Bayesian neural network for plant disease detection: A role of prior knowledge in data-scarce agricultural systems

PLANT SCIENCE TODAY
Vol 13(sp9): 1–15
https://doi.org/10.14719/pst.16499
eISSN 2348-1900

Review Article | [Agriculture] 2026, Volume 13, Issue sp9, pp. 1-15
Bayesian neural network for plant disease detection: A
role of prior knowledge in data-scarce agricultural
systems
A Tharshiniremila1, R Vasanthi1*, R Pangayar Selvi1, M Kalpana1 & C Indu Rani2
1Department of Physical Sciences and Information Technology, Agricultural Engineering College and Research Institute, Tamil Nadu
Agricultural University, Coimbatore 641 003, Tamil Nadu, India
2Department of Vegetable Science, Horticultural College and Research Institute, Tamil Nadu Agricultural University, Coimbatore 641 003,
Tamil Nadu, India
*Correspondence: R Vasanthi, vasanthi@tnau.ac.in
Abstract
Plant diseases are a major threat to global food security, causing an estimated annual economic loss of approximately
USD 220 billion. Even though deep learning models excel at image-based disease detection, the requirement for many
labelled images with a quantitative prediction uncertainty can be problematic in real-world agricultural applications. Bayesian
neural networks (BNNs) overcome these limitations by incorporating prior knowledge from plant pathology, symptom
morphology and environmental factors within a probabilistic learning framework. This approach not only improves prediction
accuracy but also yields calibrated uncertainty estimates alongside each prediction. This review synthesises recent studies
on BNNs, focusing on domain-knowledge, data-driven and functional priors for plant disease detection. The evidence gained
from the studies reviewed indicates that BNNs consistently outperform deterministic models, especially in low-data regimes,
as may be the case for smallholder farming characterised by heterogeneous environments and limited annotation resources.
The review also highlights challenges related to computational complexity, prior selection and edge deployment, while
outlining future research opportunities for robust agricultural artificial intelligence (AI) systems.
Keywords: Bayesian neural networks; data scarcity; plant disease detection; prior knowledge; uncertainty quantification
Received: 02 July 2026. Accepted: 15 September 2026. Available online: Version 1.0: 27 October 2026.
Cite this article : Tharshiniremila A, Vasanthi R, Pangayar S R, Kalpana M, Indu R C. Bayesian neural network for plant
disease detection: A role of prior knowledge in data-scarce agricultural systems . Plant Science Today. 2026; 13(sp9): 1-15.
https://doi.org/10.14719/pst.16499

1. Introduction
   Plant diseases are among the major constraints
   limiting global agricultural productivity and food
   security. Agricultural biodiversity plays a critical role in
   supplying food and raw materials to humans, making
   agricultural biodiversity an essential component of
   human well-being (1, 2). One of the most critical
   factors affecting food production is plant disease (3).
   Plant diseases arise from biotic pathogens such as
   fungi, bacteria, viruses and nematodes, while abiotic
   stresses including nutrient imbalance, temperature
   extremes and soil conditions may produce similar
   symptoms. Plant diseases may reduce plant gro wth,
   yield, quality and overall productivity (4). The agricultural
   industry is also plagued by plant diseases, causing an
   estimated annual global economic loss of USD 220
   billion, primarily driven by bacteria, fungi, nematodes
   and viruses (5). A prominent historical example is the
   Irish potato famine (1845–49), which caused
   approximately 1.2 million deaths (6). Such massive
   losses emphasise the importance of timely and
   precise disease diagnosis to reduce the damage and
   avoid massive outbreaks of diseases that pose a
   threat to world food security, especially to small
   farmers whose livelihoods depend on healthy crops
   (7).
   Conventional methods for identifying diseases,
   which often involve manual examination, are typically
   time-consuming and labour-intensive and subject to
   human error unless trained experts are involved (8).
   This has been an impetus for interest in automated,
   accurate and scalable diagnostic systems. Deep
   learning has become one of the most powerful and
   accurate tools of image classification and there are
   considerable prospects for improving the control of
   plant diseases and food security in general (9).
   However, the prerequisite for abundant, high-quality
   imagery often limits the utility of deep learning models
   when evaluating low-resolution inputs or navigating
   challenging field environments characterised by
   cluttered backgrounds, shadows and non-uniform
   lighting that are common in real-life agricultural

Plant Science Today | 2026 | Volume 13(sp9) | Page 2
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
conditions (10). Nevertheless, the fact that these
models fail to measure uncertainty tends to
undermine the practical use of those models in real-
world agricultural contexts, which are essential to
make sound decisions, explaining the high stakes of
their application (11, 12).
These limitations highlight the need for more
robust and data-efficient models when limited data
and computational resources are needed, where
Bayesian neural networks (BNNs) offer a promising
solution (13). The BNNs provide a principled
methodology to incorporate existing background
knowledge and quantify uncertainty in the models,
which is especially beneficial in the field of data
scarcity, such as agricultural disease detection where
large, high-quality annotated datasets are often
unavailable (13). Thus, techniques that can utilise
previous information to regularise the learning
process as well as approximate uncertainty play a
crucial role in creating dependable diagnostic tools in
precision agriculture (14). It can be used to make
more informed inferences since the inherent ability of
BNNs to condition their parameters with prior
distributions enables them to make inferences when
the empirical data is sparse, hence minimising
overfitting and enhancing generalisation abilities (15).
Moreover, BNNs allow parameters to be modelled as
probability distributions and not treated as point
estimates, which allows for a more calibrated
prediction of uncertainty-an important drawback of
deterministic deep learning models that often give
overly confident and poorly calibrated predictions
(16).
The purpose of this review is to fill this gap by: (1)
Reviewing the current state of deep learning and
Bayesian methods in detecting plant disease; (2)
Examining how prior knowledge may be used to
promote the robustness of models, quantifying their
uncertainty and generalising in situations of data
scarcity and local environmental heterogeneity and
(3) Determining the open challenges and research
opportunities in deploying the previous knowledge
BNNs in actual farming systems, particularly in
resource constrained areas.
2. Methodology
Literature for this review was collected by using
combinations of keywords such as “Bayesian neural
network,” “plant disease detection,” “uncertainty
quantification,” “prior knowledge,” and “Bayesian
optimisation” in the Google Scholar search, along
with crop names such as tomato, rice, maize, potato,
chilli and durian. The search was limited to articles
published between 2015–25, with a special emphasis
on the last five years to cover the latest progress in
Bayesian deep learning for agricultural applications.
Studies were included if they employed Bayesian
methods, Bayesian optimisation, or probabilistic deep
learning for plant disease detection or classification,
or if they discussed uncertainty quantification,
integration of prior knowledge, or foundational
Bayesian neural network theory relevant to this
domain. A few studies from adjacent imaging areas,
e.g., medical imaging, were also included when they
provided methodological insights that could be
immediately applied to data-scarce, high-stakes
classification problems analogous to plant disease
diagnosis. Studies not related to plant disease
detection, without any Bayesian or probabilistic
component and with no clear methodological
relevance to the scope of the review were excluded.
As the review is a narrative rather than a systematic
review, no formal preferred reporting items for
systematic reviews and meta-analyses (PRISMA)-
style screening protocol was applied; instead, the
above search strategy was used to ensure broad and
representative coverage of the field.
Fundamentals of Bayesian neural networks
Bayesian neural networks are neural networks based
on Bayesian probability theory that explicitly model
uncertainty in network parameters, making them
particularly suitable for applications such as plant
disease detection where datasets are limited and
noisy (17–19). In contrast to classical neural
networks, which learn fixed-point approximations of
network weights, BNNs consider network weights as
random variables with probability distributions (20).
The use of this probabilistic method enables BNNs to
measure uncertainty in their predictions, which
enables estimation of predictive uncertainty, an
important feature for decision-making in agricultural
diagnostics (21).
Fundamentally, BNNs utilise Bayes' theorem
to incorporate the initial belief, or the prior distribution,
with the available evidence, which is represented by
the likelihood functions, to obtain a posterior
distribution (22). In this framework, the prior
distribution expresses any existing knowledge or
assumptions about the network’s weights before any
training data is considered (20). The probability of
obtaining the provided dataset under model
parameters is estimated by the likelihood function
(23). The aggregate of these leads to the posterior
distribution, which reflects the revised belief of the
parameters based on the observed data, hence
giving a measure of uncertainty in the predictions
made by the model (24).
Mathematically, Bayes' theorem is expressed as:
Here, P (W∣D) represents the posterior
distribution of the network weights W given the
observed data D. P (D∣W) is the likelihood, P (W) is
the prior distribution and P (D) is the evidence or
marginal likelihood that normalises the posterior
distribution (19, 24).

