You are on page 1of 8

Detection and Quantitative Prediction of

Diplocarpon earlianum Infection Rate in


Strawberry Leaves using Population-based
2022 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS) | 978-1-6654-8684-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/IEMTRONICS55184.2022.9795744

Recurrent Neural Network


Oliver John Alajas1*, Ronnie Concepcion II2, Argel Bandala1, Edwin Sybingco1, Ryan Rhay Vicerra2, Elmer P. Dadios2, Christan Hail
Mendigoria1, Heinrick Aquino1, Leonard Ambata1, Bernardo Duarte3
1Department of Electronics and Computer Engineering, De La Salle University, Manila, Philippines
2Department of Manufacturing Engineering and Management, De La Salle University, Manila, Philippines
3MARE - Marine and Environmental Sciences Centre, Faculty of Sciences, University of Lisbon Campo Grande, Lisbon, Portugal

{oliver_alajas*, ronnie.concepcion, argel.bandala, edwin.sybingco, ryan.vicerra, elmer.dadios, christan_mendigoria, heinrick_aquino,


leonard.ambata}@dlsu.edu.ph, baduarte@fc.ul.pt

Abstract—Fragaria ananassa, a member of the rose family’s parts of a plant, leaves tend to be the most frequently observed
flowering plants, commonly recognized as strawberry, is prone part for the detection of infection [1]. Various studies have
to Diplocarpon earlianum infection that causes leaf scorch. been conducted that use digital image processing methods to
Assessment via visual inspection of strawberries by farmers is detect and classify leaf diseases [2, 3]. Using a vision-based
normally ineffective, destructive, and laborious. To address this approach to detect the disease by utilizing leaf images is one
challenge, the use of integrated computer vision and machine of the best options researchers tend to go to [3].
learning techniques was done to classify a healthy from a scorch-
infected strawberry leaf image and to estimate the leaf region Fragaria ananassa which is a common type of strawberry
infection rate (LRIR). A dataset made up of 204 normally species that bears strawberry fruit is a good source of vitamin
healthy and 161 scorch-infected strawberry leaf images was C for the improved immune system, potassium for lowering
used. Images were initially preprocessed and segmented via blood pressure, antioxidants for eliminating free radicals, and
graph-cut segmentation to extract the region of interest for reducing inflammation, flavonoids such as anthocyanins,
feature extraction and selection. The hybrid combination of quercetin, and kaempferol for cancer prevention. It also
neighborhood and principal component analysis (NCA-PCA) promotes healthy eyesight because it contains ellagic acid and
was used to select desirable features. Multigene genetic phenolic phytochemicals. The fruit has a low glycemic index
programming (MGGP) was used to formulate the fitness and, it contains dietary fiber that helps regulate blood sugar
function that will be essential for determining the optimized
levels [4]. However, it is prone to Diplocarpon Earlianum
neuron configurations of the recurrent neural network (RNN)
infection which is also known as leaf scorch. It is
through genetic algorithm (GA), and cuckoo search algorithm
(CSA), and artificial bee colony (ABC). Four classification
characterized by small purplish blemishes that appear on the
machine learning models were configured in which the top side of the leaf. Scorched-infected strawberry farms
classification tree (CTree) bested other detection models with an typically suffer from a decline in their strawberry produce’s
accuracy of 100% and exhibited the shortest inference time of quality and taste. Late diagnosis of scorch infection is one of
14.746 s. The developed ABC-RNN3 model outperformed GA- the main reasons for this. Moreover, monitoring and visually
RNN3 and CSA-RNN3 in performing non-invasive LRIR inspecting these fruits are typically done manually which is
prediction with an R2 value of 0.948. With the use of the NCA- quite ineffective due to its dependence on the observer’s bias,
PCA-CTree3-ABC-RNN3 hybrid model, for crop disease expertise, and observation skills.
detection and infection rate prediction, plant disease assessment
proved to be more efficient and labor cost-effective than manual A study conducted by [5] used the hybrid model utilizing
disease inspection methods. linear discriminant analysis and decision tree to detect and
quantify the damage of bacteria-infected grape leaves, [6]
Keywords—bio-inspired optimization, computational used a gaussian quantum-behaved particle swarm and
intelligence, computer vision, digital agriculture, leaf scorch, recurrent neural network to assess the disease of corn leaves
plant pathology, strawberry leaf spot suffering from leaf spots, and [7] classified three rice leaf
disease using a hybrid machine learning and deep neural
I. INTRODUCTION network. Genetic programming, genetic algorithm, and
Detection and identification of plant diseases is a machine learning models were implemented by [8-12] with
commonly experienced problem in the agricultural crop the utilization of lettuce features to monitor the growth stages
production industry since it directly affects the quality of of lettuce. Musa acuminata’s post-harvest analysis was done
products being sold to local food markets. Researchers found by [13] using hybrid machine learning and deep transfer
that while these diseases show symptoms through different networks. Meanwhile, research by [14] used an adaptive

