An Edge-Deployable Intelligent Decision Support System for Chili Leaf Disease Diagnosis Integrating Lightweight Convolutional Neural Network and Large Language Model-Based Agronomic Assistance

Agriculture constitutes one of the most fundamental pillars of the Indonesian national economy, contributing substantially to gross domestic product, employment, and rural food security. Among the horticultural commodities cultivated across the Indonesian archipelago, chili (Capsicum annuum L.) holds a position of particular strategic importance. Chili peppers are widely consumed in Indonesian cuisine and serve as a primary ingredient in domestic food manufacturing, making them a price-sensitive commodity that directly influences the national Consumer Price Index and food inflation rate [1]. Any disruption to chili production, whether from climate anomalies, pest pressure, or disease outbreaks, can trigger cascading economic effects across the agricultural supply chain.

Despite the economic significance of chili cultivation, the crop remains highly susceptible to a broad spectrum of leaf diseases caused by bacterial, fungal, viral, and nutritional agents. These diseases, including bacterial spot caused by Xanthomonas campestris, Cercospora leaf spot caused by Cercospora capsici, curl virus transmitted by Bemisia tabaci whitefly, and physiological nutrient deficiency disorders, frequently emerge during the vegetative and early generative stages of growth [2, 3]. The visual symptoms of these diseases often overlap considerably, especially in early infection phases when leaf lesion sizes and color patterns are insufficient to distinguish one pathogen from another with certainty. Globally, plant diseases and pests are estimated to reduce crop yields by 20–40% annually, a figure consistent with field observations in Indonesian chili farming contexts [4, 5].

Current agricultural disease management practices in Indonesia remain predominantly dependent on manual visual inspection conducted by farmers, agricultural extension officers, or agronomic consultants. This approach suffers from inherent limitations: it is subjective in nature, strongly dependent on the experience and contextual knowledge of the inspector, inconsistent under variable field lighting and leaf orientation conditions, and impossible to scale for simultaneous monitoring across large cultivation areas [6]. Furthermore, smallholder farmers, who constitute the majority of Indonesian chili producers, frequently lack access to trained agricultural diagnosticians, resulting in misdiagnosis, inappropriate fungicide or bactericide applications, and ultimately increased production costs with reduced efficacy.

The emergence of Agriculture 4.0 has catalyzed the integration of digital intelligence technologies, including artificial intelligence (AI), Internet of Things (IoT), Remote Sensing, and Computer Vision, into plant health monitoring workflows [7, 8]. Among deep learning architectures, Convolutional Neural Networks (CNNs) have demonstrated consistently superior performance in extracting high-level visual features such as color gradients, lesion texture, and morphological patterns from digital leaf images, significantly outperforming traditional machine learning classifiers that require hand-engineered feature descriptors [9, 10]. Multiple studies have confirmed CNN accuracy in the range of 95-99% for plant leaf disease classification under controlled experimental conditions [11].

However, a critical limitation of existing CNN-based plant disease detection systems is that they return only a disease class label and a confidence score, without providing farmers with any explanation of the disease mechanism, its biological cause, or its recommended management strategy. This output gap represents a significant barrier to practical adoption, since a label alone is insufficient to guide a smallholder farmer's treatment decision. Large Language Models (LLMs) represent a compelling solution to this gap, given their ability to generate fluent, contextually appropriate natural-language text that can translate classification outputs into actionable agronomic advice [12, 13].

While CNN-based mobile plant disease detection applications and LLM-powered agricultural advisory systems have each received independent attention in the literature, their systematic integration into a unified, on-device mobile system with structured prompt engineering and rigorously validated LLM outputs remains underexplored. The present study directly addresses this gap through three key contributions that distinguish it from prior work: (1) the CNN model is optimized and converted to TensorFlow Lite (TFLite) format for offline on-device inference, eliminating the cloud dependency and connectivity constraints that limit the usability of server-side solutions in rural agricultural environments; (2) a structured prompt-engineering pipeline is designed to translate CNN classification outputs into semantically constrained LLM queries, reducing the risk of hallucination and ensuring agronomically grounded responses; and (3) the quality and reliability of LLM-generated outputs are validated through expert review by certified agronomists using a structured Likert evaluation protocol, a validation dimension that is absent from most comparable integrated systems.

