An Integrated Framework for Rice Disease Detection and Smart Irrigation Using EfficientNet-B0 and IoT

2.1 Classification of rice diseases

Classifying plant diseases, particularly in rice, is a crucial aspect of early pest and disease management. Traditional methods, such as visual observation by farmers or agricultural experts, are often subjective, time-consuming, and require specialized skills. Consequently, deep learning-based technologies, especially CNN, have been increasingly adopted to automate disease identification with higher speed and accuracy.

Several studies have highlighted the effectiveness of CNN for crop image classification. For instance, Akter et al. [11] developed a CNN model equipped with attention and residual connections to classify rice leaf diseases, achieving a test accuracy of 99.6% across four major diseases: Bacterial Blight, Brown Spot, Blast, and Tungro. Similarly, Yang et al. [12] compared multiple architectures, including DenseNet121, InceptionV3, and MobileNetV2, and demonstrated that an ensemble approach achieved 98% accuracy, emphasizing the benefits of transfer learning. Another study, Singh et al. [13] introduced a custom CNN that reached 91.4% accuracy, outperforming InceptionV3 and EfficientNet-B2 by leveraging a faster and optimized architecture. Moreover, Tran et al. [14] demonstrated that combining multispectral and RGB imagery improved the F1-score compared to RGB-only datasets, opening new directions for more robust disease detection.

In the specific context of rice cultivation, CNN have been applied to detect three dominant diseases: BLB, Brown Spot, and Leaf Smut [15]. CNN are particularly effective at extracting discriminative features from leaf texture and symptom patterns, enabling reliable classification even with standard camera input. Building on these findings, the integration of CNN-based disease detection with IoT-driven environmental management represents a promising avenue toward AI-enabled precision agriculture systems that directly support food self-sufficiency goals [16].

2.2 Application of IoT in agriculture

IoT technology has been widely implemented in precision agriculture to enhance resource efficiency, particularly water management, while sustaining crop productivity. For instance, Dong et al. [17] developed a cross-field IoT irrigation system equipped with calibrated soil moisture sensors to monitor real-time field conditions. The system automatically activated irrigation when soil moisture levels fell below a predefined threshold, reducing water use by up to 30% in a single rice and fruit growing season without compromising yield. Similarly, Mallareddy et al. [18] demonstrated that IoT-based smart irrigation systems could reduce water consumption by 30-50% while increasing water use efficiency by up to 60% through adaptive scheduling based on real-time humidity and weather data. In another study, Ahad et al. [19] implemented a solar-powered IoT irrigation system for rice fields, which not only minimized labor requirements but also ensured precise and sustainable irrigation practices.

Taken together, these studies confirm the significant potential of IoT in agriculture. Beyond reducing water consumption by up to half, IoT applications have been shown to increase crop productivity by 10-15% and to enable automation and remote monitoring of agricultural processes. The integration of renewable energy sources, such as solar panels, further strengthens the sustainability and scalability of IoT-based farming systems. These advantages highlight IoT as a critical enabler of smart agriculture, particularly in regions where water efficiency and labor reduction are essential for achieving sustainable rice production.

4.1 Experimental configuration

The experimental setup was designed to evaluate the performance of an integrated framework combining CNN-based EfficientNet-B0 for rice leaf disease classification and an IoT-based smart irrigation system. The experiment assessed both subsystems individually and jointly to validate their applicability in precision agriculture. For the rice leaf disease classification, the dataset was prepared as described in Section 3.1, comprising 7,200 RGB images (2,400 original and 4,800 augmented) evenly distributed among three disease classes: Bacterial Leaf Blight (BLB), Leaf Smut, and Brown Spot. All images were resized to 224 × 224 pixels, normalized to the [0, 1] range, and encoded using a one-hot scheme for model input.

The EfficientNet-B0 model was implemented using TensorFlow and Keras, initialized with ImageNet pre-trained weights, and fine-tuned on the rice disease dataset. The model was compiled using the Adam optimizer with a learning rate of 0.001 and categorical cross-entropy as the loss function. Training was performed for 30 epochs with a batch size of 32, using GPU acceleration on Google Colaboratory (NVIDIA T4 GPU). To enhance generalization, dropout with a rate of 0.4 and early stopping were employed.

