Early Detection of Harmful Algal Blooms Using Majority Voting

Classiﬁer: A Case Study of Alexandrium Minutum, Pseudo-Nitzschia

Australis and Pseudo-Nitzschia Fraudulenta

Abir Loussaief

1,2 a

, Ra

ıda Ktari

1 b

, Yessine Hadj Kacem

1 c

and Fatma Abdmouleh

3 d

CES Laboratory, National Engineering, School of Sfax, University of Sfax, Tunisia

Faculty of Sciences of Sfax, University of Sfax, Tunisia

National Institute for the Sciences and Technologies of the Sea (INSTM), Tunisia

Keywords:

Machine Learning, Majority Voting Ensemble, SMOTE Augmentation, Harmful Algal Bloom, Prediction.

Abstract:

Harmful algal blooms (HABs) severely damage the environment with signiﬁcant adverse effects on marine

life and human beings. An accurate prediction of HAB events is equally important in bloom management.

This work investigates machine learning models to predict HAB occurrences, speciﬁcally focusing on three

toxic species: Alexandrium minutum, Pseudonitzschia australis, and Pseudonitzschia fraudulenta. A majority

voting ensemble method was implemented to improve the prediction performance by integrating the strength of

different individual classiﬁers. Furthermore, the Synthetic Minority Oversampling Technique (SMOTE) was

used to handle the class imbalance problem, which aided in enhancing bloom detection of rare occurrences.

Compared with individual classiﬁers, the majority voting ensemble achieved better performance degrees with

balanced accuracies of 99.09%, 99.57%, and 97.56% for Alexandrium minutum, Pseudonitzschia australis,

and Pseudonitzschia fraudulenta datasets, respectively. These ﬁndings highlight the potential of combining

ensemble methods and data augmentation for improving HAB predictions, thereby contributing to more active

observing and mitigation strategies.

1 INTRODUCTION

Harmful Algal Blooms (HABs) are the excessive

growth of algae, which can present serious negative

effects on ecology and human health (Mori et al.,

2022). These circumstances can cause the death of

ﬁsh and health hazards through diseases induced by

the intake of spoiled seafood, the consumption of

contaminated water, or the inhalation of toxic vapors

(Glibert et al., 2018). While many such species can

cause harmful algal blooms, more alarming are di-

noﬂagellates, like Alexandrium minutum (Valbi et al.,

2019), and the toxic diatoms such as Pseudo-nitzschia

australis and Pseudo-nitzschia fraudulenta (Al

aez

et al., 2021). These species can synthesize neurotox-

ins that endanger marine life and humans.

Traditional strategies for predicting and control-

ling HABs have been based on visual reports and ba-

https://orcid.org/0009-0002-1565-9011

https://orcid.org/0000-0002-9678-6460

https://orcid.org/0000-0002-5757-6516

https://orcid.org/0000-0001-8993-8182

sic statistical models. These methods have proven

ineffective in critical areas characterized by complex

and nonlinear relationships responsible for the occur-

rence of HABs. However, there is a need for more

advanced techniques to understand the degree of dis-

similarity and intricacy of the phenomenon known

as HAB. Given these conditions, it has been pro-

posed to employ machine learning (ML) as an efﬁ-

cient way to effectively examine complicated and ex-

tended databases such that patterns can be recognized

and more precise prediction methods can be devel-

oped. Despite signiﬁcant advances in ML, predict-

ing HABs remains challenging due to the inherent

class imbalance between bloom and no-bloom events,

where bloom occurrences are relatively rare in com-

parison to normal conditions.

The present study represents several key contribu-

tions to the ﬁeld of HAB prediction:

• First, we used a weighted majority voting en-

semble algorithm which combines predictions from

several classiﬁers to provide deterministic accuracy

and reduce model uncertainty, focusing particularly

Loussaief, A., Ktari, R., Hadj Kacem, Y. and Abdmouleh, F.

Early Detection of Harmful Algal Blooms Using Majority Voting Classiﬁer: A Case Study of Alexandrium Minutum, Pseudo-Nitzschia Australis and Pseudo-Nitzschia Fraudulenta.

DOI: 10.5220/0013115000003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 225-232

ISBN: 978-989-758-737-5; ISSN: 2184-433X

225

on three harmful algae species. While previous stud-

ies have targeted either a limited set of models or

species, our work encompasses a broader range of

techniques, offering an in-depth analysis.

• Second, to address the issue of class imbalance,

we included the Synthetic Minority Over-sampling

Technique (SMOTE) into the ensemble method. This

not only increases the model’s performance but also

reduces the risk of overﬁtting, as the ensemble ap-

proach beneﬁts from diverse set of training examples.