978-1-6654-8684-2/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
neuro-fuzzy inference system to classify three different lettuce will stretch from a 0 to 255 range. The contrast settings used
seed varieties. On the other hand, research by [15] examined were as follows: (0.2, 0.9), (0, 1). An image output sample
the use of lettuce images to indirectly predict the amount of together with its corresponding histogram is presented in Fig.
nitrate present in aquaponic water flowing in a crop chamber 2. To make the image enhancement task easier, a looping
system of a vertical farm with the use of a recurrent neural program was designed to preprocess all 365 images
network that is optimized by genetic algorithm. simultaneously with the use of an image batch processor, a
built-in application inside MATLAB software.
In this research, machine learning (ML) algorithms such
as support vector machine (SVM), Naive Bayes (NB),
classification tree (CTree), and linear discriminant analysis
(LDA), were implemented to detect a scorch-infected leaf.
This detected leaf is then examined to quantify the leaf region
infection rate (LRIR) in terms of the area of the whole leaf
with the use of a three-layered recurrent neural network
(RNN) that is optimized by three bio-inspired optimization
algorithms: genetic algorithm (GA), cuckoo search algorithm
(CSA), and artificial bee colony (ABC). These models are
responsible for the infection rate prediction process of the
study. This study contributes to the: (1) development of an
efficient and non-invasive vision-based detection of
strawberry scorch leaf disease that would help the farmer to
employ better agricultural practices; (2) development of an
innovative and accurate technique in predicting the leaf
infection rate that would help plant pathologist on how to treat
strawberry leaves with such severity of infection; and (3)
determination and analysis of most significant leaf phenotypes
in terms of morphological, texture and spectral traits that have
high relevance to leaf scorch disease.
II. MATERIALS AND METHODS
For this study, the realization of the model framework
design is presented in Fig. 1. Input images of two classes of
leaves (healthy and scorched-infected) were enhanced with
the use of contrast improvement to make a distinct identity
between the two and, to make the segmentation procedure
much simpler. Two types of the strawberry leaf are being
initialized as an input for image processing: healthy leaf and
scorch-infected leaf. These images were segmented to extract Fig. 1. Methodological framework for the strawberry leaf scorch
classification and infection rate prediction using bio-inspired
the leaf area of both classes (whole and scorched area) for the (population-based) optimization of recurrent neural network
calculation of the scorch infection rate. Once features are
extracted, an optimized recurrent neural network will do the
job of infection rate prediction. To accomplish such a
challenging task, MATLAB R2020b software was used to
perform vision-based leaf inspection, computational
processing, model creation, optimization, classification, and
estimation.
A. Strawberry Leaf Dataset Information
The dataset is composed of 365 images with a 1:1 aspect
ratio and 256-by-256-pixel density. The number of the healthy
leaf is 204 images while the scorched-infected leaf is 161
images. The dataset can be found on [16].
B. Image Enhancement (Contrast Improvement) Fig. 2. Image histogram stretching of scorch-infected strawberry leaf for
Completely raw images of the strawberry leaf (healthy and better segmentation results
scorched) were enhanced by adjusting the contrast of the
images through histogram equalization. By applying the imadj C. Strawberry Leaf Feature Extraction, Selection, and
built-in command of MATLAB, the images’ top 1% and Dataset Normalization
bottom 1% of all pixel values were adjusted from a default To establish a model capable of detecting and computing
value of 0 to 1 for low-in, high-in, and low-out, high-out the amount of infection, eighteen features were extracted
values. By doing this, the color map of the image histogram from the input images. It consists of color features (RGB,

