The Journal of Experimental Life Science p-ISSN: 2087-2852
e-ISSN: 2338-1655
RESEARCH ARTICLE
Hybrid VGG16AeLSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening Indah Yanti 1*.
Marjono1.
Antariksa2.
Andi Kurniawan3.
Syaiful Anam1.
Hilmi Aziz Bukhori1 1Department of Mathematics.
Faculty of Science.
Technology and Mathematics.
Universitas Brawijaya.
Malang, 65145.
Indonesia 2Department of Architecture.
Faculty of Engineering.
Universitas Brawijaya.
Malang, 65145.
Indonesia 3Department of Aquatic Resource Management.
Faculty of Fisheries and Marine Science.
Universitas Brawijaya.
Malang, 65145.
Indonesia *Correspondence:
Indah Yanti Email : indah_yanti@ub.
Address : Universitas Brawijaya.
Veteran Malang, 65145 Article History:
Submitted: 2026-02-18 Revised: 2026-06-11 Accepted: 2026-06-29 Available online: 2026-06-30 DOI: 10.
21776/ub.
Abstract While conventional culture-based, biochemical, and molecular identification methods remain fundamental in microbiology, computational image analysis can serve as an exploratory, supplementary tool for studying bacterial morphology.
This proof-of-concept study evaluates a hybrid VGG16-LSTM model for microscopic bacterial image classification.
This study utilized a small subset of the DIBaS dataset consisting of six bacterial classes: Acinetobacter baumannii.
Escherichia coli.
Lactobacillus plantarum.
Micrococcus spp.
Propionibacterium acnes, and Pseudomonas aeruginosa.
The total dataset size is highly constrained at 124 images, with a correspondingly small number of images per class ranging from 20 to 23.
The dataset was divided into training, validation, and testing subsets in a 70:20:10 ratio.
All microscopic images were resized to 224 y 224 pixels, normalized, and dynamically augmented during training to improve data variability under these limited-sample conditions.
A pre-trained VGG16 network was employed to extract spatial image features, and the final convolutional feature map was reshaped into a spatial sequence and processed using an LSTM layer for further feature learning.
Three optimization algorithms, namely Adam.
RMSprop, and stochastic gradient descent, were compared.
Among them, the RMSpropoptimized model exhibited the highest metrics after 50 epochs, achieving 92.
accuracy, 95.
24% precision, 92.
86% recall, and a 92.
38% F1-score.
however, these performance indicators must be interpreted with caution as they are based on a severely limited test set .
pproximately 12 image.
Some misclassifications occurred between P.
aeruginosa and E.
coli, which may be attributed to their shared gram-negative rod-shaped morphology.
This finding highlights the inherent biological challenge of distinguishing visually similar bacterial species from microscopic image analysis alone, underscoring that such models remain strictly experimental and are not suited for clinical-grade diagnosis or field-ready environmental screening.
Keywords: bacterial classification.
environmental monitoring.
VGG16-LSTM.
INTRODUCTION
Environmental pollution by petroleum hydrocarbons, pesticides, heavy metals, and other persistent pollutants remains a major ecological and public health concern.
Conventional remediation approaches can be expensive, energy-intensive, and may cause secondary environmental impacts.
Consequently, microbial bioremediation is increasingly recognized as a This microorganisms can transform, immobilize, degrade, or detoxify pollutants through diverse metabolic and enzymatic mechanisms .
this context, the ability to recognize and monitor bacteria associated with environmental samples is an important step in microbial screening, contamination assessment, and bioremediation Bacteria contribute to bioremediation through their rapid growth, metabolic diversity, adaptability to environmental stressors, and ability to interact with pollutants.
Genera such as How to cite: Yanti.
Marjono.
Antariksa.
Kurniawan.
Anam.
Bukhori.
Hybrid VGG16-LSTM classification of microscopic bacterial images for environmental microbioogy screening.
The Journal of Experimental Life Science, 16.
, 80-88.