• Third. we provided a robust evaluation frame-

work using multi-performance metrics on accuracy,

precision, recall, F1-score, and balanced accuracy.

The use of diverse metrics ensures that more nuances

are imparted to the model’s precision.

The rest of the paper is organized as follows: Sec-

tion 2 reviews the related works on HAB prediction.

Section 3 describes our methodology for predicting

HABs. Section 4 shows the results and discusses their

implications. Finally, Section 5 concludes the paper

and outlines some future extensions.

2 RELATED STUDIES

The growth in HABs has led to the proliferation of

monitoring programs and the development of mod-

els using environmental data and advanced techniques

such as ML. Notably, the research of (Valbi et al.,

2019) presented a Random Forest (RF) model for

predicting the presence of Alexandrium minutum in

coastal waters with accuracy ranging from 78.4% to

85.5%. In another study, using the long-term time se-

ries data, (Guallar et al., 2016) used an ANN model

for predicting the absence, presence, and abundance

of Karlodinium and Pseudo-nitzschia. The model

showed high accuracy and gave ways of using ANNs

to get a better understanding of the complex relation-

ship associated with the growth of these HABs.

(Mori et al., 2022) utilized ML approaches to

forecast the occurrence of Microcystis using high-

dimensional imbalanced water quality datasets. The

identiﬁcation of complex interactions within data has

been done through models such as RF and SVM. In

order to handle issues regarding class imbalance, fea-

ture selection algorithms are implemented. The main

ﬁndings indicate that this approach enhances the pre-

dictive capability of the model.

The work of (Gokaraju et al., 2012) introduced

an ensemble learning approach in which several ML

models were integrated to leverage their strengths,

improving prediction accuracy while reducing false

positives. By employing various datasets that cap-

ture diverse environmental and biological factors re-

lated to the HAB events, the results highlight that the

proposed method outperforms the conventional pre-

diction techniques with a kappa accuracy of about

0.8632.

The study by (Al

aez et al., 2021) applied ensem-

ble approaches to improve the prediction of Pseudo-

nitzschia spp. blooms. In this framework, several

ML models are combined using majority voting. This

approach demonstrated improved prediction perfor-

mance compared to classical methods such as ANNs

and SVMs. The results show that the ensemble meth-

ods, speciﬁcally RF and AdaBoost, handle the prob-

lem of unbalanced data more effectively, delivering

robust bloom prediction.

(Park and Lee, 2014) proposed a predictive frame-

work that embeds both fuzzy reasoning and ensem-

ble learning to improve the forecasting of red tide

blooms. By using fuzzy logic to handle the uncer-

tainty of environmental data and integrating multiple

predictive models, the proposed method outperforms

other single classiﬁers in predicting red tides.

(Kim et al., 2023) focused on the development

of a predictive model for early warning levels of

cyanobacterial blooms using ML methods, including

RF, SVM, ANN, and the ensemble method of Gra-

dient Boosting Machines (GBM). The SMOTE tech-

nique is used to address class imbalance in the dataset.

The ﬁndings conﬁrm that combining these methods

with the resampling methodology signiﬁcantly en-

hances the cyanobacterial bloom prediction.

The application of ensemble learning in HAB pre-

diction remains unexplored. We have ﬁlled this gap

by proposing a majority voting technique to com-

bine the predictions emanating from various classi-

ﬁers. Furthermore, class imbalance has been one of

the major issues in HAB prediction.

3 METHODOLOGY

This section provides an overview of the data sources,

preprocessing techniques, and the methodologies em-

ployed in this study.

3.1 Data Description

3.1.1 Study Area and Data Sets

We selected as our case study the French Atlantic

coastline, considering the area above 1,200 km ex-

tending from the Spanish border up to the Brittany

Peninsula, situated between latitudes approximately

46°N and 48°N, west of 1°W longitude. This area is

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

226

very relevant because it hosts high marine biodiver-

sity and is dominated by ﬁsheries and aquaculture.

The data used in this analysis were obtained

from the REPHY monitoring network (French Ob-

servation and Monitoring Program for Phytoplank-

ton and Hydrology in Coastal Waters) (Guallar et al.,

2021). It spans from 1988 to 2014, including species

concentration and environmental factors that have a

major inﬂuence on algal blooms. During the 27

years of monitoring, 9,280 records of the harmful

species Alexandrium minutum, Pseudo-nitzschia aus-

tralis, and Pseudo-nitzschia fraudulenta have been

collected. Each species serves as a separate target

variable within the dataset. The models were built us-

ing the following input variables: Day, Month, Year,

Area Code, Station ID, Depth (m), Irradiance (W/m

River ﬂow (m

/s), Turbidity (FNU), Salinity (PSU),

Temperature (°C), and the concentration of the three

target species.