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
HSV, L a*b*, YCbCr), textural features (correlation, optimization). NB has a 0.0512 kernel and cosine distance.
contrast, energy, entropy, homogeneity), and morphological LDA was configured with a delta of 0.0293 and gamma of
features (leaf vegetative pixel area) (Fig. 3). These features 0.6766. Also, Bayesian optimization was used for the
are the primary basis for the prediction model capable of hyperparameter tuning for all the machine learning
estimating the leaf region infection rate of a strawberry leaf algorithms mentioned. To make the classifier model robust
infected by Diplocarpon Earlianum. Using the extracted and unbiased, the image dataset was partitioned into three
morphological area of the scorch-infected leaf (Ascorched) and subparts: training data (56%), validation data (24%), and
the whole leaf area (Awhole leaf), the leaf region infection rate testing data (20%). This data partitioning technique is called
(LRIR) can be solved using (1). stratified sampling in statistical jargon. The purpose of this
step is to lower the error rate in the classification task of the
LRIR = (Ascorched /Awhole leaf) x 100 (1) model.
The assessment of each of these models will be based on
the following criteria: accuracy, inference time, Matthew’s
correlation coefficient (MCC), precision, recall, fall-out, and
hamming loss.

Fig. 3. Segmentation of the scorched-infected strawberry leaf (a) raw leaf


image (b) pre-processed leaf image (c) lazy snapping annotation for the
scorch-infected area (d) lazy snapping for the whole leaf area (e) scorch-leaf
masked region (f) whole-leaf masked region (g) binary masked image of leaf
scorched areas (h) binary masked image of leaf whole area

Since there are 18 features, the model might take quite Fig. 4. Characterization of color-texture-morphological features based on
some time to predict the correct output, hence, only NCA-PCA feature selection
significant features should be considered for the model
development. The most significant leaf features were selected E. Objective Function Creation using Genetic Programing
by using a hybrid approach. It is done by using neighborhood One of the ways to create an objective function for
and principal component analysis (NCA-PCA). PCA will unconstrained optimization problems is using genetic
give the number of significant features that need to be programming (GP). It is a type of computational evolutionary
included. On the other hand, NCA is used to rank the features
based on a particular feature’s variability and impact on the algorithm that combines the use of a regression tree and a
prediction results. This process gave out a three-feature genetic algorithm. This union of two powerful algorithms is
vector of R, G, and a*After extracting the important features, also known as multiple-gene symbolic regression. This is
they were further improved by applying data normalization. used to define the mathematical model to be minimized by
This step is implemented to promote proper scaling of data the optimization algorithms to define the best possible
and, to prevent the unequal scattering of data. The combination of a three-layered neural network which will be
normalization method, Z-score data normalization [17], was used to predict the infection rate that a strawberry leaf
used. It is primarily based on getting column vector data’s contains. GPTIPSv2, a genetic programming tool that has
mean and standard deviation (Fig. 4). To do this, the data ( built-in compatibility with MATLAB, is utilized to generate
of a certain feature (e.g., contrast) is subtracted by the mean
derived from that column vector ( . Lastly, the value is the objective function. The hyperparameters used are as
divided to the standard deviation of that column vector ( . follows: population size of 50, maximum generations of 50,
training instances of 25, tournament size of 50, elite fraction
of 0.1, Pareto tournament probability of 0.2, maximum genes
Normalized value = (2) of 10, maximum tree depth of 5, infinite number of total
D. Strawberry Leaf Health Status Identification using nodes, crossover probability of 0.84, mutation probability of
Machine Learning Model 0.14 and expressional complexity measure.
The fitness function (3) is based on three input variables,
For the detection task, machine learning algorithms such represented by N1, N2, and N3 for each layer of a three-layered
as support vector machines (SVM), Naïve Bayes (NB), neural network. The output parameter is a single variable that
classification tree (CTree), and linear discriminant analysis will provide the root mean square error (RMSE) value which
(LDA) were considered. Optimization parameters for each will help the optimization algorithms figure out the right
algorithm are as follows. SVM has a box constraint (1.6545), combination of neurons in consideration with the smallest
kernel (3.507), bias (1.0384), and solver (sequential minimal RMSE value.

