plants and weeds.
4. Discussion
According to the results of case I, the recognition rate of rice plants was high, at 90.67%. However, it was difficult to produce slipping characteristics when rice plants were touched using the extreme end of gasbag. Therefore, the recognition rate was influenced due to less obvious slipping characteristics.
According to the results of case II, the recognition rate of rice plants was as high as 98%, representing the highest rate among the three cases (
Figure 8) and meaning that the recognition of rice plants using this method was reasonably good. The high recognition rate was due to the obvious slipping characteristics produced by the middle part of gasbag touching the rice plants.
According to the results of case III, the recognition rate of rice plants was slightly lower than in case II, at 96%. There were two situations that caused recognition error related to the process of the gasbag touching the rice plants. The first involves the extreme root of the gasbag touching the rice plants, and the second is the gasbag touching weak rice plants (with a tiller number of less than three). In these situations, the bending stiffness of the root of gasbag was large due to the cantilever structure with no obvious deformation of the gasbag, resulting in a lack of obvious slipping characteristics. Therefore, recognition error occurred.
In summary, the average recognition rate of the experiment was 94.89%, which was relatively high due to the obvious mechanical identification properties of rice plants with long transplantation periods; the recognition rate was relatively high when there were obvious slipping characteristics of mutual friction (i.e., the middle of the gasbag touched the rice plants). Therefore, parameters such as the length of the gasbag should be optimized according to the rice planting pattern. In this paper, the three categories of features selected for neural network are commonly used as signal analysis features [
36,
37,
38]. The results showed that the accuracy of the proposed sensor for rice recognition was satisfactory. However, the sensor needed a certain sliding friction with the rice plant to achieve more accurate identification. Therefore, the recognition had a short delay which should be considered in the operation of guiding weeding parts or herbicide sprinklers to avoid rice plants. In addition, the sensor was mainly used for the recognition of rice plants during the period of weed control, and the recognition error was allowed to be increased in the late stage of rice growth when the mechanical and height differences between rice plants and weeds were not obvious. It should be mentioned that, in the process of feature selection using genetic algorithm, some features may not be selected, which reduces the accuracy of selecting the best features to some extent. Therefore, more features selection methods need to be carried out to optimize features. More recognition models such as decision tree and support vector machine should be compared, and the best recognition model can be used by comparing the recognition results. In future studies, the parameters that can better reflect the tactile signal characteristics should be introduced to further improve the recognition accuracy of sensor. In addition, appropriate transplanting periods should be selected according to the agronomy of rice plant growth (bending strength of the stem). The sensor should be waterproofed in the future to facilitate field measurement.
5. Conclusions
In this study, a rice plant recognition sensor was developed using a tactile method and machine learning algorithm. Tactile information was acquired from voltage signals of an air-pressure sensor in a gasbag which touched rice plants. During data processing, three algorithms were used to extract 13 features of tactile voltage signals, and an optimum set of features (variance, kurtosis, waveform factor, box dimension, and hurst exponent) was selected using a genetic algorithm. A rice plant and weed classifier was built using a BP neural network. The rice recognition rates for the three testing sets were 95.3%, 95.1%, and 94.9%.
Based on the proposed classifier, an experiment with three case was designed according to the different positions of the gasbag touching the rice plants and weeds. The best recognition performance was achieved by the middle of gasbag touching the rice plants, with the recognition rate being as high as 98%. The second-best recognition performance was achieved by the root of gasbag touching the rice plants, at 96%. When the end of the gasbag touched the rice plants, the recognition rate was the lowest that was observed in the experiment, at 90.67%. The dataset in this paper were obtained from a single rice variety, so the data of the corresponding varieties need to be obtained to train the classifier for the recognition of other rice varieties. The experiment proved that tactile-based recognition of rice plants is a promising method.