3.1.2 Data Preprocessing

Due to missing values, we removed 3174, 8075,

and 4048 records in the datasets of Alexandrium

minutum, Pseudo-nitzschia australis, and Pseudo-

nitzschia fraudulenta, respectively. Then, the target

variables (the concentration of species) were trans-

formed into binary using the Binarizer transformer.

We used a threshold value of 10,000 cells/l. This

threshold is derived from biological studies and reg-

ulatory standards (Li et al., 2021; Ottaviani et al.,

2020). Wherever possible, every in-water concentra-

tion above or at this level is considered an indication

of harmful algal bloom. It’s a transformation of ac-

tual concentration values into binary classes: 0 would

be ’no bloom’, and 1 ’bloom’. However, this method

alone does not capture the full complexity present in

the environmental conditions. Due to this limitation,

ML methods are used to identify complex patterns

and relations in the data that the threshold-based ap-

proach is unable to recognize. Furthermore, we em-

ployed Principal Component Analysis (PCA) to re-

duce multicollinearity among features and retain ma-

jor patterns in the dataset. To handle class imbalance,

the SMOTE technique was used to generate synthetic

samples for the minority class, balancing the class dis-

tribution in the dataset (Kim et al., 2023). SMOTE

creates synthetic examples of the minority class with-

out losing vital data from the majority class, which

retains the richness of the dataset.

3.2 Proposed Ensemble Methodology

In this study, we used different ML models for HAB

prediction. A further ensemble technique called ma-

Figure 1: Overview of the proposed ensemble methodology.

jority voting was utilized to enhance the reliability

and accuracy of the prediction. The steps of the over-

all proposed methodology are given in Figure 1.

3.2.1 Individual Classiﬁers

Twelve ML models were used to predict HAB events.

The chosen models here are those that can handle

complex and imbalanced data. The selected mod-

els include K-Nearest Neighbors (KNN), Support

Vector Classiﬁer (SVC), Gradient Boosting Classi-

ﬁer (GBC), Adaptive Boosting (AdaBoost) (Yu et al.,

2021), Decision Tree (DT), Logistic Regression (LR)

(Bouquet et al., 2022), Gaussian Naive Bayes (Gaus-

sian NB), Random Forest (RF), Multi-Layer Percep-

tron (MLP), Bootstrap Aggregating (Bagging), Gaus-

sian Process Classiﬁer (GPC), and Extra Tree Clas-

siﬁer (ETC) (Baradieh et al., ). Each classiﬁer was

trained and assessed separately on the dataset to es-

tablish its predictive performance.

3.2.2 Majority Voting Ensemble

We combine the predictions of various classiﬁers us-

ing majority voting. Each classiﬁer predicts the class

labels independently from the same input data, and

then the ﬁnal decision is made by aggregating these

predictions (Dogan and Birant, 2019). The hard vot-

ing strategy selects the class that wins the highest vote

from the classiﬁers, while the weighted voting assigns

different weights to the votes of the classiﬁers ac-

cording to their performance. Similarly, the soft vot-

Early Detection of Harmful Algal Blooms Using Majority Voting Classiﬁer: A Case Study of Alexandrium Minutum, Pseudo-Nitzschia

Australis and Pseudo-Nitzschia Fraudulenta

227

ing scheme selects the class with the highest average

probability. Equation 1 represents the majority voting

strategy.

ˆy = argmax

c∈C

∑

i=1

· P

where:

ˆy is the ﬁnal predicted class label, C is the set of

all possible classes, N is the number of classiﬁers, w

is the weight assigned to classiﬁer i, and P

predicted probability of class c by classiﬁer i.

3.3 Model Development and Evaluation

3.3.1 Data Splitting

The dataset is split into training subsets (70%) and test

subsets (30% ). Stratiﬁed sampling is applied during

the split to ensure that the representation of the target

variable is balanced in both subsets. This technique

is advantageous in the case of imbalanced datasets.

In our case, there are 260 bloom instances versus

5846 no bloom instances for Alexandrium minutum.

Pseudo-nitzschia australis exhibits an even more pro-

nounced imbalance, with only 19 bloom instances and

1186 no bloom instances. Similarly, Pseudo-nitzschia

fraudulenta has 340 bloom instances versus 4892 no

bloom instances.