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
properties related to the food source such as nectar taste,
Root Mean Square Error = ( , , (3) energy richness, the difficulty of energy extraction, and
closeness to the beehive. The hyperparameters for the
F. Population-based Optimization of Neurons in Recurrent
implementation of the algorithm are as follows: food source
Neural Network
is 100, maximum iteration of 100, the population size of 50,
1) Genetic Algorithm (GA) and a limit of 150 for the scouting phase.
Genetic algorithm (GA) is one of the well-known bio-
inspired algorithms recognized in the field of artificial G. Strawberry Diplocarpon earlianum Leaf Region
intelligence. It is derived from the biological evolution theory Infection Prediction using Computational Intelligence
called Darwinism, which is primarily based on population The leaf region infection rate (LRIR) is numerically
growth reproduction that introduces the concept of gene computed using the developed recurrent neural network
combination and mutation. The hyperparameters are as (RNN) with three types of optimization methods: GA-RNN,
follows: population size is 50, the number of generations is CSA-RNN, and ABC-RNN. The RNN model was built
50, and the number of runs is 10. The selection parameters through MATLAB by using the built-in function newelm.
are as follows: tournament size of 50, Pareto tournament For the configuration of this network, the training algorithm
probability of 0.2, the mutation rate of 0.1, crossover rate of used is scaled conjugate gradient (SCG), 10,000 iterations
1, constraint tolerance value of 1e-6. Moreover, a nonlinear (epoch), goal parameter is set to 1e-7, and the transfer function
constraint algorithm and selection roulette were designation was ‘tansig’, ‘purelin’, and ‘log’.
implemented. The fitness limit is set to zero with a functional To evaluate the computational intelligence models,
tolerance of 1e-6. The maximum generation allowed is set to parameters such as root mean square (RMSE), coefficient of
100. determination (R2), and mean absolute error (MAE) were
2) Cuckoo Search Algorithm (CSA) recorded. An ideal value of 1 for R2 is desired while a value
The cuckoo search algorithm (CSA) is a type of bio- of 0 is targeted for MAE and RMSE.
inspired algorithm that mimics the behavior of cuckoo birds
in population reproduction [18]. This algorithm utilizes the III. RESULTS AND DISCUSSION
Levy flight theory to generate a new candidate solution A. Strawberry Leaf Health Status Identification using
(eggs). The concept is based on the way some particular Machine Learning Models
cuckoo bird species engaged in obligate brood parasitism. It can be seen in the confusion matrix summary in Fig. 5
The way they do it is by laying their eggs on the nest of a that SVM, NB, and CTree were able to classify the scorched
different species of bird (host bird). They simply rely on other strawberry leaf from the healthy ones with 100% accuracy.
host birds to take care of their offspring. The algorithm relies LDA has the lowest accuracy of 97.37%. In terms of the
on the fact that there is at least a 10% probability that the host amount of inference time, CTree outperformed all other three
bird will discover that the cuckoo bird’s egg is not of its own algorithms. CTree is 66.53% quicker compared to SVM,
and thus, has a two-choice: to remove the cuckoo egg from 39.80% quicker compared to NB, and 9.62% faster than
its nest or carry its eggs (excluding the cuckoo egg) and bring LDA. This result is based on the testing phase of the model
it to a newly created nest. The survivability of the cuckoo bird developed with the use of stratified sampling. It is quite
solely depends on their eggs not being discovered by the host noticeable that the reduction of features used resulted in much
bird and, for the host bird to consider the cuckoo egg as its less computational time as seen in Table 1. By doing so, the
own. The ideal scenario is that the cuckoo egg is not improved performance of the models was able to recognize
discovered and, the cuckoo egg will hatch and grow together Diplocarpon earlianum leaf infection with much lesser time
with the host bird’s real offspring. Here, the population size and computational cost without sacrificing the level of
for the algorithm is set with a value of 20, the discovery rate accuracy.
probability of alien eggs (or solutions) of 0.25, and the
maximum iteration of 100 as the convergence criteria. The B. Objective Function Based on Genetic Programming
population stated is the number of host nests per generation. To create a robust model, function settings for the
This means that a nest can contain only one cuckoo egg. GPIPSv2 were set to times, plus, minus, square, sqrt, sin, cos,
add3, mult3, cube, log, abs, and neg. These were utilized to
3) Artificial Bee Colony Algorithm (ABC) generate a function. The parameters for the programming
The artificial bee colony algorithm (ABC) is a swarm- process were set to 0.1, 0.84, and 0.14 for the ERC probability,
based metaheuristic algorithm that is primarily based on the crossover probability, and mutation probability respectively.
honeybees’ intelligent foraging behavior [19]. The algorithm The derived fitness functions were ranked based on their
can be explained on the principle of how bees coordinate with coefficient of determination (R2) in descending order. There
other bees to find the best possible food source they could were 500 models generated in this process. Seventeen models
find that is close to their beehive. The main components of were shown based on highest R2 values (model 351, 370, 380,
this algorithm are food sources, employed foragers, and 379, 378, 262, 276, 294, 257, 192, 185, 285,12, 37, 437, 72,
unemployed foragers (scout bees and onlooker bees). In and 87). Model 351 fitness function was the one chosen
selecting a food source, forager bees evaluate several among other models based on its R2 value of 0.939. The
equation presented below in (2) with the use of variable N