The application is designed for deployment on Android smartphones, which are the most widely accessible and affordable computing devices among Indonesian smallholder farmers. The Flutter cross-platform framework was employed for application development to ensure broad compatibility across Android device specifications and application programming interface (API) levels [14]. The system aims to democratize expert-level disease diagnosis, enabling real-time, field-based identification and management guidance without requiring specialized agricultural training or reliable internet connectivity.

The remainder of this article is organized as follows. Section 2 presents a comparative review of related work and positions the current study's contributions. Section 3 details the research methodology, including dataset description, CNN architecture design, LLM integration with prompt engineering, and evaluation framework. Section 4 reports the experimental results and discussion. Section 5 concludes the study and identifies directions for future research.

This research adopts a Research and Development (R&D) methodology adapted to the development of intelligent mobile agricultural systems. The R&D cycle encompasses five sequential phases: (1) requirement and system design analysis; (2) dataset preparation and preprocessing; (3) CNN model design, training, and TFLite conversion; (4) LLM integration with structured prompt engineering and hallucination mitigation; and (5) Android application implementation and multi-dimensional evaluation. This methodology was selected because it enables iterative refinement across both the machine learning model components and the mobile application interface, ensuring that the final system satisfies both technical performance criteria and practical usability requirements for field deployment [4].

3.1 System requirement analysis

The system was designed around four operational requirements identified through a preliminary stakeholder needs assessment. First, the classification module must identify seven chili leaf conditions, namely healthy leaf, bacterial spot, Cercospora leaf spot, curl virus, white spot, nutrient deficiency, and whitefly infestation, from a single leaf photograph taken under uncontrolled field conditions. Second, the classification must execute locally on the device without internet connectivity to ensure usability in areas with limited network infrastructure. Third, when internet connectivity is available, the system must generate natural-language explanations of the identified disease and context-appropriate management recommendations. Fourth, the application must maintain a persistent history of previous analyses to support longitudinal plant health monitoring within a growing season. These requirements drove the architectural decision to combine an on-device TFLite CNN with an API-based LLM and a local database for history storage.

3.2 Dataset description and preprocessing

The experimental dataset was sourced from the publicly available Chili Plant Leaf Disease and Growth Stage Dataset published on Mendeley Data by Nirob et al. [15]. This dataset was selected based on its comprehensive coverage of clinically relevant chili disease categories prevalent in Southeast Asian cultivation environments, its inclusion of images captured under diverse natural lighting conditions, and its public accessibility for reproducibility. The full dataset comprises 6,714 labeled images distributed across seven disease and health categories.

Prior to training, all images underwent a standardized preprocessing pipeline to ensure consistency of input representation. Each image was resized to 224 × 224 pixels to conform to the CNN input layer specification while preserving sufficient spatial resolution for texture and lesion pattern discrimination. Pixel intensity values were normalized to the [0, 1] range by dividing by 255, which stabilizes gradient descent convergence by reducing the scale disparity between input features. Duplicate images were identified and removed using perceptual hash comparison to prevent data leakage between training and evaluation sets.

To improve model generalization and reduce susceptibility to overfitting on the training distribution, a comprehensive data augmentation strategy was applied dynamically during training. The augmentation operations included: random rotation in the range of ±30°, horizontal and vertical flipping, random zoom in the range of ±20%, and brightness adjustment in the range of ±20%. These augmentations simulate the natural variability in leaf orientation, camera angle, ambient lighting, and partial occlusion that occur in realistic field photography conditions, effectively expanding the functional diversity of the training distribution without requiring additional data collection.

Table 2. Dataset class distribution across training, validation, and test splits

The dataset was partitioned into training (70%), validation (15%), and test (15%) subsets using stratified random sampling to ensure proportional class representation across all splits. Stratified sampling was critical given potential class imbalance in the source dataset, as it prevents any single class from being systematically underrepresented in the evaluation set. Table 2 presents the resulting per-class image counts across dataset splits.

As illustrated in Table 2, the class distribution is approximately balanced across the seven categories, with total counts ranging from 871 (whitefly) to 1,025 (healthy leaf). The maximum inter-class imbalance ratio is 1.18:1, which falls within the acceptable range for CNN training without requiring oversampling corrections. This near-balanced distribution contributes to the stability of macro-averaged evaluation metrics across classes.