DOI: 10.
21776/ub.
This is an open access article distributed under the Creative Commons Attribution 4.
0 International License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original works is properly cited.
A2026 The Author.
Exp.
Life Sci.
, 2026.
Vol.
No.
2, 80-88.
https://jels.
id/index.
php/jels Hybrid VGG16-LSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening (Yanti, et al.
Pseudomonas.
Acinetobacter, and Micrococcus transformation, hydrocarbon degradation, metal tolerance, and survival in contaminated environments .
,5,7-.
Other bacteria, including Escherichia coli, are important in environmental microbiology as indicators of fecal contamination and water quality risk .
Applied microbiology also includes bacteria relevant to biosorption, fermentation biotechnology, microbial safety, and comparative physiological studies, such as Lactobacillus plantarum and Cutibacterium acnes .
Despite the importance of bacterial identification, routine recognition remains challenging in laboratories with limited resources.
Culture-based and biochemical methods are fundamental but often require repeated subculturing, multiple phenotypic tests, and expert interpretation .
Molecular methods can improve taxonomic resolution, but they are often more costly and may not be available for preliminary screening in small labs.
These limitations create a need for complementary tools to support rapid, reproducible microscopic screening prior to confirmatory tests.
Digital microscopy offers a practical route for microscopic images contain information on the cell shape, arrangement, staining pattern, texture, and local spatial organization.
Manual interpretation of such images can be affected by subjectivity, observer fatigue, image quality, and similarity among the bacterial morphologies.
Therefore, automated image-based analysis has become increasingly relevant in microbiology, particularly for microorganism detection, interpretation .
Convolutional Neural Networks (CNN.
have been widely applied to microscopic image classification because they can learn spatial features directly from image data .
However, many machine learning studies adequately explaining the biological context of the selected organisms or the task's relevance to applied life sciences.
For a journal in experimental life science, an image-classification model should be positioned not only as an artificial intelligence method but also as a microbiological support tool with a clear biological rationale and realistic boundaries of interpretation.
Fine-grained classification is biologically challenging because Exp.
Life Sci.
, 2026.
Vol.
No.
2, 80-88.
different taxa may share similar cell shapes and staining appearances.
For example.
and E.
coli are Gram-negative rod-shaped bacteria that may display overlapping visual features under light microscopy .
A hybrid CNN-LSTM architecture may help address this challenge by combining CNN-based spatial feature extraction with LSTM-based learning over reshaped spatial feature sequences.
In this design.
VGG16 extracts local visual patterns, whereas the LSTM layer models dependencies across the sequence of feature vectors generated from the convolutional feature map .
,32-.
This study aimed to develop and evaluate a VGG16-LSTM model as a complementary imagebased screening tool for bacterial microscopic images relevant to applied microbiology, environmental monitoring, and preliminary bioremediation-oriented assessment.
The study used selected classes from the DIBaS dataset .
and compared Adam.
RMSProp, and stochastic gradient descent (SGD) optimizers under smallsample learning conditions with on-the-fly The model was not designed to replace culture-based biochemical.
MALDI-TOF, or molecular identification methods.
Instead, it was positioned as a preliminary decision-support approach that may help accelerate bacterial recognition in microbiology and environmental life science laboratories when rapid visual screening is needed.
MATERIAL AND METHOD
Dataset This study used a subset of digital microscopic images of bacteria from the Digital Images of Bacteria Species (DIBaS) dataset .
The DIBaS dataset contains microscopic images of 33 bacterial species.
Six classes were selected for the present study to represent biologically diverse bacteria relevant to environmental monitoring, applied microbiology, biotechnology, and comparative bacterial recognition.
The selected classes included Gram-negative rods.
Grampositive rods.
Gram-positive cocci, and aerotolerant anaerobic Gram-positive bacteria.
The aim was not to claim that all selected bacteria are primary bioremediation agents but to develop an image-based screening framework across biologically diverse bacterial classes.