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
represents the number of neurons needed. Additionally, the lower than CSA-RNN3. Moreover, ABC-RNN3 has the
subscripts (1, 2, and 3) added under the variable N specify lowest MAE: 0.584% lower than GA-RNN3, and 9.415%
what layer is being identified. lower than CSA-RNN3. Among the developed models, CSA-
RNN3 performed the least while GA-RNN3 is the second-best
RMSE = 0.00407N1 + 0.0223N2 + 0.0262N3 + 9.42e- among the three algorithms in predicting the leaf region
6N32N22 - 0.00112N1N2 + 6.34e-4N1N3 - 0.00497N2N3 + infection rate (LRIR) of a Diplocarpon earlianum infected
1.98e-5N1N22 - 9.96e-6N2N32 - 8.49e-5N22N3 - 1.78e-6N23N3 strawberry leaf.
+ 4.24e-5N12 + 0.00466N22 - 9.86e-5N23 - 0.00521N32 +
6.46e-7N24 + 9.42e-6N34 - 1.63e-5N1 N2 N3 - 0.0209 (4)

Fig. 6. Fitness curve for GA-RNN optimization for the optimal three-
layered neural network combination

Fig. 5. Confusion matrix data summary for the testing phase of detecting
strawberry Diplocarpon Earlianum leaf infection three significant features
(a) SVM (b) NB (c) CTree (d) LDA

C. Recurrent Neural Network Configuration based on


Optimization Results
Based on the optimization results, GA-RNN has a three-
layered-hidden-layer combination of 95.9915, 32.3209,
13.9978 (rounded off to 96, 32, 14), CSA-RNN has a three-
layered-hidden-layer combination of 109.9901, 40, 15.2434
(rounded off to 110, 40, 15), and ABC-RNN has a three-
layered-hidden-layer combination of 96.0986, 33.3731, Fig. 7. Fitness curve for CSA-RNN optimization for the optimal three-
16.4391 (rounded off to 96, 33, 16). The fitness value for the layered neural network combination
GA-RNN is -0.0017, CSA-RNN has -0.5729, and ABC-RNN
has -0.5605. The fitness curve of each algorithm is presented
in Fig. 6, Fig. 7, and Fig. 8.
D. Strawberry Diplocarpon Earlianum Leaf Region
Infection Prediction using Computational Intelligence
A three-layered recurrent neural network (RNN3) is
responsible for the quantitative prediction of the amount of
prediction. The subscript of three indicates the number of
layers used inside this type of neural network. Based on Table
2, it can be noted that the ABC-RNN3 has the lowest MAE
and RMSE for the testing phase. It also has the highest R2
value which is closer to 1. It bested all other machine learning
models used in this study. To compare ABC-RNN3 has the
highest R2: 1.328% higher than GA-RNN3, and 3.834%
higher than CSA-RNN3. Additionally, ABC-RNN3 has the
lowest RSME: 11.238% lower than GA-RNN3, and 19.490% Fig. 8. Fitness curve for ABC-RNN optimization for the optimal three-
layered neural network combination