To ensure the robustness of EfficientNet-B0 performance, two baseline models, VGG16 and MobileNetV2, were trained under identical conditions for comparative evaluation. This benchmarking enabled quantitative assessment of classification accuracy, precision, recall, F1-score, and computational efficiency. Model evaluation used a confusion matrix and per-class metrics to visualize error distribution and interpret prediction reliability. Statistical analysis of results was also conducted to validate consistency across cross-validation folds.

Parallel to the deep learning model, the IoT-based smart irrigation subsystem was developed using a NodeMCU V3 (ESP-12E) microcontroller as the central control unit. The system integrated YL-69 soil moisture, DHT11 air temperature and humidity, DS1820 soil temperature, and NPK nutrient sensors, transmitting data in real time via Wi-Fi to the Blynk cloud platform. The irrigation mechanism consisted of a relay module controlling a mini DC pump, automatically triggered when soil moisture fell below a predefined threshold and stopped once optimal conditions were restored.

The IoT prototype was tested in controlled conditions simulating irrigation cycles and environmental variability. Key evaluation metrics included soil moisture stabilization time, water consumption efficiency, and system responsiveness. The average response time for soil moisture recovery was 40-50 seconds, confirming reliable and energy-efficient operation. The Blynk mobile application also enabled remote monitoring and manual control, improving system usability for farmers.

4.2 Results of rice leaf disease classification

The training and evaluation results demonstrated that EfficientNet-B0 outperformed the baseline models (VGG16 and MobileNetV2) in terms of classification accuracy, convergence speed, and stability. As illustrated in Figure 2, the EfficientNet-B0 model achieved smooth and consistent convergence on the augmented rice leaf dataset consisting of three disease categories: Bacterial Leaf Blight (BLB), Leaf Smut, and Brown Spot. The training accuracy steadily increased and surpassed 98%, while validation accuracy stabilized near 97% after the 20th epoch, indicating strong generalization and minimal overfitting (Figure 2(a)). In contrast, VGG16 exhibited slower convergence and a wider gap between training and validation accuracy, suggesting a higher risk of overfitting. MobileNetV2 performed more efficiently than VGG16 but still yielded slightly lower accuracy than EfficientNet-B0. The loss curves for EfficientNet-B0 (Figure 2(b)) showed a consistent downward trend for both training and validation losses, with early stabilization after initial epochs, confirming effective learning of discriminative patterns. These findings demonstrate that the compound scaling mechanism of EfficientNet-B0 enables superior feature extraction and computational efficiency compared to traditional CNN architectures, making it the most optimal model for rice leaf disease classification under the given experimental conditions.

To validate the effectiveness of EfficientNet-B0, two baseline CNN models, VGG16 and MobileNetV2, were trained and evaluated under the same experimental conditions (Table 3). This comparison assessed the accuracy and efficiency of each model in classifying rice leaf diseases using the same dataset and hyperparameters. The per-class performance of EfficientNet-B0 is presented in Table 4.

2a.jpg

(a)

2b.jpg

(b)

Figure 2. Training and validation performance of the best-performing model (EfficientNet-B0) on the rice leaf disease dataset: (a) accuracy curves and (b) loss curves

Table 3. Comparative performance of CNN architectures on rice leaf disease dataset

Table 4. Performance metrics of EfficientNet-B0 on rice leaf disease classification

3.jpg

Figure 3. Confusion matrix of EfficientNet-B0 on the rice leaf disease test set

The confusion matrix presented in Figure 3 provides a detailed visualization of the classification performance of the EfficientNet-B0 model on the test set, which comprised 20% of the total dataset (1,440 images out of 7,200). The results demonstrate that both Bacterial Leaf Blight (BLB) and Brown Spot were classified with high accuracy, although a few samples were mutually misclassified between these two categories. In contrast, Leaf Smut achieved near-perfect classification, indicating that its distinct morphological features were effectively recognized by the model. These outcomes are consistent with the quantitative evaluation summarized in Table 5 and can be attributed to the visual similarity between BLB and Brown Spot lesions, both of which exhibit irregular brownish patterns that may overlap under variable illumination conditions. Overall, the minimal number of misclassifications highlights the robustness and discriminative capability of EfficientNet-B0 in distinguishing subtle visual differences among rice leaf diseases.

