Abstract
Crop and weed identification remains a challenge for unmanned weed control. Due to the small range between the chopping tine and the important crop location, weed identification against the annual crops must be extremely exact. This study endeavor included a literature evaluation, which included the most important 50 research publications in IEEE, Science Direct, and Springer journals. From 2012 until 2022, all of these papers are gathered. In fact, the diagnosis steps include: preprocessing, feature extraction, and crop/weed classification. This research analyzes the 50 research articles in several aspects, such as the dataset used for evaluations, different strategies used for pre-processing, feature extraction, and classification to get a clear picture of them. Furthermore, each work’s high performance in accuracy, sensitivity, and precision is demonstrated. Furthermore, the present hurdles in crop and weed identification are described, which serve as a benchmark for upcoming researchers.
Nomenclature
Artificial Intelligence Bag-of-Features Back-Propagation Neural Network conditional Generative Adversarial Network Covariance Matrix Adaptation Evolution Strategy Convolutional Neural Network determinant of the Hessian detector Deep Neural Network difference of Gaussian in multi-scale regions Fully Convolutional Networks Systems Genetic Algorithm Graph Convolution Network Gray Level Co Occurrence Matrix Gaussian Mixture Model Graphic Processing Unit Histogram of oriented Gradients Hue, Saturation, Value Internet of Things K Means Clustering K-Nearest Neighbor Local Binary Pattern Linear Discriminant Analysis Local Weighted Maximum Discriminant Projections Multi Wavelength Laser line Profile Normalized Difference Vegetation Index Near-Infrared Reflectance Neural Network Object Based Image Analysis Principal Component Analysis Pyramid Histogram Of visual Words Partial Least Squares Discriminant Analysis pyramid scene parsing network Pulse Width Modulation Quadratic Discriminant Analysis Question Interpretation Space Radial Basis Function Region based Convolutional Neural Network Random Forest Red Green Blue Region of Interest Ratio Vegetation Index Shape Based Weed Detection Semantic segmentation model Stem Emerging Points Scale-Invariant Feature Transform Simple Linear Iterative Clustering Spatial Pyramid Matching Site Specific Weed Management Support Vector Machine TWo-way INdicator SPecies ANalysis Unmanned Aerial Vehicle Universal Function Approximation Block Visual Geometry Group Wireless Sensor Network You only look once v3
Introduction
Cultivating the land, crop cultivation, and agriculture are the terms used to describe animal farming. It comprises both the preparation and distribution of plant and animal products for human use. Agriculture is the major factor of the world’s food and textiles. It is the primary raw material supplier. Agriculture provides several raw resources, such as cotton, sugarcane, timber, and soybean oil. These substances are vital to big businesses in the circumstances that several people are unaware of pharmaceutical manufacture, fuel oil production, polymer production, and so on. Gardening, farm machinery, herbicides, insecticides, storage, supply chain, food production, dairy trade, plant breeding, pisciculture, agro-based, grains, aquaculture, fowl, cattle breeding, pet food, and plant production are among the 17 important sectors in India’s agrarian economy [19,29].
Weed management [45] is the difficult work done in the agriculture to overcome or to increase crop output. The development of an autonomous weed management system that improves cultivation accuracy while reducing the need for manual labor is a major problem. Weed killer, resistance of fresh pesticide filers, removal of high density polyethylene and expensive biodegradable plastics make machinery relationship between row cropping and manually picking the best choices in specialized crops like tomatoes. Manual cultivation has become more expensive due to manpower shortages in many richest nations. Shape based characteristics rely on a segmentation algorithm that properly detects the crop segment profile. As a result, the categorization is heavily reliant on the plant boundaries, which have poorly defined pixels. Categorization becomes more vulnerable to variations in picture collections.