Figure 1 illustrates the complete dataset preprocessing and class organization workflow. Raw input images from each of the seven disease categories are subjected to pixel normalization (scaling to [0, 1]), resized to 224 × 224 pixels, and augmented. The resulting preprocessed images are then partitioned into training, validation, and test subsets before being fed into the CNN training pipeline.

7-1.png

Figure 1. Dataset preprocessing workflow for chili leaf disease classification

3.3 Convolutional Neural Network model architecture and training configuration

The CNN model was designed and implemented using TensorFlow 2.x with the Keras high-level API. The architectural design philosophy prioritized a balance between classification accuracy and computational efficiency, as the model must execute within the memory and processing constraints of mid-range Android smartphones. The adopted architecture follows a sequential convolutional-pooling feature extraction backbone with a fully connected classification head, as summarized in Table 3.

Table 3. Convolutional Neural Network (CNN) architecture -layer-by-layer summary

The architecture comprises three convolutional blocks, each consisting of a Conv2D layer followed by a MaxPooling2D layer. The first convolutional block applies 32 filters of size 3 × 3 with ReLU activation and same-padding, preserving the spatial dimensions at the input resolution of 224 × 224 pixels. MaxPooling2D with a 2 × 2 pool size and stride 2 then halves both spatial dimensions to 112 × 112, progressively concentrating spatial information while reducing computational load. The second and third convolutional blocks apply 64 and 128 filters, respectively, following the same convolution-pooling structure. This progressive increase in filter depth allows the network to learn increasingly abstract feature representations, from low-level edge and color features in the early layers to disease-specific texture and lesion pattern features in the deeper layers.

Following the third pooling layer, the 28 × 28 × 128 feature maps are flattened into a 100,352-dimensional vector and passed through a Dense layer with 256 units and ReLU activation. A Dropout layer with a rate of 0.5 is applied after the Dense layer to randomly deactivate half of the neurons during each training step, which is the primary regularization mechanism for preventing overfitting given the relatively large parameter count of the Dense layer. The output layer consists of 7 units with Softmax activation, producing a probability distribution over the seven disease classes.

Figure 2 illustrates the CNN architecture adopted in this study, showing the sequential arrangement of convolutional, pooling, and fully connected layers from the 224 × 224 × 3 input to the 7-class probability output vector.

7-2.png

Figure 2. Convolutional Neural Network (CNN)-based chili leaf disease classification architecture

The model was compiled with the Adam optimizer [21] at an initial learning rate of 0.001, using categorical cross-entropy as the loss function. The training progression is visualized in Figure 3, which demonstrates the model's learning trajectory across epochs with convergence occurring at epoch 58. The batch size was set to 32, and training was conducted for a maximum of 100 epochs. Two callbacks were employed to prevent overfitting and improve training efficiency: Early Stopping with a patience of 10 epochs monitoring validation loss, which terminated training when validation performance ceased to improve; and ReduceLROnPlateau with a factor of 0.5 and patience of 5 epochs, which halved the learning rate upon validation loss plateaus to facilitate finer convergence. Training was executed on a Google Colab environment with an NVIDIA Tesla T4 GPU, completing in approximately 3.5 hours. The trained model was subsequently converted to TFLite format using the TFLite Converter with Float32 precision, producing a model file of 98.3 MB for on-device deployment.

7-3.png

Figure 3. Training and validation accuracy and loss curves (early stopping at epoch 58)

3.4 Large Language Model integration and structured prompt engineering

The LLM component of the system was designed to address the critical limitation of label-only classification output by generating natural-language explanations that describe the identified disease, its causal organism, characteristic visual symptoms, and recommended control and prevention measures appropriate to smallholder chili farming contexts in Indonesia. GPT-4o (OpenAI API, model version: GPT-4o-2024-08-06) was selected as the underlying LLM for this integration, based on its demonstrated performance in domain-specific knowledge generation tasks, its structured output reliability, and its support for system-level role definition that enables behavior constraint through prompt engineering [12, 19].