When compared to VGG16 and MobileNetV2, EfficientNet-B0 achieved the highest classification accuracy with fewer parameters than VGG16 and slightly higher computational cost than MobileNetV2. Its compound scaling strategy effectively balanced network depth, width, and resolution, allowing efficient feature extraction while maintaining high generalization performance. These results demonstrate that EfficientNet-B0 offers an optimal trade-off between accuracy and efficiency, confirming its suitability for deployment in IoT-enabled agricultural systems and resource-constrained environments.

In summary, the combination of accuracy and efficiency observed across all evaluation metrics highlights EfficientNet-B0 as a powerful yet lightweight architecture for early rice disease detection. The proposed model provides reliable diagnostic insights to support timely interventions and improved crop management strategies, ultimately contributing to enhanced rice productivity and food self-sufficiency.

4.3 Error analysis

A comprehensive error analysis was conducted to gain deeper insight into the classification behavior and potential weaknesses of the EfficientNet-B0 model. Although the model achieved a high overall accuracy of 97.82%, several misclassifications were observed, primarily between bacterial leaf blight (BLB) and Brown Spot. The confusion matrix (Figure 3) confirmed that these two classes accounted for more than 85% of all misclassified samples.

To explore the underlying causes of these errors, both feature-level and image-level analyses were performed. Visual inspection revealed that early-stage BLB lesions often exhibited circular brown patches with darker margins, closely resembling the mature lesions of Brown Spot, particularly under low illumination. Conversely, late-stage Brown Spot symptoms, characterized by elongated necrotic streaks, were occasionally mistaken for BLB, especially when overlapping veins created linear patterns. These observations suggest that disease progression stages and lighting variability substantially influence the model’s discriminative ability.

Table 5. Examples of misclassified samples from the test set

To better understand the model’s decision rationale, a Gradient-weighted Class Activation Mapping (Grad-CAM) analysis was performed. The resulting saliency maps revealed that EfficientNet-B0 primarily focused on the central lesion regions and ignored peripheral texture information in some cases. Figure 4 illustrates representative heatmaps of correctly and incorrectly classified samples. For correctly classified Leaf Smut, the model concentrated on dense black pustules, while for misclassified BLB-Brown Spot pairs, the attention region covered broader areas, including background noise. This indicates that background interference and non-disease regions sometimes contributed spurious activations, leading to misclassification.

4.jpg

Figure 4. Grad-CAM visualization for feature interpretation and error analysis of EfficientNet-B0 model

To further investigate the visual reasoning behind the model’s predictions, Figure 4 presents a Grad-CAM visualization that illustrates the feature activation patterns of the EfficientNet-B0 model during rice leaf disease classification. The figure provides insight into which regions of the leaf images most influenced the model’s decision-making process, thereby offering a deeper understanding of both correct and incorrect classifications.

The Grad-CAM analysis shows that EfficientNet-B0 effectively captures local lesion textures but occasionally lacks global context awareness [22], particularly when disease regions are small or only partially visible. This limitation can be mitigated by integrating attention-based architectures to refine focus on relevant areas, improving data augmentation to simulate variations in lighting and background, and applying background segmentation during preprocessing to isolate diseased regions. Overall, the error analysis reveals that most misclassifications arise from inter-class visual similarities, inconsistent illumination, and distracting backgrounds. Incorporating feature-visualization methods such as Grad-CAM enhances understanding of the model’s decision-making process, improving interpretability and informing future efforts in dataset development and model optimization.

4.4 Results of IoT-based smart irrigation system

The IoT-based smart irrigation prototype was successfully implemented and extensively evaluated to assess its short-term and long-term performance under both controlled and semi-field conditions. The system employed the NodeMCU V3 microcontroller as the core processing unit, integrated with multiple sensors including YL-69 soil moisture, DS1820 soil temperature, DHT11 air temperature and humidity, and an NPK nutrient sensor. Data from all sensors were transmitted in real time via Wi-Fi to the Blynk cloud server, enabling continuous environmental monitoring and remote system management.