Plant Science Today | 2026 | Volume 13(sp9) | Page 3
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
The BNNs naturally represent two kinds of
uncertainty, namely aleatoric and epistemic
uncertainty (24). Aleatoric uncertainty refers to the
natural noise present in leaf images such as uneven
lighting, shadows, or overlapping symptoms and it is
usually handled by using probabilistic models or
output layers that adjust to these varying noise levels
(25). Epistemic uncertainty is a sign of ignorance
because of a small or biased amount of data and is
represented within the posterior over weights; it
decreases as additional labelled plant pictures are
viewed (17, 25). In plant disease detection, epistemic
uncertainty helps identify cases that differ from the
training data, such as new cultivars or unseen
diseases, making the system safer and supporting
better decision-making for farmers (26).
The general problem of exact Bayesian
inference in deep networks is intractable; practical
approximations are used instead. Variational
inference tackles this by fitting a simpler distribution,
often a mean-field Gaussian, to the true posterior by
minimising the KL divergence, which allows the
model to be trained efficiently with stochastic or
natural-gradient methods (18, 19). Another approach
is Markov Chain Monte Carlo (MCMC), including
methods like Hamiltonian Monte Carlo and stochastic
gradient Langevin dynamics (SGLD), which is
asymptotically exact and more computationally
intensive, though SGLD has been used to detect
plant diseases to get improved calibrated uncertainty
on out-of-sample data (26). Monte Carlo Dropout has
become a popular approximation because it provides
uncertainty estimation with minimal modifications to
conventional neural network architectures (17, 27).
The BNNs provide better calibrated confidence
estimates than deterministic neural networks and
generally improve model generalisation when training
data are limited or distribution shifts occur, making
them particularly suitable for plant disease detection
in data-scarce agricultural environments (18).
Prior knowledge in Bayesian neural networks
Prior knowledge in BNNs refers to information about
the model, data, or task that is available before
observing the training data and is represented as
prior probability distributions over model parameters
(17). It is a formal model of expert knowledge of plant
pathology, symptom morphology, pathogen life cycles
and environmental cues in the form of probability
distributions over model parameters (28). The ability
to regularise the learning process and estimate the
epistemic uncertainty makes the BNNs especially
appropriate in the case of limited or imbalanced
datasets, by combining prior knowledge (26).
Prior knowledge may be incorporated (i) in
weight space, through the selection of structured or
hierarchical priors (ii) in function space, by matching a
desired Gaussian process (GP) prior or physics
based functional prior and then adjusting weight
priors to replicate that prior (iii) as domain knowledge
priors, where any constraint that can be expressed as
a loss (e.g., physics rules, fairness, healthcare
knowledge) is encoded as a prior that assign high
probability to parameter configurations that satisfy
those constraints and (iv) via data driven priors, using
generative models or external summary information
to code the data manifold (24, 29).
Encoding plant knowledge as priors
Plant pathology knowledge can be encoded as a prior
structure over disease-relevant features, attention
patterns or parameter groups that are tuned to the
morphology of symptoms and not incidental plant
traits. Because visual diagnosis depends on lesion
colour, texture, form and spot patterns, BNN priors
can target filters or latent variables that highlight
these symptom characteristics before training (8, 30).
Priors can also guide attention to disease-relevant
areas, decreasing background and non-symptom bias
in complex imagery (31). In tomato leaf disease
diagnosis, for instance, a BNN could assign stronger
priors to latent variables for spot texture and symptom
regions, then update these with labelled tomato
images to yield posterior distributions over disease
classes and confidence (32).
Agronomic knowledge comprises host crop
identity, cultivar/genotype effects, growth stage,
sowing time and management practices that can
affect disease susceptibility under different production
systems (33). While the number of cases that can be
substantiated with direct evidence is limited,
hierarchical priors are supported by the Bayesian
methodology, which offers statistical strength for a
group of related crops/cultivars or crop developmental
stages while preserving statistical variation between
the groups (34, 35). This method is a synthesis of
knowledge, which can be represented as cultivar–
stage-specific prior distributions of the risks of
diseases, updated with field observations to provide
posterior risk for selected cultivar stage combinations.
Environmental conditions like temperature, relative
humidity, rainfall, leaf wetness and their timing can
significantly affect pathogen development and
disease emergence (36, 37). One way to incorporate
this knowledge would be through the prior disease-
risk distribution or structured latent variables in BNNs
that would give higher prior probability to
environmental conditions that favour infection
(33, 38). In the case of grape downy mildew,
humidity- and rainfall-conducive conditions may
receive increased prior outbreak (baseline) risk and
then this risk may be updated with freshly observed
environmental conditions to give posterior disease
risk (36). This represents a methodological synthesis
of established environmental disease relationships
and Bayesian prior methodology rather than a directly
reported implementation of environmental priors in a
plant disease BNN.

Plant Science Today | 2026 | Volume 13(sp9) | Page 4
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
Bayesian updating and benefits
These informative priors are updated as image
evidence and auxiliary agronomic or environmental data
are collected throughout the training process and result
in posterior distributions of weights, latent disease
states, or decision layers instead of just point estimates
(29). The probabilistic updating is well suited to the
agricultural context where your images are likely to have
noise, limited labels and varying acquisition conditions
across the domain, where you need strong models to
distinguish the disease cues from the background
artefacts and you don't know what the acquisition
conditions might have been (39). In addition to providing
a measure of predictive uncertainty, these methods
have been linked to better minority-class performance,
less overfitting, increased reliability and enhanced
generalisation in less-than-optimal or limited-data
settings (8, 39). In that regard, the encoding of plant
pathological, agronomic and environmental knowledge
into priors does not preclude data-driven learning but
aims to focus posterior learning on biologically plausible
solutions, a task for which these constraints are of high
value (40).
Systematic gains of domain knowledge priors
over vague priors, like isotropic GP or generic GP
priors, are seen in a broader Bayesian deep learning
framework, with better results on tasks, since
encapsulated knowledge is inherited into the posterior
samples (41). The accuracy of prediction with
informative Gaussian scale mixture priors tuned to
both feature sparsity and signal-to-noise ratio is better
than vague priors and in certain scientific models,
even better than cross-validation tuned to
hyperparameters (42). When general pre-trained
knowledge is used along with domain-specific fine-
tuned models in an ensemble, the false positive rates
of unknown diseases are significantly lower, once
again demonstrating the advantage of integrating
prior knowledge at the model level in plant disease
anomaly detection (43).
Low-data agriculture should not ignore
domain knowledge, which allows deep models to
overfit spurious backgrounds and shapes,
miscalibrate uncertainty and break on new diseases
and field domains; knowledge-directed and structured
methods make these models more robust (26).
Overall, integrating biological knowledge into BNNs
improves robustness, enhances model interpretability
and increases predictive reliability under data-scarce
agricultural conditions.
Bayesian neural network architecture and
implementation
Data preparation and preprocessing
Benchmark datasets such as PlantVillage provide
large collections of annotated leaf images that are
widely used for training and evaluating Bayesian plant
disease detection models (44, 45). Pre-processing
usually consists of background removal or
suppression, noise removal, colour normalisation and
image resising to the input size requirement of the
CNN. Other studies also involve leaf tiling or texture-
based cropping to avoid overfitting of a model on leaf
shape instead of disease symptoms (46). Data
augmentation methods are also crucial in increasing
the diversity of data and improving the generalisation
and strength of the model, especially when using
small datasets. These include rotations, flips, scaling
and variations in brightness or contrast (3). Typical
dataset splits of training, validation and a test include
70 : 20 : 10, 80 : 10 : 10 or 60 : 20 : 20 (47). It is
necessary to divide the data into training, validation
and test sets appropriately to evaluate the
performance of the model and reduce the threat of
over fitting (8).
Model architecture
In many Bayesian neural networks for detecting plant
diseases, a CNN backbone is used to extract
features, with a Bayesian framework applied to the
final dense layers. Neural architecture models like
ResNet, DenseNet, MobileNet, or even custom CNNs
have been proven to be more effective in leaf disease
classification due to their ability to automatically
discover hierarchical texture and colour patterns
related to disease lesions (4, 48). Bayesian plant
disease models are usually trained on a deterministic
convolutional backbone, where the visual priors are
pretrained and the weights of the Bayesian fully
connected layers are random variables (26, 48).
These provide variational parameters such as mean
and variance to provide approximate posteriors,
producing a predictive distribution on the classes
instead of point estimates (26, 44). Regularisation is
standardised in many cases using isotropic Gaussian
priors when data is limited; in advanced versions,
attention or multi-scale convolutions can be used to
capture the localised symptoms, but such notions
may still be limited to having a Bayesian output to
quantify uncertainty (49). Prior knowledge and
Bayesian fully connected layers are integrated into a
convolutional neural network (CNN) based
architecture to generate probabilistic disease
predictions (Fig. 1).
Training framework
Bayesian neural networks to detect plant diseases
are trained using approximate Bayesian inference,
which balances the data-adjustment and prior belief
of the data through the minimisation of negative log-
likelihood (NLL) and Kullback–Leibler (KL) divergence
on the variational posterior, which results in the
evidence lower bound objective (26, 50). This
encourages uncomplicated, data-approximating
weight distributions, which is essential in constrained
labelled pictures.