However, in this study, particular emphasis was placed on bacterial species commonly associated with bioremediation (Table .
To address this, we used a hybrid framework to experimentally investigate a Deep Learning approach.
Hybrid VGG16-LSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening (Yanti, et al.
Table 1.
Composition of the bacterial colony image dataset Genera Species Acinetobacter Escherichia Lactobacillus Micrococcus Propionibacterium Pseudomonas Number Images Sources: DIBaS dataset .
These research phases included dataset generation, image pre-processing with both the one-hot and RGB representations, and dynamic .
n-the-fl.
data augmentation for all Furthermore, we implemented a Hybrid VGG16-LSTM model and compared its performance against 3 various optimizers.
Data Preprocessing The raw image dataset must undergo the following preprocessing before being fed as input to the model: Data Splitting to prevent data leakage among the training, validation, and test subsets, we followed the 70, 20, 10 rule.
Training Set .
%): This was used to directly train the model's weights.
Validation Set .
%): This subset was used as the model benchmark during the training process, and Early Stopping was applied.
Testing Set .
%): This was reserved for the final test of how well a trained model performed on new data.
Image Normalization: The pixel values of the images were normalized from the original range .
, .
to the standardized range .
, .
to reduce the scale disparity for faster convergence of the gradient.
Image Rescaling: All images were rescaled to 224y224 pixels, as per the input layer specifications of the VGG16 architecture.
Microscopic Data Augmentation Strategies Because bacterial microscopic images generally do not have a fixed topAebottom or leftAe right orientation, horizontal and vertical flipping augmentation operations.
We used an On-the-fly Augmentation method.
This processing method consists of the free production of manipulated images without generating additional storage files from a previous execution epoch.
The augmentation parameters are Geometric Rotation .
A), to simulate the non-uniform orientation of small objects.
Vertical and Horizontal Flip, as no AutopAy or AubottomAy is defined in microscopic object and therefore in could be zoom range of up to 20%, to model variations on microscope lens magnifying Brightness .
, due to the uneven intensity of microscope lamp.
Shift & Shear reshaping position the image, relocating the object at different place.
Table 2.
Sequential model Layer .
input_layer (InputLaye.
block1_conv1 (Conv2D) block1_conv2 (Conv2D) block1_pool (MaxPooling2D) block2_conv1 (Conv2D) block2_conv2 (Conv2D) block2_pool (MaxPooling2D) block3_conv1 (Conv2D) block3_conv2 (Conv2D) block3_conv3 (Conv2D) block3_pool (MaxPooling2D) block4_conv1 (Conv2D) block4_conv2 (Conv2D) block4_conv3 (Conv2D) block4_pool (MaxPooling2D) block5_conv1 (Conv2D) block5_conv2 (Conv2D) block5_conv3 (Conv2D) block5_pool (MaxPooling2D) reshape (Reshap.
lstm (LSTM) dropout (Dropou.
dense (Dens.
dropout_1 (Dropou.
dense_1 (Dens.
Output Shape (None, 224, 224, .
(None, 224, 224, .
(None, 224, 224, .
(None, 112, 112, .
(None, 112, 112, .
(None, 112, 112, .
(None, 56, 56, .
(None, 56, 56, .
(None, 56, 56, .
(None, 56, 56, .
(None, 28, 28, .
(None, 28, 28, .
(None, 28, 28, .
(None, 14, 14, .
(None, 14, 14, .
(None, 14, 14, .
(None, 14, 14, .
(None, 14, 14, .
(None, 7, 7, .
(None, 49, .
(None, .
(None, .
(None, .
(None, .
(None, .
Param# 1,792 36,928 73,856 147,584 295,168 590,080 590,080 1,180,160 2,359,808 2,359,808 2,359,808 2,359,808 2,359,808 328,192 8,256 Exp.
Life Sci.
, 2026.
Vol.
No.
2, 80-88.
Hybrid VGG16-LSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening (Yanti, et al.