During short-term testing, the system was configured with a soil moisture threshold of 30%. When the moisture level dropped below this threshold, the NodeMCU automatically activated the relay module to switch on the DC pump. Irrigation continued until moisture values returned to the optimal range (30-35%), after which the pump was deactivated automatically. Table 6 presents sample results from these test runs, confirming that the control logic consistently responded within an average of 40-50 seconds and stabilized soil conditions effectively.

Table 6. Sample results of IoT-based smart irrigation system performance

Beyond short-term evaluation, a longer-term assessment was conducted over 30 days under field conditions in Pagar Alam, South Sumatra, Indonesia. Measurements were taken across different times of day (morning, afternoon, evening) and weather conditions (sunny, cloudy, light rain). The system consistently maintained soil moisture between 30-45%, as shown in Figure 5, even under environmental fluctuations. The average daily water use efficiency (WUE) improved by approximately 35% compared to manual irrigation practices, attributed to more precise control and reduced water wastage.

The NPK sensor provided continuous feedback on nutrient concentrations (Nitrogen, Phosphorus, and Potassium), which were monitored to evaluate potential correlations between nutrient availability and irrigation cycles. It was observed that balanced irrigation contributed to more stable NPK levels, preventing nutrient leaching due to overwatering. These data, when visualized through the Blynk mobile interface, helped farmers make more informed decisions regarding fertilizer scheduling and nutrient management.

In addition to maintaining optimal soil moisture, the IoT system demonstrated notable adaptability across different soil textures (loamy and clayey), showing consistent response times and stable moisture regulation. Over the month-long field validation, the system exhibited 99.2% operational uptime, with no significant connectivity interruptions or sensor calibration drift. Early crop growth observations further indicated improved tiller formation and healthier leaf color compared to manually irrigated plots, suggesting potential positive impacts on yield performance.

5.jpg

Figure 5. Daily soil moisture levels recorded during testing across morning, afternoon, and evening sessions

The enhanced evaluation confirms that the IoT-based irrigation system not only performs reliably under controlled conditions but also sustains long-term stability and adaptability in diverse agricultural environments. The integration of moisture, temperature, humidity, and nutrient monitoring enables more holistic irrigation control, directly supporting precision agriculture objectives. Future research may extend the validation period to include full cropping cycles and yield quantification to further confirm the impact of the system on rice productivity and water efficiency.

4.5 Integrated framework performance

The proposed system achieves full integration between the CNN-based EfficientNet-B0 rice disease detection model and the IoT-based smart irrigation subsystem through a shared decision-support workflow. The integration is not merely parallel operation but a synergistic interaction where disease detection results dynamically influence irrigation control and nutrient management decisions.

In this integrated configuration, the EfficientNet-B0 model processes rice leaf images captured periodically via a mobile device or camera module. The model classifies the image into one of three categories (BLB, Leaf Smut, or Brown Spot) and transmits the diagnosis result to the IoT control unit via a cloud-based API. The IoT subsystem receives this disease status signal and adapts its irrigation logic accordingly:

Ÿ When a disease is detected, particularly BLB or Brown Spot that tends to thrive under high humidity, the system temporarily delays irrigation to prevent excess moisture that could worsen the infection.

Ÿ When no disease is detected or when the detected disease benefits from consistent hydration (e.g., recovery after treatment), the system resumes normal irrigation cycles based on the soil moisture threshold.

Ÿ The IoT module also records nutrient sensor (NPK) readings, which, combined with disease data, allow the system to suggest corrective fertilizer actions through the dashboard interface.

This two-way information flow between the AI model and IoT sensors forms a closed-loop adaptive system, where disease diagnosis directly modifies environmental control parameters. The integration ensures that both biotic stress (from disease) and abiotic stress (from water and nutrient conditions) are managed cohesively rather than separately.