Plant Science Today | 2026 | Volume 13(sp9) | Page 5
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
Fig. 1. Bayesian neural network (BNN) architecture.
Approximations: Due to the intractability of
exact inference, variational inference (optimising non-
robust distributions such as factorised Gaussians
using SGD on NLL+KL), as well as Monte Carlo
dropout (sampling posterior subnetworks during test-
time, dropouting), have been proposed (26, 44);
stochastic gradient Langevin dynamics are also
more accurate in reflecting uncertainty (26).
Hyperparameters, such as learning rates, dropout,
prior variances, depth and activations, are optimised
efficiently using Bayesian optimisation to learn small
CNNs that satisfy the constraints of data and
computation (8, 46). Knowledge is embedded with
priors, architecture, transfer, knowledge, approximate
inference, as well as Bayesian tuning, which
guarantee robustness in data-poor environments.
Evaluation and inference of model
Bayesian models are compared to deterministic
CNNs in terms of standard measures: accuracy,
precision, recall, F1 and area under the curve (AUC),
which are compared on datasets like PlantVillage and
field collections (51, 52). Bayesian models assume
uncertainty measures, typically predictive entropy or
variance, between many stochastic forward passes,
which are used to identify hard or out-of-distribution
images as well as minimise overconfident errors (48).
The pre-processing of the image is done at
test time and the result is fed through the CNN
backbone. Several probability vectors are generated
by repeatedly sampling the Bayesian head (Monte
Carlo sampling); the average of the probability
vectors provides the final class probabilities and the
uncertainty is determined by the dispersion of the
probabilities (48). To support real-time applications on
phones or edge devices, the backbone can be made
lightweight (e.g. MobileNet) and the latency
constrained by the number of samples to ensure low
latency but useful uncertainty information (53).
Implementation
The BNN models are essential in creating models
that embrace the uncertainty and prior information,
particularly in agricultural systems that are data
sparse. Each framework has varied goals,
programme settings and interpretation abilities. Thus,
there are many different tools and platforms to furnish
researchers with an opportunity to implement
Bayesian deep learning experiments, with the most
used frameworks summarised in Table 1.
Bayesian techniques are built into proprietary tools
such as MATLAB and open-source frameworks such
as Pyro, TensorFlow Probability, PyMC, RStan and
NumPyro are more flexible and scalable. Such
frameworks justify techniques such as MCMC and
variational inference. They are chosen based on the
data availability, computing resources and the degree
of uncertainty estimation needed in use, such as plant
disease detection. The complete workflow from image
acquisition and preprocessing to Bayesian inference
and uncertainty-based disease prediction are
illustrated in Fig. 2.

Plant Science Today | 2026 | Volume 13(sp9) | Page 6
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
Table 1. Overview of Bayesian neural networks software framework
Frameworks Open source Programming environment Main inference method Typical use difficulty
MATLAB
Toolbox No MATLAB Various, including MCMC and Variational Bayes via toolboxes high
Pyro Yes Python(PyTorch) MCMC and stochastic variational inference Moderate to high
TensorFlow Probability Yes Python(TensorFlow) VI, MC sampling, Bayesian layers Moderate
Rstan Yes R Hamiltonian Monte Carlo (HMC), Variational Inference High
NumPyro Yes Python (JAX) NUTS, SVI Moderate
PyMC Yes Python(Aesara) MCMC, VI Moderate
CNN: convolutional neural network; ELISA: enzyme-linked immunosorbent assay; GAN: generative adversarial network; Gabor:
Gabor filter; GLCM: grey-level co-occurrence matrix; k-NN: k-nearest neighbours; LBP: local binary pattern; MC: Monte Carlo; PCR:
polymerase chain reaction; qPCR: quantitative polymerase chain reaction; ResNet: residual network; SGLD: stochastic gradient
Langevin dynamics; SVM: support vector machine; ViT: vision transformer.
Fig. 2. General workflow of a Bayesian neural network-based plant disease detection pipeline.
Current approaches in plant disease detection
Traditional plant disease diagnosis primarily relies on
expert visual inspection of symptoms such as leaf
spots, chlorosis and wilting. Although effective, these
methods are labour-intensive, subjective and often
detect diseases only after visible symptoms appear
(4, 54). Recent laboratory analyses such as culturing,
enzyme-linked immunosorbent assay (ELISA) and
polymerase chain reaction / quantitative polymerase
chain reaction (PCR/qPCR) are highly specific, but
require infrastructure, trained personnel and time,
which also makes it hard to scale up (4). Simple
imaging and simple colouration can identify lesions
but typically assist, as opposed to substituting, visual
experts (54). Overall, these approaches are labour-
intensive, late-stage and challenging to scale up in
large amounts (4, 54).