Hybrid Model Architecture (VGG16-LSTM) The hybrid CNN-LSTM model is introduced by feature-learning advantages of CNNs and the sequential processing advantages of LSTMs.
A pre-trained VGG16 model with ImageNet weights was utilized as the base feature extractor.
To preserve the generalized visual features learned from ImageNet and prevent overfitting, the entire VGG16 convolutional base was frozen .
, its layer weights were set to non-trainabl.
Consequently, only the weights of the newly appended LSTM layers were updated during the training process.
The output from our final convolutional block, which has a shape of .
, 7, .
, is reshaped into a .
, .
feature vector by a reshape layer.
This vector is subsequently fed into the LSTM, treating the spatial side information as a continuous feature sequence.
Sequence Learning (RNN - LSTM) A single LSTM unit with 128 neuron units was used to learn the feature correlations of VGG16.
Dropout .
was set to prevent overfitting.
The classifier (Fully Connecte.
was a 64-dense-layer ReLU.
An output layer using softmax activation for multiclass categorization.
The Hybrid model was experimentally tested under three optimization conditions with these hyperparameter values.
For experimentation, they evaluated the hybrid model under three optimizer scenarios with such hyper-parametrization.
Early Stopping was added to the training of the model by tracking the Validation Loss.
After five passes without a reduction in the validation loss, training was stopped, and the best weights were returned.
RESULT
This section presents a comprehensive experimental evaluation of the proposed CNNAe LSTM hybrid model (VGG16AeLSTM), combined with on-the-fly data augmentation, for classifying microscopic images of bacteria.
Given the limited number of samples per class .
pproximately 20 image.
, dynamic augmentation was employed during training to increase data variability, mitigate overfitting, and improve model The model performance was evaluated using three optimization algorithms.
Adam.
RMSProp, and Stochastic Gradient Descent (SGD), to identify the most effective optimization strategy under this small-sample learning regime.
The training stability and convergence behavior were analyzed using the validation accuracy and loss curves at training epochs 10, 20, 30, 40, and 50.
As shown in Figure 1, each optimizer exhibits distinct learning dynamics when trained in conjunction with on-the-fly data Figure 1.
Validation Accuracy Comparison across Adam.
RMSProp, and SGD Optimizers (A) epoch = 10.
(B) epoch = 20.
(C) epoch = 30.
(D) epoch = 40.
and (E) epoch = 50 Exp.
Life Sci.
, 2026.
Vol.
No.
2, 80-88.
Hybrid VGG16-LSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening (Yanti, et al.
The Adam optimizer demonstrated rapid performance improvement during the early training phase, reaching a validation accuracy above 80% by epoch 30.
This behavior indicates that Adam benefits from the increased data variability introduced by on-the-fly augmentation, enabling fast adaptation to diverse augmented However, after this point, the performance stagnated, and no further improvement was observed after epoch 50.
This plateau suggests that Adam converges prematurely to a local minimum, potentially limiting its ability to exploit the additional augmented variations introduced in the later In contrast.
SGD exhibits pronounced instability throughout the training process, characterized by large oscillations in validation A sharp drop in accuracy observed around epoch 40 .
own to 64.
29%) indicates an unstable gradient update.
Despite the regularization effect of on-the-fly augmentation.
SGDAos fixed learning rate appears insufficient to handle the increased stochasticity of dynamically augmented inputs, leading to overshooting during In contrast.
RMSProp exhibited a gradual yet highly stable learning trajectory.
Although its initial convergence rate is slower than that of Adam.
RMSProp continuously benefits from the diversity of augmented samples and achieves a significant performance gain between epochs 30 Ultimately.
RMSProp attained the highest and most s3 validation accuracy by the end of training, indicating superior compatibility with both the CNNAeLSTM architecture and the on-thefly augmentation strategy.
A quantitative comparison of the classification performance across epochs 10 to 50 is summarized in Table 3.
