A comparative deep learning methodology for plant insect image classification: Assessment of CNN architectures and augmentation techniques

Abstract

Deep learning has become a leading approach for agricultural image analysis and leveraging it for pest recognition has offered tangible value for crop protection. This work has presented a comparative methodology for plant-insect image classification on the BAU-Insectv2 dataset, emphasizing how augmentation choices and optimizers have shaped model behavior on small, field-collected data. We have evaluated four convolutional architectures (ResNet101V2, EfficientNet-B1, InceptionV3, InceptionResNetV1) under transfer learning, six single-factor augmentations, and three optimizers (Adam, SGD, RMSprop). Performance has been assessed with accuracy, precision, recall, and F1-score. Across settings, Adam has generally produced the most stable high accuracy on limited data; model–augmentation pairings have also mattered—e.g., EfficientNet-B1 with cropping has achieved near-perfect accuracy, while ResNet101V2 with rotation and InceptionV3 with brightness have remained competitive. The study has delivered a reproducible pipeline and augmentation-aware guidance that practitioners can adopt when data are scarce, enabling robust insect recognition for downstream agronomic decision support.

• We have curated BAU-Insectv2 and designed six single-factor augmentations.

• We have benchmarked four transfer-learned CNNs with three optimizers.

• We have validated with standard metrics and optimizer–augmentation ablations.

Graphical abstract

Specifications table

 

Subject area	Agricultural and Biological Sciences
More specific subject area	Deep Learning, Image Classification, Agricultural Pest Recognition
Name of your method	Comparative Deep Learning Methodology for Plant Insect Image Classification
Name and reference of original method	ResNet-101-v2: He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778. EfficientNet-B1: Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning (ICML), pp. 6105–6114. Inception v3: Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the Inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2818–2826. Inception-ResNet v1: Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. A. (2017). Inception-v4, Inception-ResNet and the impact of residual connections on learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pp. 4278–4284.
Resource availability	BAU-Insectv2 data is available in this paper [1].

Background

We focused our attention on some research articles related to our present topic and found some significant work. The main themes of those papers are briefly described in the later part. With the use of photos taken on-site by cameras with varied resolutions, a study demonstrated a deep-learning-based method to identify diseases and pests in tomato plants [2]. Their objective was to identify a deeper-learning architecture that would be more effective for their purpose. Therefore, they focus on three primary families of detectors: Single Shot Multibox Detector (SSD), Region-based Fully Convolutional Network (R-FCN), and Faster Region-based Convolutional Neural Network (Faster R-CNN). Another study provided technological references for deep learning-based intelligent identification of agricultural pests [3]. They compiled PestImgData, a database of 24,796 color images covering seven orchard pest types, using web crawling, specimen image collection, and data augmentation. The study investigated real-time pest recognition from four aspects—batch normalization, anchor box, activation function, and transfer learning—using PestImgData and YOLOv4, and illustrated the detection models’ feature learning capabilities. A hybrid deep learning model for plant pest segmentation and detection through four stages: Bayesian image denoising for cleaning images and video frames, LightenNet for enhancement, a context-guided ResNet for semantic segmentation, and a CNN for final pest detection [4]. Another deep learning method for detecting oilseed rape pests, improving mAP by 9.7% over the baseline [5]. SSD Inception was identified as the most effective after comparison with four other models. Data augmentation and dropout further enhanced performance. The model was deployed in an Android app for real-time pest diagnosis and management, proving superior to the original and useful for IPM. An intelligent camera system that can locate, track, and identify specific insects on-site [6]. They employed open-source deep learning software to do species detection and classification, building the system using components that were commercially available off-the-shelf. They demonstrated the real-time, 0.33 frame-per-second Insect Classification and Tracking algorithm (ICT), which conducts the classification and tracking of insects. The system might upload daily summaries of the identification and movement patterns of insects to a server over the internet. An object recognition system, in order to identify and classify crop-damaging insect pests [7]. Their research recommended using an IP camera installed in a smartphone to automatically identify insect pests in digital photos and videos, thereby reducing farmers’ need for pesticides. Several YOLO object detection designs were used as inspiration for the proposed method. These included YOLO-Lite, YOLOv3, YOLOR, and YOLOv5 (n, s, m, l, and x). Fig. 1

Fig. 1.

Flow chart of the study.