Plant Science Today | 2026 | Volume 13(sp9) | Page 7
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
Classical machine learning presented semi-
manual pipelines, which are combined with features
under the control of humans and typical classifiers.
Based on the images of the leaves, the researchers
derive colour histograms, vegetation indices and
texture descriptors (e.g., grey-level co-occurrence
matrices, Local Binary Patterns, Gabor filters), as well
as shape or key point-style features to describe the
lesion size and distribution (4, 55). These feature
vectors can then be classified with support vector
machines, random forests, k-nearest neighbours,
naive Bayes or decision trees and are often preceded
by the steps of segmentation or clustering that isolate
diseased regions (4, 55). On relatively controlled and
balanced datasets, these pipelines typically achieve
accuracies between 85 % and 93 %, significantly
outperforming untrained human observation but
subject to variations in illumination, background
noise, or changes in datasets (54). This reliance on
features which are specified by experts and the use
of controlled environments encouraged the shift to
deep learning (54).
Deep learning, particularly through CNNs,
has emerged as the most prevalent paradigm for the
detection of plant diseases using image-based
approaches. The CNNs are trained on pixel-based
hierarchical features without the need to design features
manually and embrace subtle lesion patterns (54, 56).
ImageNet-pretrained models like VGG, ResNet,
DenseNet, EfficientNet, Inception and MobileNet are
fine-tuned on plant datasets, an idea that has shown a
significant enhancement of performance and data
efficiency (56, 57). On such big, recollected datasets as
PlantVillage, CNN-based classifiers are often able to
gain 96–99 % accuracy and some successful
architectures claim 99.5 % success with dozens of crop-
disease classes (54, 56). Newer surveys also indicate
extensions of plain classification, such as object
detection (YOLO, Faster R-CNN), how to jointly
segment lesions to count them and multimodal
combination of red, green and blue (RGB) with
multispectral or hyperspectral data to detect lesions
earlier and before they cause symptoms (45, 58). At the
same time, vision transformers (ViT) and related
architectures have become competitive or superior
alternatives, based on self-attention, which is able to
model long-range dependencies and complex lesion
distributions (54). These architectures capture long-
distance dependencies and complex lesions distribution
result in 97–99 % accuracy on PlantVillage and similar
datasets, sometimes outperforming strong baseline
CNNs, especially with challenging backgrounds and
lighting variance (59). Surveys of plant protection using
computer vision indicate that ViTs, CNN-ViT hybrids and
generative adversarial network (GAN)-based data
augmentation enhance robustness under variable
lighting and background and detect early when the
symptoms are weak (54). Yet, these deep architectures
often demand a huge amount of labelled data and
computation, which are still problems for smallholders
and data-scarce crops (26).
Despite the remarkable performance of CNNs
and ViT, their dependence on large annotated
datasets and inability to quantify predictive
uncertainty remain major challenges. Bayesian neural
networks address these limitations by incorporating
prior knowledge and probabilistic inference. In
smallholder data-sparse agriculture, the task of
detecting diseases in plants requires data-efficient
predictor models that incorporate prior knowledge,
such as regional epidemiology, climate and crop
cycles, to more effectively regularise and make fewer-
label predictions (45, 54). Bayesian neural networks
address missing elements to make safer decisions in
agriculture, with lightweight designs being most
appropriate for edge devices and uncertainty
quantification.
The main methods of plant disease detection
vary in terms of accuracy, scalability, computational
demands and handling of uncertainty. Briefly
summarise what readers should observe from the
comparison presented in Table 2.
Bayesian neural networks in plant disease
detection
Bayesian convolutional neural networks (BCNNs)
have been claimed to be more resistant to overfitting
on small datasets of plant diseases than conventional
CNNs and to provide predictive distributions, the
standard deviation of which measures confidence in
diagnosis (32).
Table 3 shows that the important observation
from the reviewed studies is that many investigations
employ Bayesian optimisation solely for
hyperparameter tuning rather than implementing fully
BNN capable of uncertainty quantification. shows that
Bayesian methods in plant disease detection are
mainly used for hyperparameter tuning and
uncertainty estimation. The CNN models are used in
most studies with the PlantVillage dataset and are
very accurate (> 94 %). Tomato is the most
extensively studied crop in Bayesian plant disease
detection research, whereas relatively few studies
have investigated wheat, maize, pepper, or durian
(Fig. 3). This imbalance highlights opportunities for
extending BNN research to underrepresented crops.
Critical synthesis: Comparative accuracy and
uncertainty calibration across Bayesian neural
networks methods
While reported accuracy is not necessarily
distinguishing, BNNs' accuracy range is substantially
overlapping that of classical ML (85–93 %), CNNs
(96–99 %) and ViTs (97–99 %) (Table 2), meaning
that the evidence currently available does not support
the accuracy advantage of BNNs. The studies in
Table 3 are not all the same class of Bayesian
methods either. Only a subset of those studies, such
as Monte Carlo (MC) dropout, stochastic gradient
Langevin dynamics (SGLD) and Bayesian learning,
include explicit support for predictive uncertainty
estimation.

Plant Science Today | 2026 | Volume 13(sp9) | Page 8
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
Table 2. Comparison of plant disease detection approaches
Approach Representative methods Typical accuracy
Expert / Manual Scouting Visual inspection by agronomists Subjective
Slow, late-stage detection; labour-
intensive; does not scale (54).
Laboratory Diagnostics ELISA, PCR/qPCR, culturing High specificity
Requires infrastructure and trained
personnel; impractical for field-scale
use (4).
Classical Machine Learning
SVM, Random Forest, k-NN with
hand-crafted features (colour
histograms, GLCM, LBP, Gabor)
85–93 %
Sensitive to illumination and
background variation; requires expert
feature design (4, 55).
Deep Learning-CNNs
VGG, ResNet, DenseNet,
EfficientNet, MobileNet (fine-tuned on
PlantVillage)
96–99 %
Demands large, labelled datasets;
overconfident predictions; poor
uncertainty calibration (54, 56).
Vision Transformers (ViT) ViT, CNN-ViT hybrids, GAN-augmented training 97–99 %
High computational cost; data-hungry;
limited deployment on edge devices
(54, 59).
Bayesian Neural Networks MC Dropout, SGLD, Variational Inference on CNN backbones 94–99 % Approximate inference; computationally intensive; prior selection is non-trivial (26, 48).
CNN: convolutional neural network; ELISA: enzyme-linked immunosorbent assay; GAN: generative adversarial network; Gabor:
Gabor filter; GLCM: grey-level co-occurrence matrix; k-NN: k-nearest neighbours; LBP: local binary pattern; MC: Monte Carlo; PCR:
polymerase chain reaction; qPCR: quantitative polymerase chain reaction; ResNet: residual network; SGLD: stochastic gradient
Langevin dynamics; SVM: support vector machine; ViT: vision transformer.
Table 3. Application of Bayesian approaches in plant disease detection
Author Crop Dataset source No. of images Method Bayesian application Limitation Accuracy (%)
A. Bayesian deep learning approaches
(26) Multiple crops PlantVillage 54306 Bayesian deep learning
Bayesian estimation
of uncertainty using
MC Dropout and
SGLD.
SGLD has higher
computational cost
and lower accuracy in
the evaluated domain
SGD: 0.96 ±
0.00
MC Dropout:
0.94 ± 0.02
(48)
tomato, potato
and pepper
bell
PlantVillage 20639
CNN with
Bayesian
Learning
Bayesian learning
layer on top of CNN
unassessed real-field
generalisation 98.9
(60) Multiple crops PlantVillage 75 FPDNN-BR
Using Bayesian
regularisation
backpropagation to
fine-tune results.
The size of the
dataset is not very
large
97.18
B. Bayesian Optimisation-based approaches
(49) Rice PlantVillage 2370 ADSNN-BO
Used for
hyperparameter
tuning
No uncertainty
estimation 94.65
(46) Tomato PlantVillage 26148 Mobile Net
Used for
hyperparameter
tuning
Unassessed real-field
generalisation 96.31

Plant Science Today | 2026 | Volume 13(sp9) | Page 9
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
(8) Tomato PlantVillage 18159 CNN-Stacking
Used for
hyperparameter
optimisation
Optimization-level
uncertainty only 98.527
(61) Potato PlantVillage 2152
Bayesian-
optimised CNN
weighted
ensemble
Bayesian
optimisation for
ensemble weight
optimisation
Absence of test-time
augmentation, pruning
or uncertainty
estimation.
97.94
(62) Tomato PlantVillage
Not specified
(taken from
research
article)
Bayesian-
optimised Merge
Activator with
Fine-Tuned
Random Forest
(BMAFRF)
Activation of
Bayesian tuning
using Gaussian-
UCB
Absence of
explainability
mechanism because
of the black-box
nature of the deep
learning model.
99.6
(63)
Multiplecrops
(Apple,
tomato, rice)
Kaggle
Tomato –
over 20000;
Rice -120 Hy-SALDD
Bayesian
optimisation used to
tune SVM
hyperparameters
Complex multi-stage
pipeline using hybrid
segmentation (PSP-
Net + HED), BWO
feature selection and
Bayesian-optimised
SVM classifier
Apple: 98.46
Tomato: 95.89
Rice: 91.67
(64) durian Field image 2595 CNN
Hyperparameter
optimisation using
the keras-tuner
The symptoms of
Blight and
Colletotrichum are
quite similar, which
leads to
misclassification.
96.61
(65) Wheat Field images Expert knowledge Bayesian belief network Used for uncertainty prediction Subjective human inspection of lesions causes inaccuracies. 89.3
(66) Cassava and multiple crops Field images+ PlanDoc - Bayesian-optimised Vision Transformer Used for hyperparameter tuning
The model confused
similar Cassava
diseases: Mosaic,
Brown Streak, Green
Mottle.
Cassava-
95.7 %
PlantDoc-
96.2 %
(67) Maize(corn)
PlantVillage
and Google
websites
18888 VGG16
Bayesian
optimisation for
Hyperparameter
Tuning
Grid search
inefficiency and class
imbalance reduced
grey leaf accuracy.
99.21 %
(68) Multiple crops Different source 145500 5DB-DenseConvNet Used for hyperparameter optimisation Extensive use of data augmentation, rather than actual data 99.36
(69) Tomato Kaggle 11,000 modified VGG16
Bayesian
optimisation for
Hyperparameter
Tuning
Confusion between
visually similar
diseases like early
and late blight
97.1
(70) Pepper PlantVillage 2475 Inception V3 with BO Used for hyperparameter optimisation Limited to only two disease classes 99.6
(71) Rice Field images 150
Bayes’ Classifier
and Minimum
Distance
Classifier
Applied for
probability-based
classification of
plant diseases.
Accuracy is not very
high
Bayes’
Classifier-69
MDC- 81
ADSNN-BO: adaptive deep stochastic neural network with Bayesian optimisation; BMAFRF: Bayesian-optimised Merge Activator
with Fine-Tuned Random Forest; BO: Bayesian optimisation; BWO: binary whale optimisation; CNN: convolutional neural network;
FPDNN-BR: fuzzy probabilistic deep neural network with Bayesian regularisation; HED: holistically nested edge detection; Hy-
SALDD: hybrid segmentation, adaptive learning and deep disease detection; Keras Tuner: Keras hyperparameter tuning framework;
MC: Monte Carlo; MDC: minimum distance classifier; PSP-Net: pyramid scene parsing network; SGLD: stochastic gradient Langevin
dynamics; SVM: support vector machine; UCB: upper confidence bound.