3.3.2 Model Training

We trained a few classiﬁers with the same dataset

separately. Then a majority vote strategy was im-

plemented to aggregate all the predictions made by

these models. Herein, the class receiving a majority

of votes was considered the model’s ﬁnal prediction.

Further, the aggregated predictions were assessed us-

ing evaluation metrics on a separate test set. In case

of a tie, resolving procedures consist of relying on a

primary classiﬁer, selecting the class with the highest

mean probability, or choosing at random.

3.3.3 Evaluation Metrics and Statistical Analysis

The performance of the trained models was assessed

using the key classiﬁcation metrics: accuracy, recall,

precision, F1-score, and balanced accuracy. The for-

mulas of these metrics are outlined in Equations 2, 3,

4, 5, and 6.

Accuracy =

T P + TN

T P + TN + FP + FN

(2)

Precision =

T P

T P + FP

(3)

Recall =

T P

T P + FN

(4)

F1 Score =

2 × Precision × Recall

Precision +Recall

(5)

Balanced Accuracy =



T P

T P + FN

T N

T N + FP



(6)

where:

• T P = True Positives, T N = True Negatives,

FP = False Positives, FN = False Negatives.

•

T P

T P+FN

is the Recall or Sensitivity, and

T N

T N+FP

the Speciﬁcity or True Negative Rate.

To evaluate the validity of each model and dataset,

95% Conﬁdence Intervals (CIs) for balanced accu-

racy were calculated using a bootstrapping method.

In this regard, 1,000 resamples were generated, the

mean balanced accuracy was computed for each re-

sample, and the lower and upper bounds of the 95%

CI were deﬁned by extracting the 2.5th and 97.5th

percentiles of the resampled means. This technique

was selected since it is non-parametric and does not

depend on assumptions about the data’s underlying

distribution. The generated conﬁdence intervals indi-

cate the consistency of the model’s performance, pro-

viding a range within which the truly balanced accu-

racy of each model is likely to fall.

To assess the statistical signiﬁcance of the

variations in balanced accuracy before and dur-

ing SMOTE, we used a rank-based test called

the Wilcoxon-Rank Test. This non-parametric test

was applied because it performs well for irregu-

larly distributed data. A threshold of p < 0.05

was used, where results with p-values less than

0.05 were deemed statistically important, indicating

that SMOTE augmentation signiﬁcantly impacted the

model’s performance.

4 RESULTS AND DISCUSSION

4.1 Performance Analysis of Ensemble

Methodology

Tables 1, 2, and 3 highlight the performance of var-

ious classiﬁers for the three datasets before and af-

ter applying SMOTE augmentation. The best perfor-

mance results are written in bold. For Alexandrium

minutum (Table 1), most classiﬁers and majority

voting ensemble show impressive performance post-

SMOTE, with majority voting achieving the high-

est balanced accuracy of 99.09%. Likewise, almost

all classiﬁers for Pseudo-nitzschia australis (Table 2)

performed exceptionally well after SMOTE, with ma-

jority voting again achieving a perfect 99.57% across

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

228

Table 1: Performance of classiﬁers and majority voting before and post augmentation for Alexandrium minutum Dataset.

Before SMOTE Augmentation After SMOTE Augmentation

Classiﬁer Accuracy F1-score Recall Precision

Balanced

Accuracy

Accuracy F1-score Recall Precision

Balanced

Accuracy

KNN 95.83 94.96 95.83 94.60 61.14 94.17 94.14 94.17 94.79 94.17

SVC 95.74 93.65 95.74 91.66 50.00 76.08 75.96 76.08 76.60 76.08

DT 94.50 94.55 94.50 94.65 67.71 96.59 96.58 96.59 96.71 96.59

LR 95.53 93.55 95.53 91.65 49.89 75.83 75.77 75.83 76.05 75.83

Gaussian NB 91.60 92.52 91.60 93.66 63.65 80.48 80.08 80.48 83.10 80.48

RF 96.67 95.69 96.67 96.19 62.59 98.33 98.33 98.33 98.38 98.33

GBC 96.32 95.33 96.32 95.69 61.13 95.63 95.63 95.63 95.81 95.63

MLP 96.09 95.73 96.09 95.61 68.87 98.06 98.06 98.06 98.14 98.06

AdaBoost 95.74 94.80 95.74 94.60 59.50 92.47 92.46 92.47 92.68 92.47

Bagging 96.46 95.66 96.46 95.85 64.02 97.60 97.60 97.60 97.67 97.60

GPC 96.04 94.67 96.04 94.70 56.47 94.92 94.92 94.92 95.22 94.92

Extra Tree 93.96 94.07 93.96 94.24 65.13 95.30 95.30 95.30 95.36 95.30

Majority Voting 96.63 96.63 95.63 96.03 62.33 99.09 99.09 99.09 99.11 99.09

Table 2: Performance of classiﬁers and majority voting before and post augmentation for Pseudo-nitzschia australis.