Weeds are now liable for 45 percent of crop losses in the agriculture business, owing to grain competition [10,21,51]. This proportion can be reduced using an effective weed identification technology. Weeds are categorized as yearly, biannual, or perpetual based on how long they live. Annual weeds were only survive for a year and sustain in the long run during that year. Crop signaling method is used for differentiating the agricultural crops from weeds to generate a system readable crop plant for weed management system. A chemical is sprayed on agricultural land for planting, and a distinctive light beam is produced. When weed management is necessary later in the season, a weeding robot uses the crop signal to identify and accurately determine the direction of crop plants. Machine learning researchers focus on studying classification difficulties and the methods to solve them. Machine learning approaches are attracting an amount of publicity in a lot of sectors because of their promise for consistency, dependability, and reproducibility among other things. The effectiveness of machine leaning classification algorithms is determined on selecting features and attempts made to avoid over fitting. Weed detection has been done in a variety of ways including hand weeding, motorized, SSWM, and computer vision [27]. Hand weeding or handheld discharge equipment is used for weeding. Handheld weed eradication is done in small fields to identify each weed. Due to its inefficiency and expensive production costs, this approach is no longer viable for large-scale operations or newer weeding procedures.
The major contribution of this survey work is:
This research analyzes the most important 50 research articles from 2012 until 2022 in the field of crop/weed classification.
Different datasets (image, crop and weed, training, soil, broadleaf grass and so on) are manifested from the reviewed papers.
The different techniques utilized for pre-processing, feature extraction and classification are discussed.
Performance evaluation in each paper is examined, and the best findings are also discussed.
Further, current crop and weed identification issues are presented, which serve as a benchmark for future investigators.
The paper is structured as: Section 2 gives a description of literature works. Section 3 shows the analysis part, where different analyses have been portrayed. Section 4 gives the research gaps and challenges, and Section 5 concludes the paper.
Literature review
The selected 50 research articles from three standard journals, IEEE, Science Direct, and Springer are subjected to a compact survey in this study endeavor. All of the articles in this collection date from 2012 to 2022.
In 2012, Faisal Ahmed et al. [3] used a machine learning method called SVM for accurate crop and weed classification in digital photos. The goal is to see if using SVM as a classification algorithm in an integrated weed management system may result in a reasonable classification rate. In order to establish the ideal combination of parameters for the maximum prediction performance, a total of fourteen criteria that define crops and weeds in images were investigated.
In 2012, Alex Krizhevsky et al. [36] built a CNN to categorize the high resolution photos into distinct classes. To quicken up training, non-saturating neurons and an extremely fast GPU version of the convolution function were used. A recently established regularization approach dropout was used to minimize fitting problem in the fully linked layers.
In 2012, M. Montalvo et al. [44] introduced a method for crop row identification in maize ground with severe weed stress in pictures. The vision system is intended to be mounted on the back of a mobile agricultural tractor. The photos were taken under reference material for the future which is influenced by the aforementioned undesirable problems. Image classification, dual thresholding and crop row recognition were the three fundamental procedures in image analysis.
In 2013, I. Herrmann et al. [24] detected the biennial grasses and leafy weeds in wheat fields using ground-level imaging spectroscopic information with spectral and spatial resolutions. Cross-validation of least squares analysis was done using picture pixels. The accuracy of the model was done by differentiating the variances and Cohen’s Kappa scores of the cross-validation confusion matrices.
In 2014, Franc-ois-Michel De Rainville et al. [14] used system vision and morphological findings to present a weed and crop categorization system. The primary morphological traits of weeds found in maize and soybean fields were extracted using supervised and unsupervised learning approaches. Feature extraction procedure was based on spatial placement of vegetation in farms.
In 2014, Sebastian Haug et al. [23] presented a non-segmentation machine vision technique for plant categorization and its use in agriculture. In commercial farms where crop and weed plants grew close together, this technology would distinguish between crop/weed and handle plant overlap. Crop/weed segregation were allowed for weed management tactics that target particular weed treatments to save money and reduce environmental impact.
In 2016, Henrik Skov Midtiby et al. [52] investigated the possibility of applying agricultural seed type information based on simulated outcomes from a vision technology. The weed density and agricultural plant design position variability were demonstrated to affect the accuracy of position-based crop plant detection.