Plant Science Today | 2026 | Volume 13(sp9) | Page 10
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
Fig. 3. Number of plants used in Bayesian-approach-based plant disease detection studies.
However, many other studies include
Bayesian optimisation but do not estimate predictive
uncertainty. Thus, the reported accuracies should not
be taken as an indication of better Bayesian
uncertainty calibration. To the best of our knowledge,
the 54306 PlantVillage images are the only set of
images tested and previous studies shows that SGLD
achieves slightly better accuracy (0.96 ± 0.00)
compared to MC dropout (0.94 ± 0.02), but this result
is from a single study and should not be applied to
the entire class of Bayesian inference methods (26).
In addition, the literature reviewed does not always
report on the same set of metrics that are used for
calibration and this makes it difficult to directly
compare the reliability of the predictions.
The BNNs can be especially useful when data
is scarce, or the domain or objective undergoes
significant change and a point estimate represents
only part of the information, such that prior
information and uncertainty can give further insights.
The reviewed studies do not provide matched
comparisons that would quantify the possible benefits
of BNNs in these situations. Studies conducted in the
field demonstrate varying levels of accuracy (89.3 %
for the wheat Bayesian belief network and 95.7 to
96.6 % for the cassava and durian studies), which
may not be solely attributable to the use of Bayesian
inference, as the differences in the datasets, disease
categories and variety of models used make it difficult
to attribute these differences to the use of Bayesian
inference. The body of literature agrees with BNNs as
a promising framework for plant disease detection
that is uncertainty-aware and can be data-efficient but
fails to conclude that BNNs can always outperform
deterministic models. Future research should thus
compare matched models, both Bayesian and
deterministic, in a controlled setting of data scarcity
and domain shift and provide quantitative calibration
metrics, like expected calibration error (ECE), in
addition to traditional predictive performance metrics.
Challenges
One of the major issues is the complexity of
computation and implementation of Bayesian deep
learning. Exact Bayesian inference in large
convolutional neural networks is computationally
intractable, requiring approximate inference methods
such as variational inference and stochastic gradient
Langevin dynamics. These come with extra training
and inference costs compared to standard
optimisation (26, 54)
The BNNs also face significant data issues in
plant disease vision systems. The representation of
many crops, diseases and symptom stages is still
under-represented and most publicly available data
are obtained under controlled conditions such as
PlantVillage (46, 62). Bayesian models may also give
overconfident predictions that are incorrect with noisy
labels or imbalanced classes, particularly in case
priors are not tuned according to the true distribution
of diseases or classes are unbalanced (62).
Close interaction between machine-learning
specialists and plant pathologists is required to
design priors that encode agronomic knowledge (and
not arbitrary Gaussian priors) to accelerate the
adoption relative to traditional CNNs with transfer
learning pipelines (58).
Another major issue is choosing priors; the
initial assumptions concerning the parameters
required to support learning with small datasets are a
critical aspect of the BNN performance that is
beneficial when selected appropriately but detrimental
when selected improperly (72). The best decisions
would be the Laplace distributions to sparsity or
hierarchical Gaussians, which entails an in-depth

Plant Science Today | 2026 | Volume 13(sp9) | Page 11
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
understanding of plant characteristics, disease
lifestyles and local trends (73). Poorly specified prior
distributions may substantially reduce predictive
performance, regardless of network architecture (72).
Interpretability and user trust are also one of
the challenges. The predictions and uncertainty of
predictions should be made clear to farmers and
extension workers. Although Bayesian models give
confidence estimates, these are not necessarily
intuitive and have to be supplemented with
visualisation tools and easy decision thresholds to
facilitate adoption, which is emphasised by material
on the development of more practical, farmer-facing
plant disease interfaces (58, 74).
The high sampling requirements in smart
agriculture make the implementation of BNNs difficult
in edge integration with smart agriculture devices.
The computation of uncertainty needs several
posterior samples, which generates large
computation requirements that are not feasible in
real-time to use on resource-intensive hardware (75).
This is a disadvantage when it comes to quick
diagnostics of the field, where time is paramount.
Opportunities
Leveraging prior expertise in BNNs may facilitate
regularised learning and help avoid overfitting when
data are insufficient. Priors that encode agronomic
information can assist the model when labelled
images are difficult to obtain, as reported by previous
researchers (76). This might be particularly essential
for smallholder systems, where comprehensive
documents are not generally accessible (57).
The integration of BNNs and smaller CNNs
backbones or transfer learning could result in small
models that fit into mobile and edge devices. Work on
small CNNs in combination with Bayesian
optimisation and texture features has demonstrated
that more efficient architectures can still reach an
accuracy above 95 %, suggesting that Bayesian
variants of such networks could support on-device
decision support in low-resource regions (46).
Furthermore, developing hybrid physics-
informed BNNs presents a promising avenue of
research. It is possible to incorporate crop growth
models or spectral reflectance laws as inductive
biases in these models, which integrates the inductive
power of neural networks with the known scientific
laws (77). As an example, combining a light
interaction model with plant canopy or disease
progression patterns would result in a more reliable
and interpretable diagnostic system (78). Such a
strategy might help to increase tolerance to different
environmental conditions and better generalisation to
other plant varieties and diseases.
The BNNs can also be used to provide risk-
aware decision tools that integrate sensor, weather
and management data. The wider Bayesian network
applications in agriculture indicate their power in
handling uncertainty in complex agricultural
environments and integrating distinct information
sources (76). Integrating Bayesian vision models with
probabilistic decision-support systems could directly
link disease diagnosis with crop management
recommendations.
Deep learning systematic reviews also
highlight the importance of generalised models
across geography, cultivar and imaging system
models that are required in plant disease detection
(48, 79). Structured priors with agronomy and
pathology inspiration present BNNs as an
embodiment of a principled step to even more robust
and data-efficient plant health monitoring in practice
in precision agriculture (26).
Future direction
Future research should focus on developing BNNs that
explicitly incorporate agronomic knowledge, such as
lesion morphology, symptom progression and disease
epidemiology, rather than relying solely on Bayesian
optimisation for hyperparameter tuning (48, 49).
In data-scarce scenarios, it is possible to
consider Bayesian transfer learning, where priors are
placed on weights of existing backbones (such as
MobileNet, Inception) that had previously been
trained on leaf disease images and then updated with
a small number of local samples of new crops or
locations (49, 62).
The integration of BNNs and active learning
and data augmentation methods has potential:
uncertain or rare cases of disease can be requested
to be annotated and minority classes can be
represented using synthetic images or texture-based
patches, which is an extension of existing texture and
augmentation methods in closed-ended CNNs with
small networks (80).
To align with the on-farm requirements, the
research must focus on lightweight Bayesian models
executable on mobile or edge devices by beginning
with the already tested, small systems such as
depthwise-separable MobileNet or tiny CNNs to
identify plant diseases in the low-resource
environment (49).
Multi-modal and spatio-temporal Bayesian
models that incorporate images, weather, management
and sensor data are required and they would use
concepts such as weather-aware augmentation and
CNN-based forecasting to improve modelling of disease
variability due to climate change (81).
Interpretability tools (attention maps, class-
activation maps) should also be provided to ensure
that farmers can not only tell where the model is
paying attention but also its level of confidence, which
builds on interpretability analyses already done on
non-Bayesian CNNs (45, 49).
Lastly, next-generation research needs to
establish data-scarce, uncertainty-aware detection