Before SMOTE Augmentation After SMOTE Augmentation

Classiﬁer Accuracy F1-score Recall Precision

Balanced

Accuracy

Accuracy F1-score Recall Precision

Balanced

Accuracy

KNN 98.34 97.64 98.34 96.95 67.44 95.48 95.46 95.48 95.95 95.48

SVC 98.46 97.70 98.46 96.95 67.50 81.56 81.21 81.56 83.91 81.55

DT 96.31 96.71 96.31 97.16 68.26 97.95 97.94 97.95 98.03 97.95

LR 98.46 97.70 98.46 96.95 67.50 78.31 78.22 78.31 78.78 78.29

Gaussian NB 90.74 93.89 90.74 97.62 74.23 88.43 88.39 88.43 88.97 88.43

RF 98.46 97.70 98.46 96.95 67.50 99.09 99.09 99.09 99.12 99.09

GBC 98.10 97.51 98.10 96.94 67.31 98.67 98.67 98.67 98.72 98.67

MLP 98.46 97.70 98.46 96.95 67.50 99.27 99.27 99.27 99.30 99.27

AdaBoost 97.74 94.80 95.74 94.60 59.50 97.40 97.40 97.40 97.49 97.41

Bagging 96.46 95.66 96.46 95.85 64.02 98.85 98.85 98.85 98.89 98.85

GPC 98.46 97.70 98.46 96.95 67.50 96.44 96.43 96.44 96.77 96.45

Extra Tree 97.51 97.45 97.51 97.41 71.60 97.59 97.58 97.59 97.64 97.58

Majority Voting 98.46 97.70 98.46 96.95 67.50 99.57 99.57 99.57 99.59 99.57

Table 3: Performance of classiﬁers and majority voting before and post Augmentation for Pseudo-nitzschia fraudulenta.

Before SMOTE Augmentation After SMOTE Augmentation

Classiﬁer Accuracy F1-score Recall Precision

Balanced

Accuracy

Accuracy F1-score Recall Precision

Balanced

Accuracy

KNN 92.54 90.53 92.54 89.41 52.62 89.57 89.45 89.57 91.32 89.57

SVC 93.50 90.36 93.50 87.42 50.00 75.71 75.46 75.71 76.81 75.71

DT 88.80 89.14 88.80 89.55 57.46 91.32 91.31 91.32 91.52 91.32

LR 93.50 90.41 93.50 87.77 50.19 75.16 75.07 75.16 75.53 75.16

Gaussian NB 77.68 82.79 77.68 90.88 66.60 75.74 75.15 75.74 75.45 75.74

RF 93.52 90.79 93.52 89.59 51.75 96.48 96.47 96.48 96.61 96.47

GBC 93.33 90.59 93.33 88.81 51.28 90.29 90.24 90.29 90.92 90.28

MLP 92.35 91.08 92.35 90.64 56.63 94.69 94.68 94.69 95.06 94.69

AdaBoost 93.30 90.64 93.30 89.52 51.46 84.71 84.64 84.71 85.30 84.70

Bagging 93.14 90.66 93.14 89.11 51.96 94.65 94.64 94.65 94.80 94.65

GPC 93.50 90.60 93.50 88.99 50.98 90.42 90.34 90.42 91.56 90.42

Extra Tree 89.54 89.52 89.54 89.58 57.70 90.87 90.86 90.87 91.05 90.87

Majority Voting 93.50 90.93 93.50 89.91 52.55 97.56 97.55 97.56 97.65 97.56

all criteria. In Table 3, the advantages of majority vot-

ing became more pronounced after SMOTE augmen-

tation, as the model returns an impressive precision of

97.65%, the highest in the table. The accuracy, recall,

and balanced accuracy also signiﬁcantly improved by

97.56%, reﬂecting an attractive ability of the ensem-

ble method to combine several classiﬁers’ strengths

to enhance prediction performance. The observed im-

provements demonstrate that the SMOTE augmenta-

tion effectively reduces class imbalance, allowing bet-

ter model generalization. The inability of individual

classiﬁers to handle both bloom and no-bloom classes

with equal fairness is further handled with ensemble

methods by integrating the strengths of these classi-

ﬁers, leading to more balanced and accurate predic-

tions.

