Nematode Detection and Classification Using Machine Learning Techniques: A Review

Abstract

Nematode identification and quantification are critical for understanding their impact on agricultural ecosystems. However, traditional methods rely on specialised expertise in nematology, making the process costly and time-consuming. Recent developments in technologies such as Artificial Intelligence (AI) and computer vision (CV) offer promising alternatives for automating nematode identification and counting at scale. This work reviews the current literature on nematode detection using AI techniques, focusing on their application, performance, and limitations. First, we discuss various image analysis, machine learning (ML), and deep learning (DL) methods, including You Only Look Once (YOLO) models, and evaluate their effectiveness in detecting and classifying nematodes. Second, we compare and contrast the performance of ML- and DL-based approaches on different nematode datasets. Next, we highlight how these techniques can support sustainable agricultural practices and optimise crop productivity. Finally, we conclude by outlining the key opportunities and challenges in integrating ML and DL methods for precise and efficient nematode management.

1. Introduction

Nematodes are thread-shaped roundworms that are found in various habitats worldwide, including soil, freshwater, and marine ecosystems, and as parasites of plants, animals, and humans [1]. They can cause substantial economic impact by deteriorating plant health and reducing yields [2,3,4]. Approximately 40 thousand species of nematodes are known [3,5]. The majority of these nematode species are free-living and non-plant parasitic. They do not rely on other organisms for food or shelter. Instead, they feed on organic matter or bacteria in their environment. These nematodes play an important role in nutrient cycling and soil health. However, a significant number of nematode species, particularly the plant parasite nematodes, around 4100, can cause a huge loss of horticultural crops (e.g., tomato, pepper, melon) and field crops (e.g., rice, potato, soybean) [6,7,8].

Plant-parasitic nematodes (PPNs) invade root tissues, disrupting water and nutrient uptake. This leads to stunted growth, wilting, and increased susceptibility to other pathogens. Their impact is particularly severe in intensive agricultural systems, where their populations can build up rapidly under favourable conditions [9]. Although management treatments are available, effective management requires timely detection and accurate identification, which remains a challenge due to their microscopic size, morphological similarity among species, and their presence in complex soil ecosystems [10].

Various biochemical and Deoxyribonucleic Acid (DNA)-based indicators can be used for nematode identification [11,12,13]. However, these methods require a specialised infrastructure and technical expertise, which can limit their accessibility, particularly for on-farm or low-resource settings. More commonly, manual microscopic observation of soil extracts is used to determine the number of nematode cysts in debris extracted from soil samples [14]. The process of nematode extraction from soil samples for microscopic examination is well described by Hallmann et al. [15]. The Baermann funnel approach [16] is another effective technique where the soil is placed in water on a mesh, allowing nematodes to move out of the soil into the water over a certain period [17]. These manual techniques are labour-intensive and costly. Recently, remote sensing (RS) methods have been explored for nematode infestation detection and mapping in agricultural fields. Here, the spectral reflectance of healthy and infested plants is captured and analysed using statistical techniques [18]. For instance, the early detection of pine wilt disease (PWD) infestation using drone-based remote sensing was demonstrated by Syifa et al. [19] in Chuncheon City of Korea. However, information on the use of AI techniques such as machine learning (ML) and deep learning (DL) in the detection and quantification of nematodes remains scarce [20,21].

There have been several reviews on the use of machine vision coupled with microscopic imaging for diagnostic purposes. A key review covering soybean cyst nematode (SCN) detection and management was reported by Arjoune et al. [18]. The review covered ML methods for image analysis-based SCN detection; however, recent DL methods such as You Only Look Once (YOLO) were not covered. The review by Ma et al. [22] covered both traditional image processing and DL approaches in the context of diagnostic microscopic work but did not cover nematode detection in specific. Similarly, Liu et al. [23] focused on DL for general microscopy image analysis of microorganisms, without specific consideration of nematodes. Recently, Pun et al. [24] reviewed the emerging techniques such as image processing, remote sensing, and high-throughput sequencing techniques for PPN detection. However, there was a limited discussion on DL methods and their applications. The summary of highly relevant survey works is presented in

Table 1. Summary of existing reviews on nematode detection and counting.

Ref	Description	Field	Datasets	Limitation and Gaps
[18]	soybean cyst nematode (SCN) detection and management	Nematode detection	Microscopy, remote sensing and hyperspectral imaging	- Covers SCN detection and management, vegetation cycle, yield loss, and ML-based image analysis. - Does not cover recent DL methods such as You Only Look Once (YOLO).
[22]	Transition from traditional image processing and ML to DL methods.	Image analysis	Microorganism images	- Reviews 142 papers from 1985–2021. - Focuses on microorganisms, not nematodes.
[23]	Microscopy image analysis with DL.	General image analysis	Microscopy images	- Discusses segmentation, classification, and reconstruction. - Focused on general microscopy, not nematode detection.
[24]	Review of emerging techniques for PPN.	PPN identification	Microscopy, remote sensing, and hyperspectral imaging	- Covers high-throughput sequencing, metabarcoding, remote sensing, hyperspectral analysis, and image processing. - Limited discussion of DL-based detection.
[20]	Nematode identification methods.	Nematode identification	Diverse array of datasets	- Reviews traditional identification methods. - Limited coverage of DL-based detection.