The results clearly demonstrate that RMSProp is the most effective optimizer when combined with the VGG16AeLSTM model and on-the-fly data augmentation.
RMSProp achieves the highest values across all evaluation metrics, including Accuracy .
86%).
Precision .
24%).
Recall .
86%), and F1-score .
38%).
The consistently high F1-score indicates a balanced trade-off between precision and recall, confirming that the model generalizes well across bacterial classes despite the limited dataset size.
The model demonstrated a strong discriminative capability across most bacterial Classes Acinetobacter baumannii.
Lactobacillus plantarum.
Micrococcus , and Propionibacterium acnes were classified with perfect accuracy, indicating that on-the-fly augmentation successfully enhanced feature robustness for these classes.
Minor misclassifications were observed in the Pseudomonas aeruginosa class, where a small number of samples were incorrectly predicted as Escherichia coli.
This confusion is biologically plausible because both species are gram-negative, rod-shaped bacteria with highly similar microscopic morphologies.
Even with dynamic augmentation, such subtle inter-class similarities remain challenging for CNN-based feature extractors, underscoring the intrinsic difficulty of fine-grained bacterial-image classification.
The superior performance of RMSProp in this study can be attributed to its adaptive gradient normalization mechanism, which is particularly components, such as LSTM layers.
Table 3.
Classification performance metrics across optimizers under On-the-Fly data augmentation Optimizer Adam
RMSProp
SGD
Adam
RMSProp
SGD
Adam
RMSProp
SGD
Adam
RMSProp
SGD
Adam
RMSProp
SGD
Accuracy Precision Recall F1-Score Source: Data processing, 2025 Exp.
Life Sci.
, 2026.
Vol.
No.
2, 80-88.
Hybrid VGG16-LSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening (Yanti, et al.
Figure 2.
Confusion Matrix of the hybrid VGG16-LSTM model trained .
(A) Adam.
(B) RMSProp and.
(C) SGD Optimizers As shown in Figure 2, on-the-fly data augmentation introduces controlled stochasticity into the training process by presenting the model with a virtually unlimited set of biologically plausible variations.
RMSProp effectively stabilizes gradient updates under this dynamic input distribution, thereby enabling a more efficient exploration of the loss landscape.
The premature convergence observed with Adam, despite its theoretical advantages, suggests that bias-corrected momentum may constrain parameter exploration when the training data distribution changes continuously due to on-the-fly augmentation.
Consequently.
Adam tends to converge early to suboptimal solutions in this small-sample scenario.
DISCUSSION
This study repositioned microscopic bacterial image classification as a support task for applied microbiology and environmental life science, rather than as a purely computational problem.
The selected bacteria represent a mixture of organisms, and applied-microbiology classes.
This design is important because an automated imagebased system intended for environmental or bioremediation-oriented laboratories must be interpreted based on the biological properties of the organisms being classified.
The strongest biological connection to bioremediation is provided by Pseudomonas spp.
, which are widely recognized for their metabolic versatility, production, and survival in contaminated environments .
Acinetobacter spp.
Micrococcus spp.
are also environmentally relevant because members of these genera have been detected in soil, wastewater, and polluted matrices and are associated with hydrocarbon degradation, biosurfactant activity, and heavy Exp.
Life Sci.
, 2026.
Vol.
No.
2, 80-88.
metal tolerance .
In contrast.
coli is more appropriately interpreted as an environmental indicator and comparative Gram-negative rod than as a primary bioremediation organism .
Lactobacillus plantarum has applied-life-science relevance in fermentation biotechnology, probiotic applications, and biosorption-related metal binding .
Propionibacterium acnes, currently C.
acnes, should be interpreted as a Gram-positive anaerobe that increases physiological diversity rather than as a bioremediation agent .
From a modeling perspective.
RMSProp exhibited the best performance and stability.
RMSProp divides the gradient by a running average of recent gradient magnitudes, which can help stabilize the training when inputs vary dynamically owing to augmentation .