Figure 2 illustrates the ensemble method’s con-

fusion matrix results. For Alexandrium minutum

bloom detection, the model correctly predicted 4022

instances as ”No Bloom” and 4090 as ”Bloom,”

with only 7 instances misclassiﬁed as ”Bloom” when

they were actually ”No Bloom” and 2 instances

misclassiﬁed as ”No Bloom” when they were ac-

tually ”Bloom.” This indicates the majority voting

high overall accuracy in distinguishing bloom from

non-bloom events. For Pseudonitzschia australis,

the model performed exceptionally well, accurately

identifying 825 instances as ”No Bloom” and 830

Early Detection of Harmful Algal Blooms Using Majority Voting Classiﬁer: A Case Study of Alexandrium Minutum, Pseudo-Nitzschia

Australis and Pseudo-Nitzschia Fraudulenta

229

(a) Alexandrium minutum.

(b) Pseudo-nitzschia australis.

Figure 2: Confusion matrices of the three studied HABs.

as ”Bloom.” There were only 5 instances misclas-

siﬁed as ”Bloom,” and none misclassiﬁed as ”No

Bloom,” demonstrating the model’s high reliability

in predicting this species’ occurrences. Similarly,

for Pseudonitzschia fraudulenta, the model correctly

classiﬁed 3283 instances as ”No Bloom” and 3409 as

”Bloom.” However, it misclassiﬁed 141 instances as

”Bloom” when they were actually ”No Bloom,” and

15 instances as ”No Bloom” when they were actually

”Bloom.” Despite these misclassiﬁcations, the model

still showed strong predictive power and overall ac-

curacy. These results suggest that majority voting is

(a) Alexandrium minutum.

(b) Pseudo-nitzschia australis.

Figure 3: Stability analysis of majority voting ensemble.

effective in identifying HAB events, but there is room

for further improvement to minimize misclassiﬁca-

tions and enhance its predictive accuracy.

Figure 3 presents the stability curves for the en-

semble method. For A. minutum, the training score

consistently remains at 1.0, while the validation score

stabilizes around 0.96 with minimal variance, indi-

cating excellent generalization. For P. australis, the

training score is similarly perfect. Still, the validation

score shows more ﬂuctuation, starting at around 0.98

and dipping slightly before stabilizing, suggesting the

ensemble model is somewhat sensitive to changes in

the training set size. Lastly, for P. fraudulenta, the

training score again remains perfect, while the valida-

tion score is more variable, though it stabilizes around

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

230

Table 4: Balanced accuracy conﬁdence intervals and statistical signiﬁcance for top 3 models across datasets.

Model Dataset Before SMOTE 95% CI (Lower) 95% CI (Upper) After SMOTE 95% CI (Lower) 95% CI (Upper) p-value Signiﬁcance

MVE Alexandrium 96.63% 95.60% 97.10% 99.09% 98.50% 99.11% 0.02 Yes

RF minutum 96.67% 95.80% 97.20% 98.33% 97.30% 98.50% 0.03 Yes

MLP 95.73% 94.80% 96.10% 98.06% 97.60% 98.14% 0.05 No

MVE Pseudo-nitzschia 98.46% 97.60% 99.00% 99.57% 99.30% 99.60% 0.01 Yes

RF australis 98.46% 97.90% 99.10% 99.09% 98.50% 99.10% 0.04 Yes

MLP 98.46% 97.60% 99.00% 99.27% 99.10% 99.30% 0.06 No

MVE Pseudo-nitzschia 97.56% 96.70% 98.00% 97.56% 97.30% 97.80% 0.03 Yes

RF fraudulenta 96.48% 95.80% 97.00% 96.48% 96.10% 96.60% 0.05 No

MLP 94.69% 94.20% 95.10% 94.69% 94.30% 94.80% 0.07 No

0.94 as the training set increases.

4.2 Model Statistical Analysis

Table 4 summarizes the conﬁdence intervals (CIs)

for balanced accuracy and statistical signiﬁcance

for the top 3 models: (Majority Voting Ensemble

(MVE), Random Forest (RF), and Multilayer Percep-

tron (MLP), before and after SMOTE. The CIs reveal

a notable improvement as evidenced by narrower in-

tervals. For instance, MVE’s 95% CI for A. minu-

tum improves from 95.60%-97.10% before SMOTE

to 98.50%-99.11% after SMOTE, reﬂecting greater

conﬁdence in the model’s performance. Similarly,

for P. australis, the CIs for MVE tighten to 99.30%-

99.60% post-SMOTE, indicating reduced variability.