, including their short description and limitations. Summarising these reviews, we identified that they explore image-based nematode detection using ML and image processing techniques but often either lack focus on recent DL advancements or do not address nematode detection specifically.

Considering the above-mentioned limitations and gaps in the existing literature, this survey aims to provide a detailed review of recent advances in DL techniques for nematode recognition, classification, and detection. It emphasises the use of modern DL architectures with microscopic imaging for nematode identification, covering a wide spectrum of ML and DL methods, datasets, and nematode identification tasks relevant to crops and plant health monitoring.

Specifically, this survey has the following main contributions:

(I)

This survey systematically explores the nematode identification and counting using traditional and emerging techniques.

(II)

This survey compares and contrasts the most effective ML and DL models used in nematode identification, providing a performance-based comparison, particularly regarding detection accuracy, and their applicability.

The rest of the paper is organised as follows: Section 2 presents the background and research questions. Section 3 presents the research methodology employed for conducting this review. Section 4 presents the results and discussion. Section 5 summarises the findings of the survey. Section 6 presents the conclusion of the research.

2. Backgrounds and Research Questions

2.1. Microscopy Image and Nematodes

Microscope imaging reveals the elaborate and often fascinating details of the microscopic realm, providing a window into structures and organisms that are obscured to the naked eye. Microscopy image analysis by an expert nematologist is one of the most popular methods for nematode identification. Once the sample is collected from soil or plant tissue, it is observed through the microscope, and nematologists identify the nematode based on their prior knowledge of shape, size, and other morphological features (Figure 1).

Optical microscopy techniques, such as bright-field and differential interference contrast (DIC), are commonly used to capture images of stained or unstained nematodes on slides. In contrast, advanced modalities like scanning electron microscopy (SEM) provide high-resolution images of external morphology and surface details, which are useful for differentiating closely related species with subtle surface differences [26,27]. Despite their utility, manual interpretation of these images remains subjective and error-prone, especially when dealing with large sample volumes or overlapping individuals. Advanced image analysis and computer vision techniques can utilise these microscopy images to automate the process of nematode detection, potentially reducing the reliance on expert nematologists [28].

2.2. Remote Sensing Image Analysis and Nematodes

Identification of nematode infestation at the field level is crucial for effective crop management. Species-specific knowledge of nematodes enables targeted control strategies such as deployment of resistant cultivars, application of selective nematicides, rotation with non-host crops, or adaptation of fallow periods. These customised management interventions not only prevent yield losses but also minimise unnecessary chemical inputs, delay resistance development and that of nematode resistance, and contribute to sustainable pest management [29].

Nematodes primarily infect crop roots, causing significant biotic stress that disrupts crop physiological functions such as water and nutrient intake [30]. This below-ground damage manifests in visible above-ground symptoms, including stunted growth and leaf yellowing and wilting, which ultimately reduces crop yields. Remote sensing (RS) offers a non-invasive and scalable alternative for detecting nematode-induced stress by analysing changes in plant spectral reflectance, which are associated with physiological indicators like chlorophyll content and photosynthetic activity [31].

Among the various RS techniques, UAVs have emerged as a promising tool for nematode infestation detection as they can address most of the limitations associated with satellite-based RS, mostly the flexibility of handling the UAV, capturing the high spatial resolution imagery as a bird’s eye view of the agricultural field and being easy to operate and deploy [32]. For instance, Sun et al. [33] developed an effective way to identify the pine wilt nematode infestation using a drone image acquired through a UAV (Figure 2).

In summary, the RS-based approach for detecting nematode infestations offers the advantages of monitoring large agricultural fields and providing early information to the farmers for timely management and control of nematode infestation. In contrast, microscopy image analysis allows for detailed classification and precise identification of nematodes at the species level, making it suitable for fine-grained analysis.

2.3. ML and DL Models for Nematode Detection and Monitoring

ML is a branch of AI that enables computers to learn from data and make predictions or decisions without being explicitly programmed [34]. DL is a specialised subset of ML that uses multi-layered neural networks to automatically learn complex patterns, particularly effective in tasks like image recognition [35].

Advanced data-driven technologies such as AI, the Internet of Things (IoT), drones, and image analysis offer promising alternative techniques in smart agriculture (SA). In the recent past, advances in such emerging technologies have led to what is termed the ‘digital agricultural revolution’, which has led to a transition from traditional farming into ‘smart farming’ based on historical data, weather patterns, soil conditions, and crop performance and helped increase efficiency, sustainability, inclusiveness, and transparency [36].

The ML and DL techniques have been applied to various aspects of agriculture, including pest detection, disease detection, plant stress monitoring, and weed detection [37,38]. Recently, these techniques have gained significant attention in automating nematode identification and counting. This development is boosted by the increasing availability of digital data from microscope imaging and other high-throughput technologies such as drones and remote sensors [39,40,41].