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
Fig. 9. Recurrent neural network model (a) GA-optimized RNN model (b) CSA-optimized RNN model (c) ABC-optimized RNN model

TABLE I. EVALUATION SUMMARY FOR STRAWBERRY SCORCHED LEAF CLASSIFICATION USING MACHINE LEARNING ALGORITHMS

Training Validation Testing


Model Fall- F1- Hamming Inference
Accuracy Accuracy Accuracy Precision Specificity Recall MCC
out score Loss Time (s)

SVM 100.000 100.000 100.000 0.000 1.000 1.000 1.000 1.000 1.000 0.000 44.058

NB 0.985 100.000 100.000 0.000 1.000 1.000 1.000 1.000 1.000 0.000 24.495

CTree 100.000 100.000 100.000 0.000 1.000 1.000 1.000 1.000 1.000 0.000 14.746

LDA 98.100 99.010 97.370 0.028 0.972 0.974 0.976 0.974 0.947 0.027 16.316

TABLE II. EVALUATION SUMMARY FOR PREDICTING THE STRAWBERRY LEAF REGION INFECTION RATE USING INTELLIGENT MODELS

Training Validation Testing


Model
RMSE R2 MAE RMSE R2 MAE RSME R2 MAE

GA-RNN3 1.990 1.000 3.728 0.037 1.000 0.005 2.274 0.936 0.919
CAS-RNN3 0.033 1.000 0.010 0.026 1.000 0.005 2.562 0.912 0.914
ABC-RNN3 0.000 1.000 0.000 0.029 1.000 0.005 2.063 0.948 0.828

Leaf scorch detection and infection rate prediction were to the ideal combination of three-layered hidden neurons for
explored to develop a vision-based approach to identifying a which ABC-RNN3 gave out the highest R2 value and lowest
strawberry leaf’s health condition and quantifying the amount RMSE and MAE values. Given the reports from [9] that
of severity the Diplocarpon earlianum fungus had caused. explored bio-inspired optimization with the firefly algorithm,
Strawberry leaf health status was classified as good or infected exploring these population-based optimizers opens more
with the help of machine learning models such as SVM, NB, options for innovation of computationally inexpensive
CTree, and LDA which are all found to be highly accurate. intelligent models that plant pathologists can consider when it
These machine learning models were also examined by [5] to comes to quantitative assessment of leaf infection rate in
classify the leaf health status of grapes infected with black rot; strawberry leaves. Extracting and selecting the top three
[6] on the other hand, used these models to identify a corn leaf features which are R, G, and a* proves that among
that is exposed to Maize Cercospora leaf spot; and [7] to morphological, textural, and spectral traits, the most
classify three different rice leaf diseases. Among the four significant leaf phenotype that is directly relevant in
classification algorithms, CTree showed a 100% accuracy for determining a scorch-infected strawberry leaf is the spectral
this study. Using population-based optimizers such as GA, feature. Using the combination of NCA-PCA, which was also
CSA, and ABC, the recurrent neural network was calibrated explored by [7, 13], for a contrast-improved strawberry leaf

