Original papersMaize and weed classification using color indices with support vector data description in outdoor fields
Introduction
Expert systems are widely applied in agricultural studies, of which plant recognition has been increasingly applied in automatic control systems to meet the demand of automatic supervision for large areas with crops. Thus, expert systems for crop-plant recognition have become highly important to hasten and facilitate agricultural tasks.
Maize production worldwide is approximately 875 Mt (Kurtulmuş and Kavdir, 2014), and maize can be planted in any climate. Maize is agriculturally important because of its nutritional value and high consumption by humans. Excessively high or small number of maize planted per unit area will reduce agricultural production. Moreover, the maximum production of maize depends on the nutritional supply of the fields where maize grows. Several studies have shown that plant density is related to the yield in fields (Dai et al., 2015, Yang et al., 2014). Under the same fertilizer level, rational close planting and increasing the basic seedlings contribute to promote yield.
Weeds are spread in cropland in every growing period of maize, and they compete with maize for water and nutrient (Westerman et al., 2008). Weeds should be removed throughout crop growth, especially in maize seedling, to reduce loss. Using herbicides is the most common method for weeds removal because of its high efficiency in controlling weeds (Ali et al., 2014). However, blanket spraying of herbicides will leave pesticide residues and lead to environmental pollution. The proper application of herbicides should be studied to reduce the use of herbicides and protect the field environment. Thus, plant recognition is important in crop fields to locate plant more accurately and calculate the coverage of crops more precisely. In traditional methods, observing maize fields and removing weeds manually is time-consuming and vulnerable to mistakes. These limitations can be avoided using machine visual technology to detect maize in fields.
Numerous studies have applied in-field computer vision methods to recognize plants. Shape features can be used after background separation or a combination of color and shape features. Golzarian and Frick (2011) presented an approach to classify images of wheat, ryegrass and brome grass species using color, texture and shape features. These features were reduced to three descriptors by principal component analysis (PCA), and the results show accuracy of 88% and 85% in the differentiation of ryegrass and brome grass from wheat, respectively. Sujaritha et al. (2017) designed a classification system to extract leaf textures and employ a fuzzy real-time classification technique, and the system detected weeds with 92.9% accuracy over a processing time of 0.02 s. Torres-Sospedra and Nebot(2014) used a new training procedure with noisy patterns for ensembles of neural networks to detect weeds. Bansal et al. (2013) used fast Fourier transform leakage values to detect green citrus. An agricultural mobile robot for mechanical weed control was equipped with two vision systems, namely, one gray-level vision that can recognize the row structure formed by crops, and the other is color-based that can identify a single crop among weed plants (Åstrand and Baerveldt, 2002). Kurtulmuş and Kavdir (2014) proposed a computer vision algorithm to detect cutting locations of corn tassels in natural outdoor maize canopy. In this process, conventional color images and computer vision were used with a minimum number of false positives. Color (vegetable) indices, shape, and texture features were used to reduce false positives, reaching 81.6% correct detection rate for the test set. However, plant detection based on shape features could become quite complex because of variations in leaf shapes of diverse species at different growth stages. Moreover, in actual field conditions, the following complex cases may occur: (1) leaf overlap; (2) leaf orientation variation; (3) shooting angle variation, causing changes in the shape of leaves; or (4) leaf movement along with the wind, blurring boundaries in images. Most of these factors may also affect textural analysis and cause greater difficulties in analysis.
Color vegetation analysis is probably the most efficient method for machine vision-based location of crops and weeds. El-Faki et al. (2000) established a simple weed-detection method using a color machine-vision system, and four types of relative color indices formed by RGB gray levels were designed. The most effective combinations of these color indices were selected, and the final classification accuracies for wheat and soybean were 54.9% and 62.2%, respectively. Vegetation indices were used in image segmentation, which was oriented to detect crop rows in images from maize fields with high weed concentration (Montalvo et al., 2012). Raw RGB channels and extracted vegetation indices were used primarily to detect vegetation against background. Kazmi et al. (2015) used 14 vegetation indices to detect creeping thistle in sugar beet fields.