In 2016, Tsampikos Kounalakis et al. [35] introduced a new framework for weed identification that uses well-known picture attributes paired with sophisticated linear image representations. The suggested weed recognition system was based on cutting edge object and image classification approaches that take advantage of sophisticated compression and machine learning algorithms to improve performance. The developed method can be used in a variety of settings including farmlands and weeds.
In 2016, P. Lottes et al. [41] developed a system that uses a Markov random field to conduct plant identification, extraction of features, randomized woodland categorization, and flattening to provide a good prediction of crops and weeds. The system has been tested on a factual farm robot in multiple sugar beet farms demonstrating that the proposed technique was capable of reliably recognizing the weeds on the ground.
In 2016, Francisca Lopez-Granados et al. [39] created an automatic OBIA algorithm that was used on ortho image and multispectral cameras. The weed mappings were taken using a hyper spectral lens at a depth of 30 meters in both regions.
In 2017, Wolfram Strothmann et al. [53] developed a MWLP system that screened the crop plants and gathered sensor information, which produced image based information, perfectly tied spectral absorption, and scattered details for each pixel at several waveforms. It was possible to get misclassification rates that were similar to the best literature values.
In 2017, M. Louargant et al. [42] presented a simulation of the full image acquisition process to calculate the signals of the input image based on a number of criteria in order to replicate in-field acquisition circumstances. Simulations for varied vegetation rates were used to investigate the categorization potential in natural vegetation, as well as monocotyledon and dicotyledon groups. The categorization capacity in topsoil as well as monocotyledon and dicotyledon groupings was investigated by applying simulation to various vegetation principles.
In 2017, Adel Bakhshipour and Abdolabbas Jafari [7] used a combination of form traits to create a pattern for each plant species. SVM and NN were used to allow the vision system to recognize weeds based on their pattern. In sugar beet fields, four types of common weeds were investigated. Fourier descriptors and moment invariant features were among the shape feature sets.
In 2017, Adel Bakhshipour et al. [8] studied wavelet texture traits to see whether they might be used to detect weeds in a sugar beet crop. A selective technique was used to identify the spectral texture qualities for each photo sub division to be sent to a DNN. Co-occurrence wavelets were found for each multi resolution image created by a single level wavelet decomposition. The picture was segmented using the NN to label each part as weed or main crop.
In 2017, Maurilio Di Cicco et al. [15] suggested an innovative and successful method for reducing the amount of human involvement required for classification. The goal was to create huge artificial training datasets by procedurally randomizing important aspects of the specific environment. More precisely, it was feasible to create a huge number of realistic site of an artificial agricultural land with little attempt by adjusting these model and utilizing a few real world textures.
In 2017, Esmael Hamuda et al. [22] suggested a new color-based method based on morphological erosion and dilatation as well as color characteristics. Under natural lighting, this method separates cauliflower crop sections from weeds and dirt in the picture. For plant, weeds and environment discrimination, the suggested algorithm uses HSV color space. Filtering several of the HSV channels between particular values defines the ROI. The moment technique was used to monitor crops and identify the location and mass distribution of items in video frames.
In 2017, JingLei Tang et al. [54] used soybean saplings and the related weeds of theirs as the study subject and built a weed detection method using K means and a CNN. This article used k means unsupervised learning to put back the aimless starting values of typical CNN variables, grouping the benefits of layered and fine-turning of CNN parameters.
In 2018, Philipp Lottes et al. [40] developed a unique crop and weed classification system based on a CNN with an encrypt-decrypt structure that considers picture sequences and adds spatial features. This approach was capable of robustly predicting a pixel-wise categorization of the pictures into crop/weed by utilizing crop organization knowledge visible from the images taken.
In 2018, Adnan Farooq et al. [17] looked at the impact of these differences on weed detection. Weeds and crops have a high degree of spectral resemblance. In this letter, patch-based categorization techniques were used. The approaches of the CNN and the HoG were contrasted and assessed.