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
image, the accuracy of the classification results had drastically the De La Salle University - Intelligent Systems Laboratory
improved while the computational cost was lessened, which is (DLSU-ISL), Manila, Philippines.
beneficial to farmers looking to improve their farming
practices without spending too much money. Reducing the REFERENCES
features needed for the model development significantly [1] S. Kaur, S. Pandey, and S. Goel, "Plants Disease Identification and
decreases the computational time and cost as observed by the Classification Through Leaf Images: A Survey", Archives of
results comparing an eighteen-feature vector to a three-feature Computational Methods in Engineering, 2018,
https://doi:10.1007/s11831-018-9255-6
vector [8], an eighteen-feature vector to a two-feature vector
[2] G. Dhingra, V. Kumar and H. D. Joshi, "Study of digital image
[13], and a twenty-two-feature vector to a seven-feature vector processing techniques for leaf disease detection and classification,"
[14] respectively. Thus, the NCA-PCA-CTree3-ABC-RNN3 Multimedia Tools and Applications, 2017, https://doi:10.1007/s11042-
hybrid algorithm configuration provides a cost-effective 017-5445-8
approach for strawberry leaf scorch detection and infection [3] A. Cruz et. al., "Vision-Based Plant Disease Detection System Using
rate quantitative assessment which can aid farmers in Transfer and Deep Learning," American Society of Agricultural and
detecting the crop disease at an earlier stage which will Biological Engineers (ASABE) Annual International Meeting,
Spokane, Washington, July 2017, https://doi:
significantly reduce widespread bacterial infection across the 10.13031/aim.201700241
farm. Additionally, it will also give plant pathologists an idea [4] C. Sass, “The health benefits of Strawberries”, January 8, 2020,
of how the infection should be treated given that the amount Available: Health, https://www.health.com/nutrition/health-benefits-
of infection severity is numerically determined. of-strawberries [Accessed: March 1, 2022]
[5] O. J. Alajas, R. Concepcion, E. Dadios, E. Sybingco, C. H. Mendigoria,
IV. CONCLUSION and H. Aquino, "Prediction of Grape Leaf Black Rot Damaged Surface
Percentage Using Hybrid Linear Discriminant Analysis and Decision
This study proposed a new technique for detecting and Tree," 2021 International Conference on Intelligent Technologies
predicting the Diplocarpon earlianum infection rate of (CONIT), 2021.
strawberry leaves using recurrent neural network (RNN) [6] R. Concepcion, E. Dadios, J. Alejandrino, C. H. Mendigoria, H.
integrated with population-based optimizers, namely genetic Aquino, and O. J. Alajas, "Diseased Surface Assessment of Maize
algorithm (GA), and cuckoo search algorithm (CSA), and Cercospora Leaf Spot Using Hybrid Gaussian Quantum-Behaved
artificial bee colony (ABC). The raw images containing Particle Swarm and Recurrent Neural Network," 2021 IEEE
International IoT, Electronics and Mechatronics Conference
scorched strawberry leaves were segmented using graph-cut (IEMTRONICS), 2021.
segmentation through lazy snapping algorithm. Hybrid
[7] C. H. Mendigoria, R. Concepcion, A. Bandala, O. J. Alajas, H. Aquino
neighborhood component analysis (NCA) and principal and E. Dadios, "OryzaNet: Leaf Quality Assessment of Oryza sativa
component analysis (PCA) was done to select the 3 most Using Hybrid Machine Learning and Deep Neural Network," 2021
significant morpho-spectro-textural features of leaves IEEE 13th International Conference on Humanoid, Nanotechnology,
resulting in red, green, and a* components. A genetic Information Technology, Communication and Control, Environment,
and Management (HNICEM), 2021, pp. 1-6, doi:
programming-based RMSE model was constructed as a 10.1109/HNICEM54116.2021.9731957.
function of neuron density on each of the three RNN layers.
[8] R. Concepcion, E. Dadios, J. Cuello, A. Bandala, E. Sybingco, and R.
After a series of explorations in the hyperparameters of GA- R. Vicerra, "Determination of Aquaponic Water Macronutrient
RNN, CSA-RNN, and ABC-RNN, the ABC-RNN model with Concentrations Based on Lactuca Sativa Leaf Photosynthetic
inputs of the 3 most significant features outperformed other Signatures using Hybrid Gravitational Search and Recurrent Neural
models in performing non-invasive leaf region rate infection Network," Walailak Journal of Science and Technology (WJST), vol.
18, no. 10, 2021.
with an accuracy of 94.8%. Four classification machine
learning models were configured in which the classification [9] R. Concepcion and E. Dadios, "Bioinspired Optimization of
Germination Nutrients Based on Lactuca sativa Seedling Root Traits
tree (CTree) bested other detection models with an accuracy as Influenced by Seed Stratification, Fortification and Light
of 100%. Hence, this study was able to introduce a technique Spectrums," AGRIVITA Journal of Agricultural Science, vol. 43, no.
of integrating NCA, PCA, CTree, ABC and RNN for crop 1, pp. 222-232, 2021.
disease detection and infection rate prediction. For plant [10] R. Concepcion, S. Lauguico, J. Alejandrino, J. de Guia, E. Dadios, and
disease assessment, it was proven to be more efficient and A. Bandala, "Aquaphotomics determination of total organic carbon and
labor cost-effective than manual disease inspection methods. hydrogen biomarkers on aquaponic pond water and concentration
prediction using genetic programming," 2020 IEEE 8th R10
For future studies, the inclusion of other strawberry leaf Humanitarian Technology Conference (R10-HTC), 2020.
disease variants caused by other fungal bacteria is highly [11] R. Concepcion, S. Lauguico, R. R. Tobias, E. Dadios, A. Bandala, and
recommended to expand the range of diseases that can be E. Sybingco, "Genetic algorithm-based visible band tetrahedron
determined by the model. Moreover, applying the developed greenness index modeling for lettuce biophysical signature
model to an actual strawberry farm for leaf health analysis as estimation," 2020 IEEE REGION 10 CONFERENCE (TENCON),
extended research to furtherly improve the model’s efficiency 2020.
is also suggested. [12] J. Alejandrino, R. Concepcion, S. Lauguico, R. R. Tobias, V. J.
Almero, J. C. Puno, A. Bandala, E. Dadios, and R. Flores, "Visual
classification of lettuce growth stage based on morphological attributes
ACKNOWLEDGMENT using unsupervised machine learning models," 2020 IEEE REGION 10
The authors would like to express their sincerest gratitude CONFERENCE (TENCON), 2020.
for the support bestowed by Engineering Research and [13] C.H. Mendigoria, R. Concepcion II, E. Dadios, H. Aquino. O.J. Alajas
Development for Technology (ERDT) of the Department of and E. Sybingco, “Vision-based Postharvest Analysis of Musa
Acuminata Using Feature-based Machine Learning and Deep Transfer
Science of Technology (DOST) of the Philippines as well as Networks,” 9th IEEE Region 10 Humanitarian Technology Conference
(R10 HTC 2021), in press