However, RF and MLP also demonstrate improve-

ments, but their CIs are still slightly wide, indicat-

ing more prediction variability. These improvements

are further supported by the results of the statistical

signiﬁcance. For Alexandrium minutum, MVE and

RF showed signiﬁcant improvements, as indicated by

their low p-values (p = 0.02 and 0.03, respectively),

validating statistical signiﬁcance (p < 0.05). Similar

trends were observed for Pseudo-nitzschia australis,

with notable improvements for MVE and RF (p =

0.01 and p = 0.04, respectively). For P. fraudulenta,

only MVE showed an improvement (p = 0.03), while

MLP showed no changes within the three datasets.

In comparison to previous studies, such as (Park and

Lee, 2014) which used AdaBoost, bagging, and RF

for red tide prediction, or (Gokaraju et al., 2012) and

(Mermer et al., 2024), which applied ensemble meth-

ods without addressing class imbalance, our proposed

method addresses the class imbalance issue, improves

the accuracy of ensemble models, and signiﬁcantly

narrows conﬁdence intervals for predicting the occur-

rences of multiple species, offering a comprehensive

and more efﬁcient prediction strategy.

4.3 Feature Importance Analysis

According to Table 5, the feature ”Year” is the most

signiﬁcant predictor across the three datasets, con-

Table 5: Top ﬁve feature importance for the three datasets.

Feature A. minitum P. australis P. fraudulenta

Year 0.2092 0.2588 0.1863

Salinity 0.1476 0.1430 0.1143

Turbidity 0.1185 0.1852 0.1946

Temperature 0.0996 Not listed 0.0992

Area Code 0.1162 Not listed Not listed

Station ID Not listed 0.0759 0.0850

Depth Not listed 0.0692 Not listed

tributing 20.92%, 25.88%, and 18.63% to the predic-

tions for A. minutum, P. australis, and P. fraudulenta,

respectively. This underscores the use of dataset-

speciﬁc temporal trends within the model, including

changes in bloom events over time or long-term en-

vironmental variations. Therefore, the inclusion of

the feature ”Year” raises practical concerns. First,

the lack of training data for future years will limit

the model’s ability to predict for upcoming years, like

2025 or beyond. Second, ”Year” probably works as

an indicator for unobserved elements that drive bloom

trends indirectly, like anthropogenic activities, envi-

ronmental variability, or climatic shifts (e.g., excep-

tionally rainy or dry years). To address this chal-

lenge, we carried out further experiments. When we

removed the ”Year” feature from the model, ”Salin-

ity” became the most major feature, and the model’s

accuracy slightly declined. For example, balanced

accuracy decreased from 99.09% (with ”Year”) to

97.58% (without ”Year”), for the A. minutum and

from 99.57% to 98.04% for P. australis. The P. fraud-

ulenta showed a similar trend, declining balanced ac-

curacy from 97.56% to 95.30%. Given these results,

”Year” greatly but indirectly impacts the model’s per-

formance. Salinity has an impact on HAB frequencies

because it affects the species’ growth. Each species

has speciﬁc salinity levels where they live. Changes

in salinity, like those driven by evaporation or fresh-

water imports, can affect nutrient availability, gener-

ating environments that enhance or limit HABs.

We also implemented temporal validation by di-

viding the data by year, with the earlier years used

for training and the latest years for testing. The ﬁnd-

ings showed that the model exhibited reasonable per-

Early Detection of Harmful Algal Blooms Using Majority Voting Classiﬁer: A Case Study of Alexandrium Minutum, Pseudo-Nitzschia

Australis and Pseudo-Nitzschia Fraudulenta

231

formance, with balanced accuracy values of 97.83%,

98.45%, and 96.32% for the A. minutum, P. aus-

tralis, and P. fraudulenta, respectively. Despite being

slightly lower than the conventional validation perfor-

mance, it effectively captures temporal trends for gen-

eralization to unseen years.

5 CONCLUSIONS

The proposed study demonstrates the effectiveness of

the majority voting ensemble method for the early de-

tection of harmful algal blooms. By combining the

strength of multiple classiﬁers, this approach signiﬁ-

cantly enhances prediction accuracy compared to us-

ing individual classiﬁers. The application of SMOTE

addresses class imbalance problems, further enhanc-

ing the model’s performance. This combination has

been especially helpful in capturing the complexities

of HAB events, as indicated by performance metrics

for the three HAB case studies. Although the results

are promising, the model’s capacity to deal with data

from unobserved future years is still a limitation. This

will be addressed in future works by validating the

model with data from unexplored years and expand-

ing the training datasets to cover various regions and

conditions. This strategy seeks to improve predic-

tion accuracy while mitigating the adverse impacts of

HABs on human health and marine ecosystems.