In 2018, Petra Bosilj et al. [11] concentrated on pixel-based techniques for crop versus weed classification, particularly in complicated scenarios including uncorrelated plants and blockages. The advantages of attribute profiles were multi scale and content driven morphology-based labels. When combined with morphology based segmentation on a max-tree structure, the suggested classification approach was highly useful since the identical representation may be utilized for the extraction of features.
In 2018, Saad Abouzahir et al. [1] utilized the color indices to construct the graphs to distinguish three classes such as soil, soybean, and broadleaf. BPNN and SVM classifiers were used to test the feature representation. Precision agriculture was a notable standout in terms of using the latest breakthroughs in intelligent systems. Adoption of this system was encouraged by the need to minimize costs, improve healthcare quality and efficiency and hence increase the volume and quality of crops.
In 2018, Junfeng Gao et al. [20] investigated the weed and maize classification potential of an unique hyper spectral photo mosaic lens. An appropriate classifier model was created for image analysis and feature extraction. A maximum of 185 spectral characteristics were created which include reflectance and vegetation index [46] characteristics. The amount of features which were followed in the classification algorithm was determined and the accuracy-oriented feature reduction was done.
In 2018, Tsampikos Kounalakis et al. [34] suggested an efficient image based weed detection system for the Broad-leaved Dock weed management challenge. The suggested weed recognition system was built on a platform that enables the effects of different picture resolutions on classification and tracking accuracy to be investigated. It also incorporates cutting-edge object and image classification techniques such as features collection, codebook knowledge, feature encode, picture presentation, and classification.
In 2018, Wenhao Zhang et al. [60] suggested an alternative approach for identifying broad-leaf weeds in grassland which allows precise weed management with less chemical usage. To obtain high accuracy and reliability, both machine learning and deep learning approaches has been investigated and evaluated. For classifier training and method testing, in-pasture grass and weed picture data were collected.
In 2019, Shanwen Zhang et al. [59] presented a Grabcut and LWMDP algorithm for weed species detection in agriculture fields. Grabcut was used to eliminate the background followed by KMC to partition the weeds from the entire picture. The low-dimensional discriminant characteristics were extracted using LWMDP. To identify weed species, the SVM classifier [45] was used. The test results from the weed picture dataset indicate that the suggested approach is effective for predicting weed species and can meet the criteria for multi row agricultural management using machine vision.
In 2019, Aichen Wang et al. [56] highlighted weed identification progress utilizing ground based computer vision and pixel execution approaches. The four techniques for weed identification including preprocessing, fragmentation, extraction of features, and categorization. Various color criteria and categorization algorithms such as color index, threshold, and learning were developed to distinguish plants from the backdrop. The challenge with weed identification was distinguishing between crops and weeds which frequently have identical characteristics.
In 2020, Rekha Raja et al. [48] demonstrated a unique crop signaling strategy that uses computer vision technology to detect plants through weeds in difficult natural circumstances, such as excessive weed populations seen on organic farms. Crop signaling is cost effective. The signaling compound was sprayed on the crop and allowed the crop to be distinguished from weeds visually.
In 2020, Rekha Raja et al. [49] presented a revolutionary crop signaling idea to make agricultural plants as machine-understandable one by marking them at planting. Weeds were recognized as plants that lacked this crop signaling, and the automated weed blade operator was used to remove them. The SEPs of tomato plants were recognized using a machine vision method that evaluated pictures of crop plants acquired by lens of the camera with the help of an imaging chamber. To eradicate weeds, they were delivered to the robotic weed cutter management algorithm.
In 2020, Taskeen Ashraf and Yasir Niaz Khan [6] suggested two classifiers for weed density-based picture categorization. The plants were concentrating on the grass family, and it was known as nut grass in rice. Another method proposed that classify the grass density using properties such as invariant to scale and rotation moments. In terms of execution duration, two strategies comparison was done.
In 2020, Merel A.J. Hofmeijer et al. [25] investigated many types of crop variegation used in organic cropping systems and how it impacts arable weed vegetation. Crop variegation has an impact on weed species populations. The influence of regionally diverse crop varieties was reflected in this study, which emphasizes the significance of appropriate crop and weed management.