Supervised ML methods such as support vector machines (SVMs), decision trees (DTs), random forests (RFs), Naive Bayes (NB), and multi-layer perceptron (MLP) neural networks rely on manually annotated datasets that are used to train ML models for nematode identification. Meanwhile, DL approaches, particularly convolutional neural networks (CNNs), have achieved impressive results in biological image analysis, including nematode detection [25]. Their success is largely attributed to pre-trained models, originally developed on large datasets like ImageNet [42] that can be fine-tuned for nematode-specific data through transfer learning [43,44]. These DL-based approaches have helped address complex tasks such as species-level identification and morphological differentiation of nematodes. Automating the detection, identification, and counting of nematodes using ML and DL significantly improves efficiency and accuracy compared to traditional manual microscopy, which is time-consuming and costly.

Implementing ML and DL techniques for image-based nematode identification involves several key steps: data preparation, image annotation, feature extraction, model training, and evaluation [45,46]. Image annotation is a critical step, often performed using labelling tools like LabelImg, with output formats such as YOLO and Common Objects in Context (COCO). For instance, Pun et al. [25] used “LabelImg” [47] to annotate nematode eggs and saved the data in YOLO format. Additionally, image preprocessing techniques such as noise reduction, contrast enhancement, and morphological operations are applied to improve image quality and feature extraction [48].

2.4. Research Questions

We consider the following research questions to guide the review and explore the relevant literature.

RQ1: What types of nematodes and infestations have been investigated using image-based analysis methods? Since nematodes cause distinct damage patterns in crops, and different species can affect various parts of the plant (e.g., roots, stems, leaves), this question explores the range of nematode infestations successfully detected using image analysis methods. This helps assess the effectiveness of these technologies for identifying and differentiating nematode species and their specific impacts on crops.

RQ2: What are the most successful and accurate ML and DL methods for nematode recognition using different image modalities? This question seeks to compare and evaluate the performance of various data-driven techniques, particularly DL models, for detecting nematodes. This helps identifiy the most effective algorithms and methods for automating nematode detection.

RQ3: What are the key challenges and opportunities in utilising image modalities for nematode detection? This question aims to explore the limitations of existing methods across various benchmarking datasets, image modalities, and experimental setups, thereby suggesting potential avenues for future research on more precise nematode monitoring.

3. Survey Method

We employed a standard literature review approach to systematically explore existing research on DL and its potential applications in nematode identification, detection, and counting. This process involved gathering relevant research articles and synthesising the information based on the specific research questions we sought to investigate. The review adheres to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [49] to collect the pertinent publications (Figure 3). The steps involved in the review are discussed below.

3.1. Search Strategy

To address the formulated research questions, a specific search strategy was defined to narrow down the literature search from broad topics, such as “nematodes” and “image analysis,” to more specific topics, including “nematode detection” and “deep learning.” We formulated the following search query: (“Machine learning” OR “Deep learning”) AND (“Nematode”). A comprehensive search was conducted across four major databases: Scopus, Google Scholar, PubMed, and MDPI. We restricted the search to articles published between 2012 and 2025 (as machine learning and deep learning have seen significant advancements since 2012 [50]; after execution of the search query, we received the following numbers of articles MDPI: 544; Scopus: 262; PubMed: 246; and Google Scholar: 42 (accessed date: 23 January 2025). It is important to note that the keywords were searched using the default settings of the respective databases, except for Google Scholar. For Google Scholar, an advanced search feature was employed to narrow the results to articles with the keywords specifically within the title. This adjustment was necessary, as the default search settings on Google Scholar yielded an overwhelming number of results, approximately 14,400. The detail search strategy for each database is reported in

Table 2. Detailed search strategy for each database. Note that the search query used for all databases was ((“Machine learning” OR “Deep learning”) AND (“Nematode”)), and the date accessed was 23 January 2025.

Database	Search Strategy	Remarks
MDPI	(Search within = ‘All fields’, Article Types = ‘Article’ and Years = ‘2012–2025’)	Article types and Years are used as search filters.
Scopus	(Search within = ‘Article title, abstract, keywords’, Years = ‘2012–2025’, Document Types = ‘Article’, language = ‘English’)	Document type, Years, and Language are used as search filters
PubMed	(Search within = ‘All fields’, Text Availability = ‘full text’, Years = ‘2012–2025’, Language = ‘English’)	We utilise text availability as an extra search filter in this case.
Google Scholar	(Search within = ‘title’, Years = ‘2012–2025’)	Advanced setting such as with all of the words = ‘Nematodes’, with at least one of the words = (‘machine learning’, ‘deep learning’) were used as search filters.

.

3.2. Inclusion and Exclusion Criteria

The article selection process (Figure 3) began with screening the duplicate records and irrelevant literature from consideration. Then, the title, keywords, and abstract filtering were applied to filter out the irrelevant articles. Finally, the following inclusion/exclusion criteria were used to finalise the articles to be included in the survey:

(a)

Publications that are not related to automated or semi-automated detection of nematodes;

(b)

Articles that are not peer-reviewed and not available in full;

(c)

Articles that are written in a language other than English.

3.3. Data Item Extraction

Using the survey approach reported in Figure 3, a total of 52 articles were selected for final data extraction. These literatures were analysed based on the following data items:

(i)

Meta information such as publisher, title, publication date and year, etc.

(ii)

Study area, nematode types, and data acquisition modalities.

(iii)

Nematode identification task, such as classification, segmentation, and detection.

(iv)

ML and DL models and their performance metrics.