Plant Science Today | 2026 | Volume 13(sp9) | Page 12
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
standards, including cross-field and cross-season
experiments, standardised measures beyond
accuracy (calibration, reliability) and comparisons
between Bayesian and conventional deep models on
realistic but small agricultural problems (82).
Collectively, these research directions will support the
development of robust, interpretable and data-
efficient Bayesian plant disease detection systems
suitable for real-world agricultural applications.
3. Conclusion
This review highlights Bayesian neural networks
(BNNs) as a promising approach for plant disease
detection by integrating prior knowledge with
probabilistic learning to address data scarcity and
improve predictive reliability. Current limitations
include computational complexity, appropriate prior
specification and deployment on resource-
constrained devices. Addressing these challenges will
be essential for broader adoption of BNNs in
precision agriculture. To develop precision
agriculture, future studies should focus on agronomy-
encoded BNNs, transfer learning on a Bayesian
model, active learning, interpretability and
benchmarks in data-poor environments. Continued
advances in Bayesian deep learning, together with
interdisciplinary collaboration between artificial
intelligence researchers and plant scientists, will
support the development of robust, interpretable and
deployable plant disease diagnostic systems capable
of improving sustainable agricultural production.
Acknowledgements
The authors sincerely thank the Chairman and
Advisory Committee members for their valuable
feedback, constructive suggestions and continued
support, which greatly improved the clarity and quality
of this review article.
Authors’ contributions
AT prepared and drafted the manuscript. RV provided
guidance and supervision throughout the writing
process. RPS, MK and CIR critically reviewed and
revised the manuscript. All authors have read and
approved the final manuscript.
Compliance with ethical standards
Conflict of interest
Authors do not have any conflict of interests to declare.
Ethical issues
None
Declaration of generative AI and AI-assisted
technologies in the writing process
During the preparation of this work, the author used
QuillBot and Grammarly for paraphrasing, grammar
correction, language improvement, sentence
restructuring and enhancement of clarity and readability.
The author reviewed and edited the content as
necessary and take full responsibility for the content of
the publication.
References

1. Ramanjot, Mittal U, Wadhawan A, Singla J, Jhanjhi NZ,
   Ghoniem RM, et al. Plant disease detection and
   classification: A systematic literature review. Sensors.
   2023;23(10):4769. https://doi.org/10.3390/s23104769
2. Tirkey D, Singh KK, Tripathi S. Performance analysis of
   AI-based solutions for crop disease identification,
   detection, and classification. Smart Agric Technol.
   2023;5:100238.
   https://doi.org/10.1016/j.atech.2023.100238
3. Tugrul B, Elfatimi E, Eryigit R. Convolutional neural
   networks in detection of plant leaf diseases: A review.
   Agriculture.2022;12(8):1192.
   https://doi.org/10.3390/agriculture12081192
4. Demilie WB. Plant disease detection and classification
   techniques: a comparative study of the performances. J
   Big Data. 2024;11(1):5. https://doi.org/10.1186/s40537-
   023-00863-9
5. Dong J, Fuentes A, Zhou H, Jeong Y, Yoon S, Park DS.
   The impact of fine-tuning paradigms on unknown plant
   diseases recognition. Sci Rep. 2024;14(1):17900.
   https://doi.org/10.1038/s41598-024-66958-2
6. Hughes D, Salathé M. An open access repository of
   images on plant health to enable the development of
   mobile disease diagnostics. arXiv preprint
   arXiv:1511.08060.2015.
   https://doi.org/10.48550/arXiv.1511.08060
7. Guth FA, Ward S, McDonnell K. From lab to field: An
   empirical study on the generalization of convolutional
   neural networks towards crop disease detection. Eur J
   Eng Technol Res. 2023;8(2):33–40.
   https://doi.org/10.24018/ejeng.2023.8.2.2773
8. Khan B, Das S, Fahim NS, Banerjee S, Khan S, Al-
   Sadoon MK, et al. Bayesian optimized multimodal deep
   hybrid learning approach for tomato leaf disease
   classification. Sci Rep. 2024;14(1):21525.
   https://doi.org/10.1038/s41598-024-72237-x
9. Natarajan S, Chakrabarti P, Margala M. Robust diagnosis
   and meta visualizations of plant diseases through deep
   neural architecture with explainable AI. Sci Rep.
   2024;14(1):13695. https://doi.org/10.1038/s41598-024-
   64601-8
10. Hamed BS, Hussein MM, Mousa AM. Plant disease
    detection using deep learning. Int J Intell Syst Appl.
    2023;15(6):38–50.