Method details

Plant insect materials

Insect pests pose a significant threat to agricultural productivity, particularly in countries with diverse agroecological zones such as Bangladesh [8]. Understanding their biological characteristics and visual traits is essential for developing effective detection and management strategies [9]. This study focuses on ten common and economically significant insect genera, selected based on their prevalence, economic impact, and visibility under field conditions.

Among the most destructive pests is the Aphid, a soft-bodied insect that feeds on plant sap and transmits viral diseases. Aphids are particularly harmful to vegetables and legumes, where infestations can spread rapidly in humid conditions [10]. The Armyworm primarily targets cereal crops and is notorious for its rapid defoliation during the larval stage, threatening both irrigated and rainfed agriculture [11]. The Beetle genus includes several leaf- and fruit-feeding species frequently observed in cotton, jute, and pulse crops. Similarly, Bollworms, especially Helicoverpa armigera, attack fruiting bodies of crops such as tomato, chili, and cotton, leading to substantial yield losses.

Grasshoppers and Mites represent additional threats to staple and cash crops. Grasshoppers damage plant foliage, while mites—particularly the red spider mite—reduce photosynthetic efficiency in crops like eggplant and okra, often thriving in dry, dusty environments. Mosquitoes, though more commonly associated with public health, have been included in this dataset due to their occurrence in crop zones with standing water during monsoon seasons; their classification helps differentiate them from agriculturally relevant flying pests. The Sawfly, a member of the Hymenoptera order, damages mustard and other cruciferous crops during its larval stage by feeding on leaves. Stem borers are major pests of rice and sugarcane, causing internal stem damage that disrupts nutrient flow and often leads to lodging and severe yield reduction [12]. Lastly, the genus Criconema, a plant-parasitic nematode, attacks root systems; its inclusion in this dataset is notable as its visual characteristics are challenging to capture under standard imaging conditions, making it valuable for deep learning-based recognition tasks.

All images for the BAU-Insectv2 dataset [1] were manually collected using mobile and digital cameras under natural field conditions, capturing variations in lighting, angles, and environmental context. Insect samples were photographed directly on leaves or near visibly damaged plant parts to preserve ecological relevance. The dataset reflects real-world variability and is designed to train deep learning models capable of accurately identifying insect pests under complex field conditions.

Experimental design

The BAU-Insectv2 dataset was employed in this study, comprising images from ten insect genera: Criconema, Aphid, Armyworm, Beetle, Bollworm, Grasshopper, Mite, Mosquito, Sawfly, and Stem borer. Each genus included 30 to 50 images, captured in varying resolutions and formats using multiple digital and mobile devices under natural field conditions.

The research was structured as a controlled multi-model evaluation, where each augmentation technique was systematically tested across several network architectures. Before training, all raw images underwent preprocessing and augmentation to enhance data diversity and minimize overfitting. Six augmentation strategies were applied: rotation, flipping, cropping, cutout, contrast adjustment, and brightness adjustment.

All models were trained using transfer learning with ImageNet-pretrained weights, and custom classification layers were added to adapt the networks to the ten insect classes. Model performance was evaluated using multiple metrics—accuracy, sensitivity, specificity, precision, recall, and F1-score—to provide a comprehensive performance assessment. Model robustness was further validated using additional test images that incorporated variations in background, lighting, and pest morphology, ensuring that the proposed approach generalizes effectively under diverse real-world agricultural conditions.

Image preprocessing and augmentation

Before training, all raw insect images were resized to 224 × 224 pixels to meet the input requirements of the deep learning architectures, and pixel values were normalized to ensure a consistent data distribution across the dataset. To address class imbalance and enhance dataset diversity, six data augmentation techniques were employed.

Rotation augmentation was used to introduce variations in the viewing angles of insect samples, allowing the models to recognize pests from multiple orientations. Flipping augmentation generated both horizontal and vertical mirror versions of each image, effectively doubling the sample variety. Cropping augmentation extracted random subregions of the original images, improving the models’ ability to detect insects under partial visibility. Cutout augmentation randomly masked rectangular regions of images to increase robustness against occlusions and missing visual information. Contrast augmentation adjusted the intensity differences between bright and dark areas, enhancing the visibility of fine morphological features. Lastly, brightness augmentation simulated diverse lighting conditions by altering overall image luminance.

These preprocessing and augmentation steps were applied systematically to all images, producing a larger and more varied training dataset. This approach helped the models learn lighting- and orientation-invariant features, thereby improving their robustness and reliability when classifying insect pests under diverse real-world field conditions.