4. Result and Discussion

4.1. Traditional Image Analysis Methods

Digital image analysis methods offer a semi-automated approach to nematode identification by utilising morphological features observed in microscopy images. These techniques aim to extract meaningful geometric and textural features from the images and synthesise them into higher-level concepts that facilitate the identification of nematodes [51].

The image analysis methods, such as contour arc (CA), skeleton structure (SS), and extreme point (EP), are implemented to detect and count nematodes and nematode eggs [14,52,53]. CA refers to a connected curved segment of a contour extracted from a digital image, possibly representing parts of the nematode body. SS denotes the skeleton structure, which represents the medial axis while preserving the shape’s topology and connectivity. EPs are points on the nematode’s contour, such as the head and tail. Combining these features enables accurate detection and counting of nematodes in microscopy images. These methods essentially capture the shape of the object as arcs or curves and help measure the size of nematodes [54,55]. For instance, Pun et al. [53] utilised the digital image analysis technique, such as CA, showing a strong correlation ($R^2$ of 0.857) when tested on 517 microscopy images.

The summary of related studies on nematode and egg detection and counting based on digital image analysis are reported in

Table 3. Overview of image analysis methods used for nematode detection and counting. Note that EPN, RKN, and SCN represent Entomopathogenic, root-knot, and soybean cyst nematodes, respectively.

Ref	Input Images	Methods	Remarks
[52]	Nematode egg	Image analysis techniques such as CA, SS, and EP	The semi-automated counting methods for RKN eggs achieved with overall $R^2$ of 0.90.
[56]	RKN Eggs	Contour-based method	The counting of three types of RKN (M. enterolobii, M. incognita, M. javanica) was highly correlated with human raters with $R^2=0.977$, $R^2=0.990$, $R^2=0.924$, respectively.
[57]	EPN image	Standard curve method	The nematode area was estimated using image processing, and the results showed that the nematode pixel area from image analysis was well correlated with the total number of nematodes in the sample ($R^2=0.99$).
[53]	RKN image	Image analysis techniques such as CA, thin structure and skeleton graph	When these methods were tested on 517 microscopy images, the result of automated counting was highly correlated with manual counting of the nematode, with the highest $R^2=0.857$ using the CA method.
[58]	Cyst nematodes	Image analysis techniques such as thresholding colour, removing outliers, and watershed, etc.	A low-cost and open-source imaging method for nematode counting was developed.
[59]	SCN	Statistical models	Nematode infestation on soybean using machine learning and high resolution multispectral aerial imagery.
[60]	EPN images	Image analysis techniques such as skeleton and two-path analysis	Detected and counted the dead nematodes in microscopy images.
[61]	Nematode image	Image analysis techniques, such as auto-contrast technique and segmentation, are used.	An automatic and intelligent technique for nematode identification was developed using a neural network.

, which shows that these methods can produce reasonable success in automated or semi-automated counting of nematodes like Entomopathogenic nematode (EPN), RKN, and SCN in microscopy images. However, these methods rely on manually designed features such as CA, curve fitting, and skeleton graph, which may not perform well with varying conditions such as complex nematode sizes, species, overlapping nematodes, and different backgrounds.

4.2. ML-Based Methods

De Jesus et al. [62] evaluated the performance of four ML algorithms, namely RF, Stochastic gradient boosting (SGBoost), SVM, and k-nearest neighbour (KNN), for classifying free-living nematode specimens (Acantholaimus and Sabatieria). The results demonstrated that the RF algorithm produced the highest accuracy in correctly classifying free-living nematode in comparison to the algorithms: SGBoost, SVM, and KNN. Chen et al. [63] investigated the possibility of identifying the sugar beet nematodes (SBNs) using ML techniques. They implemented a new labelling strategy based on the Local Maximum Boundary Intensity (LMBI) and combined it with SVM for SBN segmentation. The SVM-based detector achieved recall ∼0.68 and precision ∼0.74, which shows the potential of ML to automate the SBN identification process. Ropelwaka et al. [64] developed ML methods to identify the four species of cyst nematodes (Globodera pallida, Globodera rostochiensis, and Heterodera schachtii). Here, they extracted the different image texture features to be utilised as input to an artificial neural network (ANN) classifier.

Besides microscopy images, RS images are utilised for nematode infestation detection in large agricultural fields. For instance, Santos et al. [31] utilised the MicaSense RedEdge sensor mounted on a drone to capture images of nematode infested areas in three soybeans fields. At each site, nematode species were extracted in the laboratory from soil collected from the geo-referenced points. Then the training samples of nematode-infected and non-infected points were identified using the threshold methods. Finally, ML methods, including RF, DT, analysis of variance (ANOVA), and logistic regression (LR), were implemented for nematode infection prediction. Among these methods, the DT algorithm achieved the highest classification accuracy (>70%) when using spectral bands, with green (560 ± 20 nm) and near-infrared (840 ± 40 nm) bands identified as the key input variables for the model. The summary of performance of ML methods for nematode image analysis tasks such as classification and segmentation is reported in

Table 4. Performance metrics of ML methods for nematode image analysis. Note that P, R, F, and Acc. represent the precision, recall, f1-score, and accuracy in percentage (%). This notation is consistently used in the subsequent tables. Note that the symbol ‘-’ represents metrics ‘not reported’.