Plant Science Today | 2026 | Volume 13(sp9) | Page 13
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
11. Zhang Y. On cold posteriors of probabilistic neural
networks: Understanding the cold posterior effect and a
new way to learn cold posteriors with tight generalization
guarantees. arXiv preprint arXiv:2410.15310. 2024.
https://doi.org/10.48550/arXiv.2410.15310
12. Le AT, Shakiba M, Ardekani I. Tomato disease detection
with lightweight recurrent and convolutional deep learning
models for sustainable and smart agriculture. Front
Sustain.2024;5:1383182.
https://doi.org/10.3389/frsus.2024.1383182
13. Xu M, Kim H, Yang J, Fuentes A, Meng Y, Yoon S, et al.
Embracing limited and imperfect training datasets:
opportunities and challenges in plant disease recognition
using deep learning. Front Plant Sci. 2023;14:1225409.
https://doi.org/10.3389/fpls.2023.1225409
14. Melki P, Bombrun L, Diallo B, Dias J, Da Costa JP. Group-
conditional conformal prediction via quantile regression
calibration for crop and weed classification. In: 2023
IEEE/CVF International Conference on Computer Vision
Workshops (ICCVW). 2023. p. 614–23.
https://doi.org/10.1109/ICCVW60793.2023.00068
15. Bargagna F, De Santi LA, Martini N, Genovesi D, Favilli B,
Vergaro G, et al. Bayesian convolutional neural networks
in medical imaging classification: a promising solution for
deep learning limits in data scarcity scenarios. J Digit
Imaging.2023;36(6):2567–77.
https://doi.org/10.1007/s10278-023-00897-8
16. Abboud Z, Lombaert H, Kadoury S. Sparse Bayesian
networks: Efficient uncertainty quantification in medical
image analysis. In: International Conference on Medical
Image Computing and Computer-Assisted Intervention.
Cham: Springer Nature Switzerland; 2024. p. 675–84.
https://doi.org/10.1007/978-3-031-72117-5_63
17. Jospin LV, Laga H, Boussaid F, Buntine W, Bennamoun
M. Hands-on Bayesian neural networks—a tutorial for
deep learning users. IEEE Comput Intell Mag.
2022;17(2):29–48.
https://doi.org/10.1109/MCI.2022.3155327
18. Magris M, Iosifidis A. Bayesian learning for neural
networks: an algorithmic survey. Artif Intell Rev.
2023;56(10):11773–823. https://doi.org/10.1007/s10462-
023-10443-1
19. Goan E, Fookes C. Bayesian neural networks: An
introduction and survey. In: Case Studies in Applied
Bayesian Data Science: CIRM Jean-Morlet Chair, Fall
2020. Cham: Springer International Publishing; 2020. p.
45–87. https://doi.org/10.1007/978-3-030-42553-1_3
20. Chang DT. Bayesian neural networks: Essentials. arXiv
preprint arXiv:2106.13594. 2021.
https://doi.org/10.48550/arXiv.2106.13594
21. Li M, Han H, Liu H, Sang W, Yuan S. Permeability
prediction and uncertainty quantification based on
Bayesian neural network and data distribution domain
transformation. Chin J Geophys. 2023;66(4):1664–80.
https://doi.org/10.6038/cjg2022P0837
22. Fan C, Gao F, Sun S, Cui F. Bayesian neural networks
and its application. In: 2008 Fourth International
Conference on Natural Computation. 2008. p. 446–50.
https://doi.org/10.1109/ICNC.2008.624
23. Raj V, Cui T, Heinonen M, Marttinen P. Incorporating
functional summary information in Bayesian neural
networks using a Dirichlet process likelihood approach. In:
International Conference on Artificial Intelligence and
Statistics. 2023. p. 6741–63.
24. Arbel J, Pitas K, Vladimirova M, Fortuin V. A primer on
Bayesian neural networks: review and debates. Stat Sci.
2026;41(2):316–53. https://doi.org/10.1214/24-STS969
25. Olivier A, Shields MD, Graham-Brady L. Bayesian neural
networks for uncertainty quantification in data-driven
materials modeling. Comput Methods Appl Mech Eng.
2021;386:114079.
https://doi.org/10.1016/j.cma.2021.114079
26. Hernández S, López JL. Uncertainty quantification for
plant disease detection using Bayesian deep learning.
Appl Soft Comput. 2020;96:106597.
https://doi.org/10.1016/j.asoc.2020.106597
27. Gal Y, Ghahramani Z. Dropout as a Bayesian
approximation: representing model uncertainty in deep
learning. In: International Conference on Machine
Learning. 2016. p. 1050–59.
28. Tran BH, Rossi S, Milios D, Filippone M. All you need is a
good functional prior for Bayesian deep learning. J Mach
Learn Res. 2022;23(74):1–56.
29. Holden M, Pereyra M, Zygalakis KC. Bayesian imaging
with data-driven priors encoded by neural networks. SIAM
J Imaging Sci. 2022;15(2):892–924.
https://doi.org/10.1137/21M1406313
30. Li L, Zhang S, Wang B. Plant disease detection and
classification by deep learning—a review. IEEE Access.
2021;9:56683–98.
https://doi.org/10.1109/ACCESS.2021.3069646
31. Rani R, Bharany S, Elkamchouchi DH, Ur Rehman A,
Singh R, Hussen S. VGG-EFFATTNNET: hybrid deep
learning model for automated Chili plant disease
classification using VGG16 and EfficientNetB0 with
attention mechanism. Food Sci Nutr. 2025;13(7):e70653.
https://doi.org/10.1002/fsn3.70653
32. Seno A, Miyagusuku R, Kurokura T, Tabata K, Ozaki K.
Tomato leaf disease diagnosis using Bayesian
convolutional neural networks. In: 2024 IEEE/SICE
International Symposium on System Integration (SII).
2024. p. 623–28.
https://doi.org/10.1109/SII58957.2024.10417302
33. Babu GR, Gokuldhev M, Brahmanandam PS. Integrating
IoT for soil monitoring and hybrid machine learning in
predicting tomato crop disease in a typical South India
station. Sensors. 2024;24(19):6177.
https://doi.org/10.3390/s24196177
34. Fortuin V. Priors in Bayesian deep learning: a review. Int
Stat Rev. 2022;90(3):563–91.
https://doi.org/10.1111/insr.12502
35. Cheng L, Yin F, Theodoridis S, Chatzis S, Chang TH.
Rethinking Bayesian learning for data analysis: the art of
prior and inference in sparsity-aware modeling. IEEE
Signal Process Mag. 2022;39(6):18–52.
https://doi.org/10.1109/MSP.2022.3198201
36. Lee S, Yun CM. A deep learning model for predicting risks
of crop pests and diseases from sequential environmental
data. Plant Methods. 2023;19(1):145.
https://doi.org/10.1186/s13007-023-01122-x
37. Kumar D, Mukhopadhyay R. Climate change and plant
pathogens: understanding dynamics, risks and mitigation
strategies. Plant Pathol. 2025;74(1):59–68.
https://doi.org/10.1111/ppa.14033
38. Logeshwaran J, Srivastava D, Kumar KS, Rex MJ, Al-
Rasheed A, Getahun M, et al. Improving crop production
using an agro-deep learning framework in precision

Plant Science Today | 2026 | Volume 13(sp9) | Page 14
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
agriculture. BMC Bioinformatics. 2024;25(1):341.
https://doi.org/10.1186/s12859-024-05970-9
39. Salman Z, Muhammad A, Han D. Plant disease
classification in the wild using vision transformers and
mixture of experts. Front Plant Sci. 2025;16:1522985.
https://doi.org/10.3389/fpls.2025.1522985
40. Moupojou E, Tagne A, Retraint F, Tadonkemwa A,
Wilfried D, Tapamo H, et al. FieldPlant: A dataset of field
plant images for plant disease detection and classification
with deep learning. IEEE Access. 2023;11:35398–410.
https://doi.org/10.1109/ACCESS.2023.3263042
41. Sam D, Pukdee R, Jeong DP, Byun Y, Kolter JZ. Bayesian
neural networks with domain knowledge priors. arXiv
preprint arXiv:2402.13410. 2024.
https://doi.org/10.48550/arXiv.2402.13410
42. Cui T, Havulinna A, Marttinen P, Kaski S. Informative
Bayesian neural network priors for weak signals. Bayesian
Anal. 2022;17(4):1121–51. https://doi.org/10.1214/21-
BA1291
43. Dong J, Zhou H, Fuentes A, Yoon S, Park DS. Enhancing
anomaly detection in plant disease recognition with
knowledge ensemble. Front Plant Sci. 2025;16:1623907.
https://doi.org/10.3389/fpls.2025.1623907
44. Eunice J, Popescu DE, Chowdary MK, Hemanth J. Deep
learning-based leaf disease detection in crops using
images for agricultural applications. Agronomy.
2022;12(10):2395.
https://doi.org/10.3390/agronomy12102395
45. Shafik W, Tufail A, De Silva Liyanage C, Apong RA. Using
transfer learning-based plant disease classification and
detection for sustainable agriculture. BMC Plant Biol.
2024;24(1):136. https://doi.org/10.1186/s12870-024-
04825-y
46. Restrepo-Arias JF, Branch-Bedoya JW, Awad G. Plant
disease detection strategy based on image texture and
Bayesian optimization with small neural networks.
Agriculture.2022;12(11):1964.
https://doi.org/10.3390/agriculture12111964
47. Lu J, Tan L, Jiang H. Review on convolutional neural
network (CNN) applied to plant leaf disease classification.
Agriculture.2021;11(8):707.
https://doi.org/10.3390/agriculture11080707
48. Sachdeva G, Singh P, Kaur P. Plant leaf disease
classification using deep convolutional neural network with
Bayesian learning. Mater Today Proc. 2021;45:5584–90.
https://doi.org/10.1016/j.matpr.2021.02.312
49. Wang Y, Wang H, Peng Z. Rice disease detection and
classification using attention based neural network and
Bayesian optimization. Expert Syst Appl.
2021;178:114770.
https://doi.org/10.1016/j.eswa.2021.114770
50. Thiagarajan P, Ghosh S. Jensen–Shannon divergence
based novel loss functions for Bayesian neural networks.
Neurocomputing.2025;618:129115.
https://doi.org/10.1016/j.neucom.2024.129115
51. Chen J, Zhang D, Nanehkaran YA. Identifying plant
diseases using deep transfer learning and enhanced
lightweight network. Multimed Tools Appl.
2020;79(41):31497–515. https://doi.org/10.1007/s11042-
020-09669-w
52. Joseph DS, Pawar PM, Chakradeo K. Real-time plant
disease dataset development and detection of plant
disease using deep learning. IEEE Access.
2024;12:16310–33.
https://doi.org/10.1109/ACCESS.2024.3358333
53. Hassan SM, Maji AK, Jasiński M, Leonowicz Z, Jasińska
E. Identification of plant-leaf diseases using CNN and
transfer-learning approach. Electronics. 2021;10(12):1388.
https://doi.org/10.3390/electronics10121388
54. Upadhyay A, Chandel NS, Singh KP, Chakraborty SK,
Nandede BM, Kumar M, et al. Deep learning and
computer vision in plant disease detection: a
comprehensive review of techniques, models, and trends
in precision agriculture. Artif Intell Rev. 2025;58(3):92.
https://doi.org/10.1007/s10462-024-11100-x
55. Sharma N, Sharma P, Kumar N. Feature engineering to
early detection of plant disease using image processing
and artificial intelligence: a comparative analysis. Int J
Latest Technol Eng Manag Appl Sci. 2025;14:1107–13.
https://doi.org/10.51583/IJLTEMAS.2025.1407000133
56. Ferentinos KP. Deep learning models for plant disease
detection and diagnosis. Comput Electron Agric.
2018;145:311–18.
https://doi.org/10.1016/j.compag.2018.01.009
57. Shoaib M, Shah B, Ei-Sappagh S, Ali A, Ullah A, Alenezi
F, et al. An advanced deep learning models-based plant
disease detection: a review of recent research. Front Plant
Sci.2023;14:1158933.
https://doi.org/10.3389/fpls.2023.1158933
58. Shafay M, Hassan T, Owais M, Hussain I, Khawaja SG,
Seneviratne L, et al. Recent advances in plant disease
detection: challenges and opportunities. Plant Methods.
2025;21(1):140. https://doi.org/10.1186/s13007-025-
01450-0
59. Kumar R, Singh V. Advancing plant disease detection: a
comparative analysis of deep learning and hybrid machine
learning models. Mach Graph Vis. 2025;34.
https://doi.org/10.22630/MGV.2025.34.3.3
60. Trivedi VK, Shukla P, Pandey A. Plant leaves disease
classification using Bayesian regularisation back
propagation deep neural network. J Phys Conf Ser.
2021;1998(1):012025. https://doi.org/10.1088/1742-
6596/1998/1/012025
61. Jain A, Dubey AK, Pan VS, Mahfoudh S, Althaqafi TA,
Arya V, et al. Bayesian optimized CNN ensemble for
efficient potato blight detection using fuzzy image
enhancement. Sci Rep. 2025;15(1):31259.
https://doi.org/10.1038/s41598-025-15940-7
62. Padmanabhuni S, Gera P. Bayesian optimized deep
learning and ensemble classification approach for multiclass
plant disease identification. Discov Sustain. 2025;6(1):709.
https://doi.org/10.1007/s43621-025-01648-1
63. Jain A, Dwivedi RK. Hybrid segmentation-based
agricultural leaf disease detection (Hy-SALDD) using black
widow optimization for feature selection, and Bayesian-
optimized SVM classification. Int J Inf Technol.
2025;17(6):3197–211. https://doi.org/10.1007/s41870-025-
02478-0
64. Söylemez İ, Nalici ME, Ünlü R. High-accuracy identification
of Durian leaf diseases: a convolutional neural network
approach validated with K-fold cross-validation and
Bayesian optimization. Appl Fruit Sci. 2025;67(6):464.
https://doi.org/10.1007/s10341-025-01698-9
65. Burney SA, Naseem J. Decision making in uncertainty: a
Bayesian network for plant disease diagnoses. Int J
Comput Sci Inf Secur. 2018:16–21.