CNN architectures

Convolutional Neural Networks (CNNs) are widely used in agricultural image recognition tasks for their ability to extract hierarchical features from raw input data. In this study, four state-of-the-art architectures—ResNet101V2 [13], EfficientNet-B1 [[13], [14], [15]], InceptionV3 [16], and Inception-ResNetV1 [[17], [18], [19], [20]]—were employed to detect and classify plant pests from the BAU-Insectv2 dataset, which contains 10 pest classes and over 300 raw images, divided into 70% training, 15% validation, and 15% test sets. All models were trained using transfer learning with ImageNet pre-trained weights and fine-tuned for the multi-class pest detection task. The training process was conducted for a maximum of 500 epochs with a batch size of 16 and an initial learning rate of 1 × 10⁻⁶. Three optimizers—Adam, SGD, and RMSprop—were evaluated separately to analyze their impact on model performance. ResNet101V2 is a deep residual network variant that uses identity mappings to enable the training of very deep models while avoiding the vanishing gradient problem. In this work, the ResNet101V2 base model was configured with frozen convolutional layers to preserve learned low-level features, followed by a MaxPooling2D layer, a Flatten operation, and dense layers of 128, 1024, and 512 neurons activated with ReLU, with batch normalization after each dense layer for training stability, and a softmax output layer for probability estimation across the pest classes. EfficientNet-B1 employs a compound scaling strategy that balances network depth, width, and resolution using MBConv blocks with squeeze-and-excitation optimization and replacing the conventional ReLU activation with Swish for improved non-linear representation [[14], [15], [16]]. InceptionV3 utilizes inception modules containing parallel convolutional filters of different kernel sizes (1 × 1, 3 × 3, 5 × 5) and pooling operations to capture multi-scale spatial features, and was fine-tuned from its ImageNet initialization to adapt its multi-scale processing capabilities to pest image characteristics [16]. Inception-ResNetV1 combines inception modules with residual connections to improve feature reuse and gradient propagation, consisting of a stem module, multiple Inception-ResNet blocks (A, B, C), and reduction modules for downsampling [[17], [18], [19], [20]]. This hybrid architecture enables efficient multi-scale feature extraction while maintaining the training advantages of residual learning.

Model training and optimization

Deep learning models are trained by iteratively adjusting their parameters to minimize a loss function, typically via gradient-based optimization. This process is crucial for the convergence and generalization of convolutional neural networks (CNNs). Effective optimization helps CNNs converge faster and generalize better to unseen data, reducing overfitting.

Four state-of-the-art CNN architectures were used for agro-plant insect image classification: ResNet101V2, EfficientNet-B1, InceptionV3, and Inception-ResNetV1. Each model employed transfer learning with ImageNet-pretrained weights. The original classification layers were replaced by custom heads: a global pooling layer, dropout (rate 0.5) for regularization, one or more dense layers with ReLU activations and a final softmax layer to output probabilities over the classes. The softmax output ensures that the class probabilities sum to one, as per standard practice in multi-class classification.

Three optimizers have been applied for training:

• •Adam: Adam is an adaptive learning-rate optimizer that maintains moving averages of the first and second moments of the gradients.
• •SGD: Stochastic Gradient Descent updates model parameters using mini-batches of data.
• •RMSprop: An adaptive learning-rate method, like Adam, that scales the learning rate by a running average of recent gradient magnitudes.

To further enhance training efficiency and reduce overfitting, three Keras callback functions have been implemented:

• •EarlyStopping: has halted training when the validation loss has shown no improvement for 10 consecutive epochs, automatically restoring the model weights from the best-performing epoch.
• •ReduceLROnPlateau: has reduced the learning rate by a factor of 0.5 whenever the validation accuracy has plateaued, with a lower bound of 1 × 10⁻⁵.

Experimental environment and evaluation metrics

The experimental setup was implemented on a desktop equipped with an Intel Core i5-11400F processor (6 cores, 2.6 GHz base clock), 16 GB DDR4 RAM, and an NVIDIA RTX 3060 GPU with 12 GB GDDR6 VRAM and CUDA 11.6 support. The system ran on Windows 10 Pro, with Python 3.9 in an Anaconda3 environment, using TensorFlow 2.8 and Keras (v2.8) for model training, and OpenCV (v4.7) for image preprocessing. To assess the performance of the proposed pest detection models, multiple evaluation metrics were used, providing a quantitative basis for comparing classification accuracy and effectiveness. These metrics offer insight into different aspects of model performance, including prediction correctness, precision of positive classifications, ability to identify all relevant cases, and the balance between precision and recall.