Ref.	Dataset	ML Methods	P	R	F	Acc.	Remarks
[31]	RS images	RF, CIT, LR	-	-	-	71.00	Multispectral imagery acquired with drone.
[62]	Microscopy image	RF, SGBoost, SVM, KNN	-	-	-	93.00	Two nematode dataset Acantholaimus (D1), Sabatieria (D2) are used.
[63]	Microscopy image	SVM	68.05	74.94	-	-	Sugar beet nematodes were analysed.
[64]	Microscopy image	ANN	89.80	-	89.70	86.30	Cyst nematodes were analysed.
[65]	Flatbed scanner image	SVM	96.73	98.15	97.13	96.21	Nematode (Caenorhabditis elegans) in Petri dish scanned photo were analysed.
[41]	HS spectra	MLC	-	-	-	93.00
[66]	Microscopy image	KNN, SVM, XGBoost	-	91.60	-	95.50	EPN were classified.
[67]	RS data	SVM, RF, KNN DT	-	-	-	72.00	Root lesion nematodes by proximal sensor were analysed.

.

In summary, the ML methods for nematode identification are focused on classification and segmentation tasks. Among them, the majority of studies tackle the classification problem, where they report the promising classification accuracy in the range of 71–100%. RF and SVM appear to provide the best performance for nematode classification.

4.3. DL-Based Methods

In this section, we report the results of various DL-based methods for nematode recognition in microscopy and other image modalities. Essentially, the image recognition task can be categorised into classification, segmentation, and detection, which are described in the following sections.

4.3.1. Object-Level Classification Methods

Classification is the fundamental task in computer vision and image analysis. The image classification deals with classifying the whole image into predefined nematode groups representing nematode species or classes. Figure 4 illustrates the nematode classification using the DL-based approach.

DL networks such as Visual Geometry Group (VGG) and Residual Network (ResNet) are applied for nematode image classification. These models are built upon the powerful framework of CNNs, which are essential for processing and analysing complex image data [50].

The image-level classification was intensively investigated for nematode image identification (

Table 5. Performance of DL methods for nematode classification. Note that P, R, F, and Acc. represent the precision, recall, f-score, and accuracy in percentage (%), respectively. Note that the symbol ‘-’ represents metrics ‘not reported’.

Ref.	Dataset	DL Methods	P	R	F	Acc.	Remarks
[68]	Microscopy image (PPN)	ResNet101v2, CoAtNet-0, EfficientNetV2B0, EfficientNetV2M	98.26	97.26	97.99	98.66	957 image of PPN from Indonesia, representing 11 classes/species
[69]	Microscopy image (Juveniles and Adult)	Xception	-	-	-	88.28	Juveniles and adults nematode image by light microscope
[70]	Microscopy image	NemaNet and other DL models	98.96	98.87	98.91	98.80	Microscopy images of different nematode species
[26]	Microscopy image (NemaRec)	ResNet101	-	-	-	54.7	Microscopy image consisting of 19 nematode species (2769 images) collected in China
[71]	Microscopy image (I-Nema)	Xception, ResNet50, ViT, and so on	-	-	-	86.78	This includes the 40 nematode species (9215 microscopy images)
[72]	Microscopy image (PPN)	EB-Net	-	-	-	71.00	It includes the PPN image of 14 species from Peru, Mexico and Europe
[46]	Microscopy image (EPN)	Custom CNN	95.66	95.56	95.56	98.52	A custom CNN was developed to classify EPN species using microscopy images

). For instance, Shabrina et al. [68] collected an image of PPN in Indonesia that included 11 species of nematodes. The various DL methods for classification, such as ResNet101v2, CoAtNet-0, EfficientNetV2B0, and EfficientNetV2M, were evaluated for performance, and the highest accuracy of 98.66% was achieved with the EfficientNet model. However, the dataset used in the study is not publicly available, so benchmarking their results is not possible. Similarly, the classification of juvenile and adult nematodes’ microscopy images captured with an OPTIKA camera using an Xception model was reported by Uhlemann et al. [69]. The Xception model achieved 88.28% for the juvenile dataset and 69.45% for the adult dataset. The results showed that the model performed quite well for juvenile image classification, whereas it struggled with adult nematode image classification. Furthermore, a microscopy image dataset of nematodes was made publicly available by Abade et al. [70] for nematode image classification benchmarking of DL models. They implemented the custom CNN (NemaNet) and compared its performance with other DL models. The NemaNet achieved the highest classification accuracy of 98.8%.

In summary, the accuracy of DL methods for nematode species classification ranges from 54.70% to 98.66% (Table 5). The performance of classification models is mainly affected by the number of species to classify, the DL architecture used, and the quantity of training data available. For instance, in the NemaRec dataset, ResNet101 achieved a relatively low accuracy of 54.70%, likely due to the high number of species (19) and limited dataset size (2796 images), which together increase classification complexity. Additionally, the high visual similarity among nematode classes further contributes to this challenge. In contrast, NemaNet, a DL model specifically designed for nematode classification, achieved significantly higher accuracy of 98.88% when trained on a dataset of 3063 microscopy images representing only five species.