Plant Science Today | 2026 | Volume 13(sp9) | Page 15
© The Author(s) 2026. Creative Commons CC BY 4.0. Published by Horizon e-
Publishing Group.
https://doi.org/10.14719/pst.16499
66. Mumtaz S, Ahmed RM, Jalal A. Bayesian-optimized
framework for segmentation guided transformer-based
leaf disease detection. In: 2025 6 th International
Conference on Innovative Computing (ICIC). 2025. p. 1–6.
67. Subramanian M, Lv NP, B J, A MB, VE S. Hyperparameter
optimization for transfer learning of VGG16 for disease
identification in corn leaves using Bayesian optimisation.
Big Data. 2022;10(3):215–29.
https://doi.org/10.1089/big.2021.0218
68. Arun Pandian J, Kanchanadevi K. An improved deep
convolutional neural network for detecting plant leaf
diseases. Concurr Comput Pract Exp. 2022;34(28):e7357.
https://doi.org/10.1002/cpe.7357
69. Arkan TM, Sugiharto A, Wibawa HA. Optimizing VGG16
architecture with Bayesian hyperparameter tuning for
tomato leaf disease classification. J Masyarakat Inform.
2025;16(2):162–73.
https://doi.org/10.14710/jmasif.16.2.73168
70. Alhashmi AA, Alohali MA, Ijaz NA, Khadidos AO,
Alghushairy O, Sayed A. Bayesian optimization with deep
learning based pepper leaf disease detection for decision-
making in the agricultural sector. AIMS Math.
2024;9(7):16826–47.
https://doi.org/10.3934/math.2024816
71. Sharma V, Mir AA, Sarwr DA. Detection of rice disease
using Bayes’ classifier and minimum distance classifier. J
Multimed Inf Syst. 2020;7(1):17–24.
https://doi.org/10.33851/JMIS.2020.7.1.17
72. Pham VC, Nguyen C, Le T, Phung D, Carneiro G, Do TT.
Model and feature diversity for Bayesian neural networks
in mutual learning. Adv Neural Inf Process Syst.
2023;36:80633–43. https://doi.org/10.52202/075280-3534
73. Liu J, Wang X. Tomato disease object detection method
combining prior knowledge attention mechanism and
multiscale features. Front Plant Sci. 2023;14:1255119.
https://doi.org/10.3389/fpls.2023.1255119
74. Jafar A, Bibi N, Naqvi RA, Sadeghi-Niaraki A, Jeong D.
Revolutionizing agriculture with artificial intelligence: plant
disease detection methods, applications, and their
limitations. Front Plant Sci. 2024;15:1356260.
https://doi.org/10.3389/fpls.2024.1356260
75. Thilagavathi T, Arockiam L, Mary IP. A smart agriculture: a
comprehensive survey on IoT-enabled plant disease
detection and agricultural automation. Int J Inf Technol Res
Appl.2024;3(2):39–47.
https://doi.org/10.59461/ijitra.v3i2.107
76. Drury B, Valverde-Rebaza J, Moura MF, de Andrade
Lopes A. A survey of the applications of Bayesian
networks in agriculture. Eng Appl Artif Intell. 2017;65:29–
42. https://doi.org/10.1016/j.engappai.2017.07.003
77. Hu X, Gu X, Sun Q, Yang Y, Qu X, Yang X, et al.
Comparison of the performance of multi-source three-
dimensional structural data in the application of monitoring
maize lodging. Comput Electron Agric. 2023;208:107782.
https://doi.org/10.1016/j.compag.2023.107782
78. Safari MM, Malian A. Plant disease mapping in paddy
growing stages using remotely sensed data. Environ Earth
Sci. 2025;84(1):1. https://doi.org/10.1007/s12665-024-
11991-7
79. Dhaka VS, Meena SV, Rani G, Sinwar D, Ijaz MF,
Woźniak M. A survey of deep convolutional neural
networks applied for prediction of plant leaf diseases.
Sensors.2021;21(14):4749.
https://doi.org/10.3390/s21144749
80. Gülmez B. A comprehensive review of convolutional
neural networks based disease detection strategies in
potato agriculture. Potato Res. 2025;68(2):1295–329.
https://doi.org/10.1007/s11540-024-09786-1
81. Dai G, Fan J, Tian Z, Wang C. PPLC-Net: neural network-
based plant disease identification model supported by
weather data augmentation and multi-level attention
mechanism. J King Saud Univ Comput Inf Sci.
2023;35(5):101555.
https://doi.org/10.1016/j.jksuci.2023.101555
82. Shoaib M, Sadeghi-Niaraki A, Ali F, Hussain I, Khalid S.
Leveraging deep learning for plant disease and pest
detection: a comprehensive review and future directions.
Front Plant Sci. 2025;16:1538163.
https://doi.org/10.3389/fpls.2025.1538163
Additional information
Peer review: Publisher thanks the Sectional Editor and anonymous
reviewers for their contribution to the peer review of this work.
Reprints & permissions information: Available at
https://horizonepublishing.com/journals/index.php/PST/open_access_policy.
Publisher’s Note: Horizon e-Publishing Group remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
Indexing: See
https://horizonepublishing.com/journals/index.php/PST/indexing_abstracting.
Copyright: © The Author(s). This is an open-access article distributed under
the Creative Commons Attribution License CC BY 4.0.
Publisher information: Plant Science Today is published by Horizon e-
Publishing Group with support from Empirion Publishers Private Limited,
Thiruvananthapuram, India.