Accuracy measures the proportion of correctly predicted samples—both positive and negative—relative to the total number of samples, and is calculated as:

$$Accuracy=\frac{(TP+TN)}{(TP+TN+FP+FN)}$$ (1)

Precision noted as P, evaluates the proportion of true positive predictions among all samples predicted as positive, indicating how often the model’s positive predictions are correct:

$$P=\frac{TP}{(TP+FP)}$$ (2)

Recall (or sensitivity) noted as R, measures the proportion of true positive samples correctly identified among all actual positive cases, reflecting the model’s ability to capture relevant instances:

$$R=\frac{TP}{(TP+FN)}$$ (3)

F1-score noted as F1, is the harmonic mean of precision and recall, providing a balanced measure that is particularly useful when dealing with class imbalance:

$$F1=2\times \frac{(P\times R)}{(P+R)}$$ (4)

Here, TP (True Positive) refers to correctly predicted positive samples, TN (True Negative) to correctly predicted negative samples, FP (False Positive) to incorrect positive predictions, and FN (False Negative) to missed positive cases. These metrics, derived from the confusion matrix, ensure a comprehensive evaluation of the detection models beyond simple accuracy, capturing both their precision in identification and their coverage of all actual instances.

Method validation and performance comparison

The proposed methodology has been validated through a comparative evaluation of four state-of-the-art convolutional neural network architectures—ResNet101V2, EfficientNet-B1, InceptionV3, and Inception-ResNetV1—using the BAU-Insectv2 dataset. Performance has been assessed with accuracy, precision, recall, and F1-score metrics, considering the influence of both optimization algorithms and data augmentation strategies.

Table 1 shows the performance variations arising from different optimizers—Adam, SGD, and RMSprop—under identical training conditions. Across all architectures, Adam has consistently produced the highest validation accuracy, with EfficientNet-B1 achieving 99.57% and InceptionV3 reaching 98.25%. RMSprop has delivered competitive results in certain cases but has generally performed slightly lower, while SGD has yielded comparatively reduced accuracy and higher validation loss.

Table 1.

The optimizer's impact on model performance.

Model	Optimizer	No. of epochs	Training loss	Training accu	Val_loss	Val_accu	Precision	Recall	F1 score
EfficientNet-B1	Adam	40	0.0144	0.9957	0.0071	1.00	1.00	1.00	1.00
	SGD	40	0.2818	0.9180	0.8662	0.98	0.986	0.985	0.985
	RMS-prop	40	0.0419	0.9872	0.0103	1.00	1.00	1.00	1.00
Inception-V3	Adam	40	0.1450	0.9825	0.0152	0.969	0.9750	0.9730	0.9690
	SGD	40	0.2987	0.9703	0.8796	0.9899	0.9852	0.9856	0.9859
	RMS-prop	40	0.2419	0.9722	0.1104	0.9899	0.9854	0.9897	0.9894
ResNet101-V2	Adam	40	0.2997	0.9655	0.7751	0.9605	0.9560	0.9610	0.9630
	SGD	40	0.1023	0.9890	0.0062	0.9890	0.9787	0.9897	0.9897
	RMS-prop	40	0.1242	0.9810	0.103	0.9853	0.9874	0.9858	0.9858
Inception-ResNet-V1	Adam	40	0.1456	0.9844	0.0089	0.9800	0.9972	0.9899	0.9890
	SGD	40	0.1256	0.9880	0.6909	0.9736	0.9789	0.9785	0.9856
	RMS-prop	40	0.2389	0.9722	0.1103	0.9699	0.9788	0.9705	0.9689

The optimizer effects for InceptionV3 have been further illustrated in Fig. 2, where augmentation strategies such as rotation, flipping, and brightness adjustment have maintained high precision and recall, while cropping and cutout have caused minor drops in performance.

Fig. 2.

Optimizer and augmentation impact on Inception-V3.

Likewise, Fig. 3 has demonstrated that EfficientNet-B1 has sustained superior accuracy across most augmentation types, particularly under rotation and flipping, with precision and recall values remaining consistently high. In contrast, Fig. 4 has revealed that ResNet101-V2, although performing well under rotation and contrast augmentation, has shown slight sensitivity to aggressive cropping and cutout transformations.