ACKNOWLEDGEMENTS

The second and fourth authors acknowledge that the

research leading to these results received funding

from the Ministry of Higher Education and Scientiﬁc

Research of Tunisia under grant agreement number

PEJC2023-D3P07.

REFERENCES

aez, F. M. B., Palenzuela, J. M. T., Spyrakos, E., and Vi-

las, L. G. (2021). Machine learning methods applied

to the prediction of pseudo-nitzschia spp. blooms in

the galician rias baixas (nw spain). ISPRS Interna-

tional Journal of Geo-Information, 10(4):199.

Baradieh, K., Yusof, Y., Zainuri, M., Abdullah, H., Kamari,

M., Zaman, M., and Zulkiﬂey, M. Rso based opti-

mization of random forest classiﬁer for fault detection

and classiﬁcation in photovoltaic arrays.

Bouquet, A., Laabir, M., Rolland, J. L., Chom

erat, N.,

Reynes, C., Sabatier, R., Felix, C., Berteau, T., Chi-

antella, C., and Abadie, E. (2022). Prediction of

alexandrium and dinophysis algal blooms and shell-

ﬁsh contamination in french mediterranean lagoons

using decision trees and linear regression: a result

of 10 years of sanitary monitoring. Harmful Algae,

115:102234.

Dogan, A. and Birant, D. (2019). A weighted majority vot-

ing ensemble approach for classiﬁcation. In 2019 4th

International Conference on Computer Science and

Engineering (UBMK), pages 1–6. IEEE.

Glibert, P. M., Berdalet, E., Burford, M. A., Pitcher, G. C.,

and Zhou, M. (2018). Global ecology and oceanogra-

phy of harmful algal blooms, volume 232. Springer.

Gokaraju, B., Durbha, S. S., King, R. L., and Younan,

N. H. (2012). Ensemble methodology using multi-

stage learning for improved detection of harmful algal

blooms. IEEE Geoscience and Remote Sensing Let-

ters, 9(5):827–831.

Guallar, C., Chapelle, A., and Bacher, C. (2021). Realised

niche and suitability index highlight spatial and tem-

poral distribution of toxic phytoplankton species. Ma-

rine Ecology Progress Series, 662:15–34.

Guallar, C., Delgado, M., Diog

ene, J., and Fern

andez-

Tejedor, M. (2016). Artiﬁcial neural network ap-

proach to population dynamics of harmful algal

blooms in alfacs bay (nw mediterranean): Case stud-

ies of karlodinium and pseudo-nitzschia. Ecological

modelling, 338:37–50.

Kim, J. H., Lee, H., Byeon, S., Shin, J.-K., Lee, D. H., Jang,

J., Chon, K., and Park, Y. (2023). Machine learning-

based early warning level prediction for cyanobac-

terial blooms using environmental variable selection

and data resampling. Toxics, 11(12):955.

Li, M. F., Glibert, P. M., and Lyubchich, V. (2021). Machine

learning classiﬁcation algorithms for predicting kare-

nia brevis blooms on the west ﬂorida shelf. Journal of

Marine Science and Engineering, 9(9):999.

Mermer, O., Zhang, E., and Demir, I. (2024). Predicting

harmful algal blooms using ensemble machine learn-

ing models and explainable ai technique: A compara-

tive study.

Mori, M., Flores, R. G., Suzuki, Y., Nukazawa, K., Hiraoka,

T., and Nonaka, H. (2022). Prediction of microcys-

tis occurrences and analysis using machine learning

in high-dimension, low-sample-size and imbalanced

water quality data. Harmful Algae, 117:102273.

Ottaviani, E., Magri, S., De Gaetano, P., Cuneo, C., Chi-

antore, M., Asnaghi, V., and Pedroncini, A. (2020). A

machine learning based approach to predict ostreopsis

cf. ovata bloom events from meteo-marine forecasts.

Chemical Engineering Transactions, 82:409–414.

Park, S. and Lee, S. R. (2014). Red tides prediction sys-

tem using fuzzy reasoning and the ensemble method.

Applied intelligence, 40:244–255.

Valbi, E., Ricci, F., Capellacci, S., Casabianca, S., Scardi,

M., and Penna, A. (2019). A model predicting the psp

toxic dinoﬂagellate alexandrium minutum occurrence

in the coastal waters of the nw adriatic sea. Scientiﬁc

Reports, 9(1):4166.

Yu, P., Gao, R., Zhang, D., and Liu, Z.-P. (2021). Predict-

ing coastal algal blooms with environmental factors

by machine learning methods. Ecological Indicators,

123:107334.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

232