In 2020, Emad Ali Alsherif [5] emphasized the weed ecology of cereal grains in comparison to other crops, as well as their distribution. There were 92 weed species found in all of the crops investigated, including cereals and other crops. Crop family had a little influence as a factor determining weed composition, according to TWINSPAN and Ward categorization.
In 2020, Ishita Dasgupta et al. [13] investigated the method, which combines IoT devices, WSN and AI techniques to provide farmers with better and more convenient crop suggestions followed by a list of variables including temp, rainfall data, overall land size available, previous crop mature origins, and perhaps other assets. In addition, drone cameras and deep learning technologies were used to detect inappropriate vegetation on crops and specifically for weed detection.
In 2020, Basil Andy Lease et al. [37] created a model with a two-level optimizing framework based on aggregation. To find the optimal rotation-invariant uniform LBP configurations, the first level uses GA optimization. The second level employs the NN that groups the CMA-ES to find the optimal pairings of casting values for each classifier’s projected result.
In 2020, Ke Xu et al. [57] suggested a weed detection system for merging the RGB picture characteristics and depth components in real wheat fields. When detecting the grass weeds that look like wheat, this approach overcomes the constraints of two-dimensional spatial data collected from RGB photos. The shade, location, structure, and complexity characteristics of weeds in wheat fields were retrieved from RGB and the underground photos were collected during the tillering and connecting periods according to the species.
In 2020, Jason Adams et al. [2] compared supervised learning approaches to frequently used thresholding techniques for segmenting plants from the background in such photos. A unique way to create correct labels was created by collecting accurate training data, which was a key barrier to employing supervised learning algorithms for segmentation.
In 2020, Mansoor Alam et al. [4] demonstrated a real time crop and weed identification system for variable-rate pesticide spraying. A RF classifier was used to detect and classify the weeds and crops. The classification model has been trained offline using the self-created dataset were deployed for testing field. Spraying of agrochemicals was accomplished using application equipment that included a PWM based fluid flow control system that is apt for applying the appropriate quantities of agrochemicals by a vision-based feedback system.
In 2020, Borja Espejo-Garcia et al. [16] introduced a crop and weed detection system that combines fine tuning and pre trained convolutional networks with machine learning classifiers trained with earlier selected strategies. To minimize overfitting and provide a stable and compatible result was the main goal of this method.
In 2020, Honghua Jiang et al. [30] suggested a weed and crop detection technique based on a GCN. The retrieved CNN features and the Euclidean distances of these CNN were used to create a GCN graph. By mixing labeled and unlabeled image qualities based on semi supervised learning and practice samples that obtain label information from labeled weed data through propagation and the GCN graph strengthened the hypothesis.
In 2020, Kavir Osorio et al. [47] introduced three approaches for weed assessment using deep learning image analysis, and these were compared to expert eye estimates. SVM with HOG as a feature representation were one way. The second solution was based on YOLOV3 which took use of its powerful object recognition architecture. To generate a decomposition, the third one used Mask R-CNN. Those approaches were supplemented with a background subtractor that used the NDVI index to expose non photosynthetic objects.
In 2020, Jie You et al. [58] presented a weed and crop separation system that performs better in challenging conditions for accurately identifying the weeds with any forms and provides excellent assistance for autonomous robots to reduce weed density. By incorporating four extra components into DNN based segmentation model, a long-term benefits was achieved.
In 2021, Tongyun Luo et al. [43] suggested a nondestructive intelligent picture recognition approach for weed seed identification in a wide range of practice. The resolution of systematic issues and concerns associated with weed seed identification might lead to improved organization and management of these harmful seeds.
In 2021, S. Imran Moazzam et al. [26] established a novel approach for weed detection and a patch based classification technique for real time aerial spraying systems. This scheme minimizes a three class pixel classification into two class crop-weed patch classification and increases the crop and weed accuracy of classification. VGG-Beet CNN was constructed for classification.