4.3.2. Pixel-Based Segmentation Methods

Unlike object-level classification methods, which assign a single label to an entire image, image segmentation classifies each pixel in the image, enabling the precise delineation of nematodes and other objects within an image. For instance, instead of classifying an entire image as containing a nematode or not, segmentation methods identify which pixels belong to the nematode, effectively separating it from the background and other structures. This process can be viewed as a form of pixel-wise classification. DL architectures commonly used for nematode segmentation in microscopy images include UNet, SegNet, and DeepLab, all of which have demonstrated strong performance in image segmentation tasks. In addition to semantic segmentation, instance segmentation methods such as Mask Region-based CNNs (Mask R-CNNs) [73] have also been applied for nematode segmentation, as they assign labels to individual object instances, facilitating tasks such as nematode counting and tracking.

Segmentation-based methods are implemented to identify the nematode species in microscopy [74,75] or remote-sensing images [76,77]. For instance, Qin et al. [76] acquired the pine wood images through the onboard UAV multispectral sensors and segmented the pine wood nematode infection area on the UAV images. They implemented a novel deep learning architecture based on a spatial-context attention network (SCANet) and compared its performance with another standard CNN-based segmentation model, such as CANet, and DeepLabV3. The highest accuracy of 79.00% for nematode detection was reported for SCAnet. Similarly, Shen et al. [77] examined the performance of UNet and DeepLabV3 for nematode-infested region identification on UAV images for pine wood nematodes and achieved the highest f-score of 99.13%. They also reported other evaluation metrics such as a recall of 88.50% (

Table 6. Performance of DL methods for nematode segmentation. The * represents the instance segmentation method. Note that the symbol ‘-’ represents metrics ‘not reported’.

Ref.	Dataset	DL Methods	P	R	F	Acc.	Remarks
[76]	RS images	SCANet, CANet, SNet, DeepLabV3+, HRNet	86.00	91.00	-	79.00	PWN image acquired by drone
[77]	RS images	VGG with UNet, ResNet50 with DeepLabV3+	-	-	88.50	99.13	Pine Wood Nematode (PWN) disease identified based on drone imagery
[74]	Microscopy images	UNet and Attention-UNet	-	-	-	85.00	Microscopy images of nematodes
[78]	RS image	UNet	66.00	74.66	69.00	-	Nematode pest detection in coffee crops using drone imagery
[73]	Microscopy image	Mask R-CNN *	96.00	95.66	95.8	-	Microscopy image fo C. elegans

).

4.3.3. Object Detection-Based Methods

The object detection task in general is a complex task within the computer vision field that combines object classification and localisation. The object of interest in the case of nematode detection is the localisation and classification of nematode objects in the microscopy images. Unlike nematode image classification, which simply assigns a label to an entire microscopy image, object detection requires not only labelling the nematode object but also identifying its location by drawing a bounding box around it. This makes it more challenging as it involves both identifying the nematodes and determining their position in the image (Figure 5).

Object detection can be achieved using two types of detectors: two-stage and single-stage detectors. Two-stage detectors first generate region proposals which are likely to contain objects, followed by a second stage that classifies these regions and refines their bounding box. The notable methods in this category include region-based CNN (R-CNN) and Faster R-CNN [79]. These methods tend to be more accurate but are computationally intensive, as each proposed region needs to undergo a CNN. In contrast, single-stage detectors perform object classification and bounding box regression in a single pass, enabling faster inference. Popular examples include YOLO and Single Shot Detectors (SSDs) [80], which offer real-time performance with a potential trade-off in accuracy.

Object detection-based methods are better suited for nematode identification than classification and segmentation, particularly for practical tasks like locating and counting nematodes [81,82]. Recently, there has been a surge of studies exploring this pipeline for nematode identification and counting. Akintayo et al. [83] identified the SCN nematodes and their eggs using the selective autoencoder (SAE) model with an overall accuracy of 94.33%. They began with the collection of soil samples from SCN-infested fields by arbitrary placement of 25.4 mm diameter probes in a zigzag 11 intervals on several farms in the state of Iowa, the United States. Furthermore, they implemented an end-to-end Convolutional Selective Autoencoder (CSAE) model for SCN detection, which resulted in 95.05% accuracy [84]. Chen et al. [85] designed a custom CNN to detect the potato cyst nematode (PCN) and reported an overall precision of 84.20%, which is relatively lower compared to other nematode detection methods [84]. The performance summary of the object detection-based methods is synthesised in

Table 7. Detection performance of DL methods for nematode datasets. The * represents mean average precision (mAP). Note that ‘CASE’ and ‘SAE’ represent Convolutional Selective Autoencoder and Selective Autoencoder, respectively. Note that the symbol ‘-’ represents metrics ‘not reported’.