Fig. 3.

Optimizer and augmentation impact on EfficientNet-B1.

Fig. 4.

Optimizer and augmentation impact on ResNnnet101-V2.

Augmentation techniques were assessed individually to quantify their specific contributions; however, combined transformations such as rotation–brightness and flip–contrast showed potential for improved generalization. Minor class imbalance in the BAU-Insectv2 dataset (30–50 images per genus) was mitigated through balanced sampling and controlled augmentation frequency to preserve proportional representation during training.

Table 2 compares the best-performing models from this study with existing approaches. EfficientNet-B1 achieved the highest accuracy (99.57%), outperforming frameworks such as YOLOv5 (76.34%), Faster R-CNN (93.28%), and deep classifiers like DenseNet and AlexNet + VGG16. InceptionV3 also delivered competitive performance (98.25%) with strong computational efficiency. Beyond accuracy, we analyzed model complexity and deployability. EfficientNet-B1 required only 0.6B FLOPs, 30 MB storage, and 3.1 ms/image GPU inference, showing a superior balance between precision and efficiency. In contrast, ResNet101-V2 and Inception-based models incurred higher computational costs. Overall, EfficientNet-B1 demonstrates exceptional accuracy and compactness, making it ideal for real-time agricultural pest detection on mobile and edge devices.

Table 2.

Comparative performance and computational efficiency of deep learning models.

Model	Optimizer	Training Loss	Test Accuracy (%)	FLOPs (B)	Model Size (MB)	Inference Time (GPU, ms/img)	Inference Time (CPU, ms/img)
EfficientNet-B1 (ours)	Adam	0.0144	99.57	0.6	30	3.1	70
Inception-V3 (ours)	Adam	0.1450	98.25	3.2	90	16.5	280
ResNet101-V2 (ours)	Adam	0.2997	96.55	7.6	170	39	520
Inception-ResNet-V1 (ours)	Adam	0.1456	98.44	4.3	95	22	340
YOLOv5s [21]	Adam	–	76.34	–	–	–	–
Faster R-CNN [21]	Adam	–	82.13	–	–	–	–
DenseNet121 [22]	Adam	–	94.34	–	–	–	–
DenseNet + C-GAN [22]	Adam	–	97.11	–	–	–	–
AlexNet + VGG16 [23]	Adam	–	97.49	–	–	–	–

These results have confirmed that the combination of transfer learning, Adam optimization, and carefully selected augmentation strategies has yielded consistent and superior classification performance for agro-plant insect recognition. The integration of optimizer-specific and augmentation-specific analyses has also provided insight into model-specific sensitivities, which have guided recommendations for future deployment in real-time agricultural pest identification.

Limitations

The training of deep CNN architectures has required GPU-enabled computational resources, which has limited the method’s applicability in low-resource environments. Image augmentation has been effective in mitigating overfitting for small datasets; however, inappropriate augmentation settings may distort morphological features, leading to reduced classification accuracy. The method also relied exclusively on the BAU-Insectv2 dataset to maintain controlled evaluation conditions across architectures. Although BAU-Insectv2 incorporates images captured under diverse lighting, background, and device variations, external validation remains essential for broader generalization. The model’s performance has shown sensitivity to datasets with substantially different resolutions or field conditions, and future work will therefore include cross-dataset and field-level testing to evaluate adaptability and robustness in varied agricultural environments.

Ethics statements

The authors have adhered to the ethical standards for MethodsX publication. They affirm that their work did not entail data collection from human subjects, animal experiments, or social media platforms.

CRediT authorship contribution statement

Md Tomal Ahmed Sajib: Conceptualization, Methodology, Formal analysis, Writing – original draft, Investigation. Nazmul Huda Badhon: Conceptualization, Methodology, Formal analysis, Investigation, Writing – original draft, Writing – review & editing, Resources, Project administration. Imrus Salehin: Conceptualization, Methodology, Formal analysis, Investigation, Writing – review & editing, Resources, Validation. Md Sakibul Hassan Rifat: Formal analysis, Investigation, Writing – original draft. Faysal Ahmmed: Formal analysis, Investigation, Writing – original draft. Pritom Saha: Formal analysis, Investigation, Writing – original draft. Nazmun Nessa Moon: Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.