In 2021, Mulham Fawakherji et al. [18] offered an alternate approach to the typical data enlargement methods which was applied on core challenge of precision gardening crop and weed segmentation. Semi-artificial examples were constructed by substituting the object classes with synthetic ones which were starting with genuine photos.
In 2021, Xiaojun Jin et al. [31] suggested a novel approach that blends deep learning with image processing technologies in an unconventional way. The first step was to use an educated Center Net model to recognize veggies and create boundary boxes around them. The leftover green items that fell out of the boundary boxes were then labeled as weeds. This method can significantly decrease the amount of the trial picture collection as well as the weed detection complexity.
In 2021, Mohana Das et al. [12] developed a deep learning segmentation model for addressing the difficulties of class imbalance that concentrates on the simple class without harming other type of classes. The suggested model was trained and tested using an initial stage canola field picture dataset. The suggested technique beats the benchmark deep learning models in terms of weed and injured canola plant segmentation according to the evaluation findings.
In 2021, Radhika Kamath et al. [32] used a strategy called semantic segmentation to categorize two different agricultural weeds such as bogbean weeds and latifolious weeds in paddy fields. The rice crop and two kinds of weeds were segmented using three models: SegNet, PSPNet, and UNet. This information might be used to prescribe appropriate weed killer to farmers which results in site-specific weed management and long-term agriculture.
In 2021, S. V. Jansi Rani et al. [28] proposed a feature extraction methods and a machine learning system for effectively distinguishing crop and weed. The weed and crop characteristics were retrieved using a graph of slopes and speeded-up significant improvements. Weed and crop categorization was done using logistic regression and SVM techniques. This model was used in a weed detection system for a farm robot.
In 2021, Shahbaz Khan et al. [33] constructed a deep learning system to recognize weeds and crops in croplands. High-resolution UAV footage contains two separate fields such as pea and strawberry, which was used to apply and assess the created system. Based on current machine learning and deep learning-based systems, the developed deep learning system outperformed all of them and can be integrated into a precision sprayer to apply the SSWM approach.
In 2022, Najmeh Razfar et al. [50] suggested a vision-based weed detection system that efficiently detects weed within a soybean crop using machine learning models. Five deep learning models including MobileNetV2, ResNet50, and three custom CNN models were used.
In 2022, Tao Tao and Xinhua Wei [55] presented a deep CNN with a SVM classifier to increase the precision of winter sowing [9,38] and weeds in farm categorization. The majority of crop picture attributes were used to identify weeds. Color, shade, structure, and form were the retrieved picture attributes. Specified features relied on human labor which was partially blind.
Analysis on collected research papers
Analysis on dataset used
In the existing literature works, various datasets such as image, crop and weed, training, soil, broadleaf grass, Bonn and Stugatt, carrot, RCB-D, LabelMe, Original+ours, Original+GAN, CWFID, Rumex 100, BLW,GW, TREC, Stack EX[L], Parrot sequoia have been gathered in order to validate the suggested prototype. Among them all, the image dataset [1,6,23,35,36,40,48,50,55,59], training dataset [3,15,18,22,33,54,56,60] and crop/weed [4,12,16,20,52,53] were the most frequently used datasets. In the gathered research studies, several data sources were selected, and their analysis is shown in Fig. 1.
Analysis on different datasets.
Analysis on pre-processing techniques
Preprocessing is commonly utilized at the beginning stages of the machine learning and AI development pipeline to ensure correct findings. The aim of the pre-processing technique is to enhance the quality of the image. Table 1 displays the various pre-processing approaches used in the gathered research publications. NDVI has been used in [53]. K-means based feature learning has been used in [50]. Diversity minimization has been used in [40]. Image to patches, class balancing has been used in [12]. One-hot encoding technique has been used in [13]. Histogram equalization and morphological dilation techniques have been used in [33]. Binarization technique has been used in [3]. Image enhancement techniques has been used in [4]. Acquisition technique has been used in [16]. NDVI, RVI have been used in [20,47]. Morphological erosion and dilation has been used in [22]. Data standardization and data whitening have been used in [54]. Down-sampling and smoothing have been used in [56]. Mosaicking has been used in [39]. Therefore, this analysis helps to understand the important pre-processing techniques utilized in the crop/weed classification systems. Also, this analysis aids the upcoming researchers to find out the different pre-processing techniques.