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.
[14] C. H. Mendigoria, H. Aquino, O. J. Alajas, R. Concepcion, E. Dadios, [17] S. H. Javaheri, M. M. Sepehri, and B. Teimourpour, “Chapter 6 -
E. Sybingco, A. Bandala, and R. R. Vicerra, “Varietal classification of Response Modeling in Direct Marketing: A Data Mining-Based
Lactuca Sativa seeds using an adaptive neuro-fuzzy inference system Approach for Target Selection,” in Data Mining Applications with R,
based on morphological phenes,” Journal of Advanced Computational Academic Press, Pages 153-180, 2014, ISBN 9780124115118,
Intelligence and Intelligent Informatics, Vol. 25, Issue 5, pp. 618-624, https://doi.org/10.1016/B978-0-12-411511-8.00006-2.
September 2021. [18] X. -S. Yang and Suash Deb, "Cuckoo Search via Lévy flights," 2009
[15] O. J. Alajas, R. Concepcion, R. R. Vicerra, A. Bandala, E. Sybingco, World Congress on Nature & Biologically Inspired Computing
E. Dadios, J. Cuello and V. Fonseca, "Indirect Prediction of Aquaponic (NaBIC), 2009, pp. 210-214, DOI: 10.1109/NABIC.2009.5393690.
Water Nitrate Concentration Using Hybrid Genetic Algorithm and [19] D. Karaboga and B. Basturk, “A powerful and efficient algorithm for
Recurrent Neural Network" 13th International Conference on numerical function optimization; artificial bee colony (ABC)
Humanoid, Nanotechnology, Information Technology, algorithm,” Journal of Global Optimization, 39(3), pp. 459-471 (2007).
Communication, and Control, Environment, and Management https://doi.org/10.1007/s10898-007-9149-x.
(HNICEM) 2021.
[16] J. Arun Pandian and G. Geetharamani, “Data for Identification of Plant
Leaf Diseases Using a 9-layer Deep Convolutional Neural Network”,
Mendeley Data, V1, 2019.

Authorized licensed use limited to: Heriot-Watt University. Downloaded on December 19,2023 at 13:42:44 UTC from IEEE Xplore. Restrictions apply.

You might also like