The proposed algorithm in this study was developed to detect maize in natural outdoor fields using conventional color pictures and computer vision with a minimum number of false positives in a real-time monitoring system. Conventional image segmentation methods based on RGB values can separate maize from background. However, these methods cannot separate maize from weeds because of the high similarities in their image RGB values. We discovered that the color indices of these plants differ in contrast. Thus, color indices were used to distinguish maize and weeds in our study. According to the characteristic of real-time supervision system, the samples in the training set of this algorithm were collected only one day during the early growth stage of maize in 2011, and the training model was used to recognize the samples after several days. The test set consisted of images for 12 days in the same year, and images in following two years were tested to verify the stability of the algorithm. As a one-class classifier, support vector data description (SVDD) was used in our task (Tax and Duin, 1999), because weed species vary in the field and the features of all kinds of weeds are difficult to be extracted. The main objective of this study was to develop and test classification models based on color indices for identifying maize and weeds using color digital images.
Section snippets
Image acquisition
RGB images were taken in a maize field located in Gucheng, Hebei province, China, using a digital camera (E450 Olympus) with a resolution of 3648 ∗ 2736 pixels and focal length of 16 mm. The camera within the protection cover was fixed on a bracket at height of 6 m from the ground. The pitch and roll angles of camera were 30° and 0°, respectively. The images used in this study were captured in July 9–July 21 in 2011, July 1–July 16 in 2012, and July 13–July 27 in 2013. The image was collected one
Image segmentation using the ExG index
The images acquired at different illumination condition and weather (including sunny, cloudy and rainy days) were shown in Fig. 2. The original RGB images contained the complex background information including soil, wheat straw, stone as shown in Fig. 2(a), (d) and (g)). After segmentation using the ExG index, the images removed the background information and the maize/weed were preserved. For the blurred images (Fig. 2(g)) resulting from the rain, the ExG index still given the satisfactory
Conclusion
In this study, we proposed a practical and efficient method to solve maize detection with SVDD and vegetation indices. This process is more practical than texture or shape features, because the color-based system has lower sensitivity to canopy overlap, leaf orientation, camera focusing, and wind effect. Only the maize features were selected as the training set can improve the efficiency and decrease the complicacy. The results show that the accuracy reached 96.74% in the early stage and the
Acknowledgment
Dr. Qibing Zhu, Dr. Min Huang and Yang Zheng gratefully acknowledge the financial support from the National Natural Science Foundation of China (Grant no. 61271384, 61275155), the Fundamental Research Funds for the Central Universities (JUSRP51730A), New Agricultural Engineering of Jiangsu Province Agricultural Commission (SXGC 2016-106), the Prospective Joint Research Foundation of Jiangsu Province of China (BY2016022-32), the 111 Project (B12018), and sponsored by Qing Lan Project.
References (21)
- et al.
Manipulation of dry matter accumulation and partitioning with plant density in relation to yield stability of cotton under intensive management
Field Crops Res.
(2015) - et al.
Detecting creeping thistle in sugar beet fields using vegetation indices
Comput. Electron. Agric.
(2015) - et al.
Detecting corn tassels using computer vision and support vector machines
Expert Syst. Appl.
(2014) - et al.
Verification of color vegetation indices for automated crop imaging applications
Comput. Electron. Agric.
(2008) - et al.
Automatic detection of crop rows in maize fields with high weeds pressure
Expert Syst. Appl.
(2012) - et al.
Weed detecting robot in sugarcane fields using fuzzy real time classifier
Comput. Electron. Agric.
(2017) - et al.
Support vector domain description
Pattern Recogn. Lett.
(1999) - et al.
Two-stage procedure based on smoothed ensembles of neural networks applied to weed detection in orange groves
Biosys. Eng.
(2014) - et al.
A detection method of weed in wheat field on machine vision
Proc. Eng.
(2011) - et al.
Effects of plants density on yield and canopy micro environment in hybrid cotton
J. Integrative Agric.
(2014)
Cited by (93)
Machine learning and handcrafted image processing methods for classifying common weeds in corn field
2023, Smart Agricultural TechnologyInstance segmentation method for weed detection using UAV imagery in soybean fields
2023, Computers and Electronics in AgricultureAttention-aided semantic segmentation network for weed identification in pineapple field
2023, Computers and Electronics in AgricultureA review of machine learning techniques for identifying weeds in corn
2023, Smart Agricultural TechnologyPerformance evaluation of deep learning object detectors for weed detection for cotton
2023, Smart Agricultural Technology