This behavior is particularly relevant when an LSTM layer is included, as its the recurrent components may be sensitive to unstable gradient updates.
The final RMSProp performance of 92.
accuracy and 92.
38% F1-score suggests that the VGG16-LSTM architecture can extract useful limited-data Adam showed faster early improvement but reached a performance plateau after epoch 30.
This finding is consistent with reports that adaptive methods may converge quickly but do not always provide the best generalization in some non-convex learning tasks .
SGD showed the least stable performance, likely because its fixed-update behavior was less able to accommodate the stochasticity introduced by the on-the-fly augmentation.
These differences indicate that the training strategy is important when microscopic datasets are small and dynamically augmented.
Hybrid VGG16-LSTM Classification of Microscopic Bacterial Images for Environmental Microbiology Screening (Yanti, et al.
The misclassification between Pseudomonas aeruginosa and Escherichia coli was one of the most biologically informative findings of this Both species are Gram-negative rods and may share similar microscopic morphologies, especially when image quality, staining variation, and field-of-view differences are present .
This observation supports the argument that microscopic image-based classification should be used as a screening tool and not as a definitive taxonomic identification method.
Confirmatory testing remains necessary, particularly when bacteria have similar morphologies or when species-level environmental, or regulatory consequences.
Compared with standalone CNN-based microorganism image studies, the VGG16-LSTM design may offer an advantage by allowing spatial dependencies in the convolutional feature map to be learned as a sequence .
,33,.
On-the-fly augmentation also reduces the risk of memorizing a small fixed set of transformed images, a problem noted in image augmentation studies involving limited datasets .
Nevertheless, the performance must be interpreted with caution because the original dataset was small and did not capture the full variability of environmental Overall, the results support the feasibility of using hybrid image-based learning as a complementary tool for preliminary bacterial screening in applied microbiology.
In a bioremediation-oriented workflow, such a system could help prioritize candidate images or guide early recognition before biochemical and molecular confirmation.
However, its practical value will depend on validation across larger datasets, multiple laboratories, environmental samples, and different staining protocols, as well as bacteria isolated from contaminated sites.
CONCLUSION
This study evaluated a hybrid VGG16-LSTM model with on-the-fly data augmentation as a complementary image-based approach for preliminary bacterial screening in applied microbiology, environmental monitoring, and bioremediation-oriented The organisms, biotechnology-related species, and comparative microscopic classes, thereby supporting the study's applied life-science .
Among the three optimizers tested.
RMSProp achieved the best performance after 50 epochs, 86% accuracy, 95.
24% precision, 92.
recall, and a 92.
38% F1-score.
These results indicate that the proposed model can extract useful microscopic features from limited bacterial image data.
Misclassification between P.
aeruginosa and E.
coli reflects the biological difficulty of distinguishing visually similar Gram-negative rodshaped bacteria using microscopic images alone.
Therefore, the model should be interpreted as a preliminary screening and decision-support tool, not as a replacement for culture-based.
MALDI-TOF, identification methods.
Overall, the proposed framework shows potential to support early bacterial recognition in applied microbiology and environmental life sciences laboratories.
Further validation using larger datasets, environmental isolates, different staining protocols, and multi-laboratory imaging conditions is required before practical bioremediation-related Author Contribution: IY: Conceptualization.
WritingAeOriginal Draft.
Methodology.
Software.
Project Administration.
M: Validation.
WritingAe Review and Supervision.
A: Methodology.
AK:
Methodology.
Conceptualization.
Validation.
WritingAeReview.
Editing, and Supervision.
SA:
Conceptualization.
Validation.
Resources.
WritingAeReview.
Editing.
Supervision.
Software, and Formal Analysis.
HAB: Software, and Formal Analysis.
Funding: This research was funded by the Faculty of Sciences.
Technology, and Mathematics, grant 05/UN10.
F09/PN/2025.
Conflict of Interest: The authors have no conflict of interest.
REFERENCES