Ref	Dataset	Methods	P	R	F	Acc.	Remarks
[84]	Microscopy image	CSAE	93.73	-	94.40	95.05	Detection and counting of SCN in a microscopic image.
[85]	Microscopy image	CNN	84.20	85.63	-	-	CNN-based on Line Annotations was implemented to detect PCN in Microscopy images.
[82]	Microscopy image	YOLO-v5s	78.10	78.30	-	-	Detection of the infective juvenile stage of EPN using a bounding box.
[81]	Microscopy image	YOLOv5	100.00	99.80	99.90	-	YOLO models with mosaic data augmentation were implemented to detect RKN.
[86]	Microcopy images	YOLOv5	85.10	75.30	-	-	Plant parasite nematode were detected in complex microscopy samples.
[76]	RS images	YOLOv5	98.70	98.10	97.30	-	Infestation of nematode on Pine wood was estimated using multi-spectral drone images.
[87]	RS images	YOLOv3	84.38	99.09	-	-	Dead pine tree detection due to pine wood nematode (PWN) using drone imagery.
[88]	RS images	Improved-YOLOv8	85.20	64.30	-	-	PWN disease tree detection using drone imagery.
[45]	Microscopy image	YOLOv8x	-	-	-	94.00	Detection and counting fo Nematode eggs.
[89]	Microscopy image	YOLOv6	96.53 *	-	-	-	AgriNema dataset.
[90]	RS image	YOLOV8	87.90	87.00	-	-	PWN infected trees were detected using drone images.

.

Another widely used object detection technique for nematode recognition is YOLO, which is a real-time object detection algorithm known for its speed and accuracy in detecting a range of objects in images using a single forward pass through a CNN. The YOLO has undergone several improvements since its initial release, resulting in multiple versions (YOLOv2, YOLOv3, and so on up to YOLOv12), each enhancing detection accuracy, speed, and architectural efficiency. For instance, Pun et al. [25] compared the various variants of YOLO models (YOLOv2 to YOLOv7) for RKN detection in tomato plants using microscopy images. They collected images of RKN using a BX53 microscope with an Olympus DP80 Camera with a 4×objective lens. The study used the YOLO (v5 to v7) models to identify RKN on tomatoes. The highest f-score of 99.9% was reported with the YOLOv5 model with the mosaic data augmentation technique.

While looking into the performance of these object detection models, it was determined that they demonstrated high accuracy in controlled settings; however, their generalisability across nematode species and varying field conditions remains limited [84]. Furthermore, the robustness of the models is highly dependent on nematode species morphology and image quality, underscoring the need for standardised datasets and cross-species validation to advance reliable nematode detection systems.

5. Summary of Findings

5.1. Nematode Identification Using Image Analysis Methods

Nematodes are diverse species that respond differently to environmental factors, and manual counting remains the most common method for enumerating them. However, accurate nematode detection is crucial for various purposes, including population enumeration and mapping infestations. In recent years, image analysis methods have been developed for nematode detection, showing a strong correlation between manual and automated detection and counting [52]. Several studies have explored the use of image-based techniques [18,22,53] for identifying a broad range of nematodes, including RKN, EPN, and free-living nematodes.

These methods can generally be divided into two categories: microscopy-based and RS-based methods. Microscopy methods focus on detecting and enumerating nematodes in high-resolution images, offering high accuracy in population enumeration. For example, some studies have reported an impressive accuracy of $R^2$ of 0.99 between manual and automated counting [52]. However, these methods often rely on manual feature extraction and may struggle in complex environments such as images with highly cluttered backgrounds. On the other hand, RS-based methods utilise data acquired from drones or proximal sensors to monitor nematode infestation in agricultural fields. These methods can identify and map the nematode infestation patterns on a large scale, enabling early intervention and management to mitigate the potential crop damage.

5.2. Recent Advances in Nematode Identification Using ML and DL

A key trend across the studies is the growing use of ML and DL techniques, which are increasingly being applied to enhance nematode detection in precision agriculture. In particular, research focused on crops such as sugarcane, soybean, potato, and tomato highlights how these technologies facilitate earlier and more precise detection of nematodes, enabling effective management of agricultural pests and nematodes.

In summary, DL methods have demonstrated superior performance in detecting, classifying, and segmenting nematodes in microscopy images. However, it is important to recognise that these methods are not without challenges, particularly when applying them under varying conditions or with different datasets. For example, while models may perform exceptionally well on training data, they may struggle to generalise to new or unseen samples due to issues like over-fitting.

Among detection models, the YOLO and its variations consistently emerge as the top performers, thanks to their real-time object detection capabilities. For the classification task, the models like NemaNet, ResNet, and EfficientNet show excellent accuracy. Finally, the DeepLabV3 and UNet are shown to be effective for segmentation, making them well suited for detailed spatial localisation of nematodes. However, performance variations across these ML and DL models can arise from factors like differences in nematode types, image acquisition platforms, modelling task types (i.e., classification or segmentation or object detection), training dataset size (more details are reported in

Table A1. Summary for nematode detection tasks, datasets, and AI methods. Note that Acc., Prec., Rec., IoU, MPA, mAP represent accuracy, precision, recall, intersection of union, mean pixel accuracy, and mean average prevision respectively.