Table 7. Comparison between integrated framework and conventional farming practices

Table 7 presents the performance comparison between the integrated framework and conventional farming practices. The integrated system achieved notable improvements: 98.42% disease detection accuracy, 30-40% water-use reduction, and approximately 35% improvement in water-use efficiency (WUE). Furthermore, the feedback mechanism between disease detection and irrigation scheduling reduced the recurrence of humidity-related diseases by 12% during controlled tests, demonstrating measurable synergy between the two modules.

These results confirm that the integration goes beyond co-existence, it represents a unified precision agriculture framework where each subsystem reinforces the other’s function. The CNN-based diagnosis informs the IoT irrigation behavior, while environmental feedback from the IoT sensors enables continuous model retraining and adaptive decision-making. The combined outcome leads to optimized yield conditions, efficient resource use, and a more resilient digital agriculture ecosystem.

4.6 Discussion and implications

The findings of this study confirm that integrating EfficientNet-B0 for rice leaf disease detection with an IoT-based smart irrigation system offers clear advantages over prior approaches. As summarized in Table 8, recent works in this domain can generally be categorized into two streams: deep learning-based disease detection and IoT-based irrigation management. However, few studies have successfully merged these two technologies into a unified, interoperable framework for decision support in precision agriculture.

Table 8. Comparison of recent studies with the proposed integrated framework

On the disease detection side, Nugroho et al. [23] demonstrated the feasibility of deploying resource-optimized CNN models on ARM-based microcontrollers, achieving real-time classification performance under strict hardware constraints. Their work effectively showed that lightweight CNNs can be adapted for edge computing environments, but it primarily emphasized model compression and inference efficiency rather than broader integration with environmental sensing or adaptive decision support. Similarly, Simhadri et al. [27] conducted a comprehensive review of deep learning architectures, identifying transfer learning and extensive data augmentation as critical strategies for improving classification performance. Meanwhile, Pai et al. [28] achieved remarkable results exceeding 98% accuracy using large, well-annotated rice disease datasets and state-of-the-art deep learning pipelines. Although these studies collectively validated the robustness of CNN-based disease classification, they remained limited to visual analytics and did not include environmental feedback or IoT-based control mechanisms.

On the irrigation management side, Morchid et al. [24] and Di Gennaro et al. [26] developed IoT-based smart irrigation systems that demonstrated substantial water savings and improved irrigation scheduling in real-world field trials. These studies effectively highlighted the potential of IoT automation to enhance resource efficiency and reduce manual intervention. However, both focused only on abiotic stress management such as soil moisture and temperature control without integrating biotic factors like disease diagnosis. As Sharma and Shivandu [25] emphasized in their review of AI and IoT integration in agriculture, most existing research still treats image-based disease detection and IoT-driven irrigation as separate and parallel developments, leaving a major gap in the realization of a unified precision agriculture ecosystem.

In contrast, the integrated framework proposed in this study bridges this gap by addressing both biotic stress (rice leaf diseases) and abiotic stress (water availability) within a unified and intelligent decision support architecture. The EfficientNet-B0 model achieved a classification accuracy of 98.42%, comparable to or exceeding other recent deep learning models, while maintaining computational efficiency suitable for deployment on low-cost hardware. At the same time, the IoT subsystem autonomously maintained soil moisture within the optimal range of 30-45% and reduced water usage by approximately 30-40% through adaptive irrigation scheduling. These complementary functionalities demonstrate the practical synergy between AI-based crop health diagnostics and IoT-based environmental control, enabling farmers to receive early warnings for disease outbreaks and respond with optimized irrigation strategies in real time.

The implications of this integration are significant. First, it illustrates how lightweight deep learning architectures such as EfficientNet-B0 can be effectively combined with affordable IoT hardware, making advanced precision agriculture technologies accessible to smallholder farmers in Southeast Asia. Second, it contributes directly to sustainable farming practices by reducing water waste, minimizing unnecessary pesticide use, and promoting environmentally efficient crop management. Third, the framework aligns with national food self-sufficiency goals by improving productivity through data-driven, automated, and adaptive farming solutions. Although current validation remains at the prototype level, the results clearly demonstrate that the proposed system establishes a meaningful connection between AI-based disease detection and IoT-enabled smart irrigation, advancing toward a more integrated and resilient model of precision agriculture.