Analysis on feature extraction techniques
Analysis on feature extraction techniques
A phase in the latent factor procedure is feature extraction, which splits and reduces a huge set of original data into tiny groups. As a consequence, processing will be more straightforward. Table 2 displays the various feature extraction approaches used in the gathered research publications. Pattern recognition has been used in [48]. Sliding window concept has been used in [49]. GLCM has been used in [6]. Skeleton based or shape based features have been used in [53]. The sugar beet and weed classification system has been used in [18]. CNN with dropout ReLU to extract robust and discriminative features has been used in [17,30]. Color feature has been used in [31]. Morphology-based features have been used in [11]. The multilevel resolution feature has been used in [12]. Distribution features of between row weeds and in row weeds in the tillering stage has been used in [57]. Forward feature selection and backward feature elimination method has been used in [32]. Row detection, row position, row relative assumptions, modeling PDF, Inferring probabilities has been used in [14]. Dataset features like temperature, precipitation, humidity etc
Analysis on performance metrics
The different performance measures of 50 articles were assessed in this part, which included a total of 50 studies. In 38 papers [1–4,6–8,13–15,20,23,24,26,30,32–37,39–41,43,44,47–50,54–56,58–60], accuracy was assessed, which was 48% of overall contribution. Precision was evaluated in 17 papers [2,12,14–16,20,22,23,28,37,40,41,43,47,49,56,58], which was amounted to 14% of the total contribution. Recall was evaluated in 12 papers [12,15,16,20,22,23,28,37,41,43,47,56], which was amounted to 15% of the total contribution. F1 score was evaluated in 7 papers [16,20,23,28,37,43,47], which was of 9%. Kappa coefficient was evaluated in 2 papers [24,33], which amounted to 3% of the total contribution, sensitivity was evaluated in 2 papers [22,47], which amounted to 15% of the total contribution. Figure 2 shows the analysis on performance.
Analysis on performance metrics.
Weeds are classified as annual, biennial, or perennial weeds based on their ontogeny. In this survey, the classification methods used in 50 papers were analyzed. A machine vision algorithm, bayesian classifier has been used in [48]. A robotic weed knife controlling algorithm has been used in [49]. SVM has been used in [6]. MWLP has been used in [53]. Vision based weed detected system has been used in [50]. AlexNet, GoogLeNet, and VGG have been used in [43]. VGG-Beet CNN has been used in [26]. A FCNS model has been used in [40]. A machine vision approach has been used in [25]. TWINSPAN and Ward classifications has been used in [5]. Position based crop plant detection method has been used in [52]. An acquisition chain model has been used in [42]. cGAN model has been used in [18]. HoG model has been used in [17]. Center Net model has been used in [31]. Pixel-based approach has been used in [11]. DeepVeg segmentation model has been used in [12]. Data gap repair algorithm has been used in [57]. SegNet, PSPNet, Unet has been used in [32]. Bayesian classification – GMM model has been used in [14]. Naive Bayes, SVM, KNN, multiple regression and K-means have been used in [13]. Pixel-level weed classification model has been used in [37]. Logistic regression and SVM algorithms have been used in [28]. VGG-SVM model has been used in [55]. SVM, KNN, and YOLO-v3 have been used in [33]. The PLS-DA classification model has been used in [24]. The PLS-DA classification model has been used in [1]. LDA, QDA have been used in [2]. A SVM based classification model has been used in [3]. A real-time computer vision based crop/weed detection system has been used in [4]. A shape based weed detection algorithm has been used in [7]. A sugar-beet detection algorithm has been used in [8]. A leaf kinematic model has been used in [15]. A fine tuning model has been used in [16]. A RF model has been used in [20]. The HSV decision tree method has been used in [22]. A machine vision approach has been used in [23]. The GCN-ResNet-101 approach has been used in [30]. PHOW, DoG, detHess method has been used in [35]. BOF, SPM have been used in [34]. ImageNet Classification has been used in [36]. Fuzzy logic classification has been used in [41]. Otsu’s method has been used in [44]. SVM, HoG,YOLOV3 has been used in [47]. The vector quantization method has been used in [54]. Ground based machine vision and image processing techniques have been used in [56]. A DNN-based segmentation model has been used in [58]. KNN, ensemble methods, complex tree and logistic regression has been used in [60]. Grabcut, LWMDP algorithm have been used in [59]. OBIA has been used in [39]. Table 3 shows the analysis of the classification models.