Ref.	Nematode	Input Type	Task Type	AI Method	Dataset Size	# Classes	Eval. Metrics	Public Avail.
[31]	Nematode	Drone	Classification	ML	-	-	Acc.	No
[62]	Marine nematode	Microscope	Classification	ML	260	-	Acc.	No
[63]	Cyst nematode	Microscope	Classification	ML	435	2	Prec.	No
[64]	Cyst nematode	Microscope	Classification	ML	-	-	Acc.	No
[65]	Nematode	Scanner	Classification	ML	40,394	2	Acc.	No
[41]	RKN	HS spectra	Classification	ML	-	8	Acc.	No
[67]	RLN	Proximal sensor	Classification	ML	-	4	Acc.	No
[68]	PPN	Microscope	Classification	DL	957	11	Acc.	Yes
[69]	EPN	Microscope	Classification	DL	188	3	Acc.	No
[70]	Phytoparasitic	Microscope	Classification	DL	3063	5	Acc.	-
[26]	Nematode	Microscope	Classification	DL	2769	19	Acc.	-
[71]	Nematode	Microscope	Classification	DL	9215	40	Acc.	Yes
[72]	Cyst-nematode	Microscope	Classification	DL	-	-	Acc.	No
[46]	Nematode	Microscope	Classification	DL	513	5	Acc.	No
[76]	PWN	Drone	Segmentation	DL	-	-	Prec., Rec., Acc.	No
[77]	PWN	Drone	Segmentation	DL	-	-	IoU, MPA, Acc.	No
[74]	Nematode	Microscope	Segmentation		4000	-	Acc.	No
[78]	Nematode pest	Drone	Segmentation	DL	-	-	Acc.	No
[73]	C. elegans	microscope	Segmentation	DL	1908	1	Prec., Rec., F-score	Yes
[84]	SCN eggs	Microscope	Detection	DL	644	-		Yes
[85]	PCN	Microscope	Detection	CNN	3376	-	Prec., Rec.	No
[82]	EPN	Microscope	Detection	DL	1135	2	Prec., Rec., mAP	No
[81]	RKN and FLN	Microscope	Detection	DL	4606	2	Prec, Rec., F-score, mAP	No
[86]	PPN	Microscope	Detection	DL	3503	-	Prec., Rec. mAP	No
[76]	PWN	Drone	Detection	DL	4862	-	Prec., Rec., Acc.	No
[87]	PWN	Drone	Detection	DL	1872	-	Prec., Rec.	No
[88]	PWN	Drone	Detection	DL	2478	-	Prec., Rec, F-score, mAP	No
[45]	RKN	Microscope	Detection	DL	4742	-	Acc.	Yes
[89]	PPN	Microscope	Detection	DL	525	-	mAP	No
[90]	PWN	Drone	Detection	DL	894	-	mAP, Prec., Rec.	No

), and issues such as image resolution, annotation accuracy, and class imbalance. These factors can affect the model’s ability to learn and generalise effectively among the different nematode species and growth stages.

5.3. Challenges and Future Directions

We summarise the challenges and opportunities in data acquisition, benchmarking, and modelling for nematode detection.

(i)

The quality of image acquired with microscope relies on factors such as specimen preparation, microscope configuration, and sensor characteristics. Additionally, the skill and expertise of operators to utilise the optimal setting of the microscope plays a crucial role in controlling image resolution, lighting, and noise. These dependencies pose challenges for achieving consistent quality across datasets. Furthermore, the overlapping of various nematode species on the images further complicates the detection process, and many methods struggle to handle such complications. There is demand for developing more robust object detection and segmentation methods capable of handling such complex structure of a nematode.

(ii)

There are a few nematode datasets publicly available for benchmarking the performance of ML and DL models on the nematode detection task. Data labelling is always a labour-intensive and costly task that demands nematology experts to create such a dataset. Despite the scarcity of large, annotated datasets for training deep learning models that hinder the performance, there are opportunities to automate the data labelling process with advanced semi-supervised or foundational AI models such as Segment Anything (SAM) [91].

(iii)

The promising result of ML and DL models in nematode detection has paved the way towards the widespread use of such emerging technologies in the near future. However, these methods should be rigorously tested before implementing them in agricultural practice as the different nematode species have different characteristics at their different growth stage, which further complicate the detection process. For instance, the ML and DL models tested on the juvenile stage may not perform well in detecting nematodes in the adult stage. This brings the opportunities to develop the multi-stage multi-modal AI methods that can leverage the information from all growth stages of nematode from multiple input modalities, such as image and other environmental factors.

6. Conclusions

In this work, we conducted a systematic review of existing literature on nematode identification using image modalities based on the PRISMA guidelines from two perspectives: (a) identification of nematodes and their infestations that have been investigated using image analysis and (b) finding the most successful and accurate DL method for detecting nematodes.

From the survey, it is evident that ML and DL hold significant potential for identifying and counting nematodes through image analysis. The advanced computational modelling approaches offer the ability to automate and enhance traditional methods, addressing the challenges, including manual identification and quantification’s time-consuming, labour-intensive nature. Leveraging emerging techniques of image processing and feature extraction in DL models can significantly enhance accuracy, scalability, and efficiency. These improvements make them essential tools for precise nematode detection, enabling earlier identification and more effective management of agricultural pests and nematodes. So, ML- and DL-based techniques have the potential to enhance nematode management, leading to sustainable agricultural practices and improved crop productivity.

This review inherits certain methodological limitations. The literature search was restricted to peer-reviewed articles published in English, thereby excluding studies available in other channels, such as grey literature, or those published in languages other than English. Despite these limitations, the study provides a comprehensive examination of ML and DL techniques utilised for nematode detection and quantification. The findings presented are expected to hold significant value for researchers and practitioners in a field that clearly requires further investigation and sustained efforts to achieve its full potential.