Analysis on classification techniques
Analysis on classification techniques
(Continued)
In this survey, the maximal accuracy of 100% has been recorded in the OBIA method [39] and the least accuracy of 73% has been recorded in [6]. The maximum precision of 100% has been recorded in the robotic weed knife controlling algorithm [49] and the least precision of 0.95% has been recorded in [2]. The maximum recall of 100% has been recorded in the RF model [20] and the least recall of 46.34% has been recorded in [56]. The maximum F1 score of 99.29 % has been recorded in the fine tuning model [16] and the least F1 score of 88% has been recorded in [47]. The maximum sensitivity of 83% has been recorded in SVM, HoG, and YOLOV3 [47]. Finally, the maximum and least specificity of 0% have been recorded in [47]. If the SVM method uses the best enhancement methods in the future, the specificity gets increased.
Research gaps and challenges
Still, automated weed management is in its infancy due to different specifications in the agricultural sectors. Some drawbacks of the existing literature are given below. The MWLP method needs technical options for enhancing speed [53]. Due to a lack of shape information, the FCNS model was unable to correctly identify crops and weeds [40]. The position-based crop plant detection method requires additional informations, such as plant morphology, and spectral characteristics, to attain enhanced results [52]. Additional parameters need to be tested in the acquisition chain model to see how they affect crop-weed discrimination [42]. Calculating fitness values repeatedly in GA may pose some computational difficulties [31]. The dataset used in the RF method was relatively limited [20]. One restriction of the machine vision method was that multiple plants of the same class that overlapped in the output label image were not split into different plant regions [23]. The fuzzy logic classification and Otsu’s method have a high processing time [41,44]. The YOLOV3 method overestimates the high weed coverage values [47].
Moreover, in most cases, irrigation efficiency decreases as a result of improper management. Additionally, the value of the land decreases. The major crop’s production has decreased. Reduce the amount of moisture in the soil. When time is of the essence, it is necessary to plant. Some methods require a better understanding of ecological functionalities and weed management strategies, even in the face of varying climatic changes. More progress is required in artificial intelligence, which must be capable of processing massive amounts of data. Automated management must also enrich the efficiency of crop yielding, and hence proper management requires with adequate information, which is missing in this scenario.
Conclusion
In this work, the most recent developments in the study of crop/weed classification have been examined, and talked about the difficulties and prospects for further study. This study provides a thorough overview of 50 research papers on crop/weed classification topic. Due to the inter-disciplinary nature of this topic, more than ten public datasets have been gathered using various modalities, and numerous weed detection techniques have been reported in various academic fields. It also discussed the pre-processing, feature extraction, and classification methods. The classification method includes machine vision approaches, machine learning approaches, deep learning approaches, pixel-level based approaches, and so on. In addition, each work’s performance metrics, such as accuracy, recall, precision, F1 metrics, kappa coefficient, sensitivity, specificity, position score, path score, and union score, have been presented. From the reviewed papers, it is clearly known that, the maximal accuracy of 100% has been recorded in the OBIA method [39] and the least accuracy of 73% has been recorded in SVM [6]. Eventually, the research gaps and challenges of the reviewed papers have been discussed.
From this work, we know that reviewed crop/weed classification models have attracted growing interest from several research communities. After going over the aforementioned methodologies, it is still necessary to improve crop/weed classification methods; this might be done by using hybrid techniques in upcoming projects. Also, this work suggests that large datasets are critical to taking this research field to a new level.