The LLM integration is implemented through a structured prompt-engineering pipeline, which constitutes the primary technical novelty of the LLM component. Rather than issuing open-ended queries to the LLM API, the system constructs a constrained prompt from a fixed template that incorporates the CNN classification output. This structured approach is motivated by established findings in the prompt engineering literature that unconstrained queries to LLMs significantly increase hallucination probability, particularly in specialized domain contexts where the LLM's internal knowledge may conflict with ground-truth agricultural science [12]. The prompt template applied in this system is as follows:

This structured prompt template constrains the LLM response space in four critical ways. First, the explicit system role definition anchors the model's persona to certified agronomic expertise, reducing the probability of generic or non-agricultural responses. Second, embedding the specific disease label from the CNN output prevents the LLM from speculating about alternative diagnoses. Third, the four-part structured output format enforces response organization and ensures that all agronomically critical information types (etiology, symptomology, management, follow-up) are consistently present. Fourth, the inclusion of the CNN confidence score as an explicit input parameter enables the LLM to conditionally escalate its response when model certainty is low, advising farmers to consult a human expert rather than acting on a potentially unreliable classification.

To further mitigate hallucination risk and improve response consistency, the following API parameters were applied: temperature = 0.3 (low randomness to favor factual precision over creative variation); maximum output tokens = 400 (preventing excessively verbose responses); and top_p = 0.9 (nucleus sampling to maintain output diversity within high-probability token space). An offline fallback mechanism was implemented for cases where internet connectivity is unavailable: pre-authored static summaries for each of the seven disease categories are stored in the application's local asset bundle and displayed in place of LLM-generated content when API access fails. Average LLM API response latency was measured at 1,240 ms under standard 4G mobile network conditions across 30 consecutive test queries, which is acceptable for non-time-critical advisory use cases.

Figure 4 illustrates the CNN and LLM integration workflow, depicting the data flow from the Android device interface through the TFLite inference engine to the cloud-based LLM API and back to the application display layer. The workflow shows three core processing layers: the application interface layer (user-facing input/output), the device layer (on-device TFLite inference), and the cloud service layer (LLM API query and response processing).

7-4.png

Figure 4. Convolutional Neural Network (CNN) and Large Language Model (LLM) integration workflow across application interface, device, and cloud service layers

3.5 Android application implementation

The Android application was developed using the Flutter framework (version 3.10) with Dart as the primary programming language, targeting Android API level 21 (Android 5.0) and above to ensure compatibility with the majority of Android devices currently deployed in Indonesian agricultural contexts [14, 20]. Flutter's widget-based cross-platform architecture was selected to facilitate potential future extension of the application to iOS platforms without substantial code restructuring.

The TFLite CNN model was integrated using the tflite_flutter plugin (version 0.10.4), which provides native bindings to the TFLite interpreter for efficient on-device inference. The model file (98.3 MB) is bundled as a local asset, enabling fully offline classification without any network request. The application maintains an SQLite local database using the sqflite package for persistent storage of analysis history records, each of which contains the leaf image thumbnail, classification result, confidence score, LLM-generated explanation (or static fallback), and analysis timestamp.

The image acquisition workflow supports two input modalities: real-time camera capture using the device's rear-facing camera through the image_picker plugin, and gallery image selection for retrospective analysis of previously captured photographs. Prior to inference, the input image undergoes the same preprocessing pipeline applied during training: resizing to 224 × 224 pixels and normalization to the [0, 1] float range. The TFLite interpreter returns a 7-element probability vector, from which the argmax operation identifies the predicted class, and the corresponding probability value is reported as the confidence score.

Figure 5 shows the activity diagram of the complete application workflow. The diagram illustrates the sequential interaction between the user, the Android application interface, the TFLite inference engine, and the LLM API, including the conditional branching logic for online versus offline operation and the history recording functionality.

7-5.png

Figure 5. Activity diagram of Android-based chili leaf disease classification system

3.6 Evaluation framework

The evaluation of the proposed system was conducted across four complementary dimensions to provide a comprehensive assessment of both technical performance and practical deployment viability.

•Classification performance evaluation: The CNN model was evaluated on the held-out test set of 1,003 images using standard classification metrics: accuracy, precision, recall, and F1-score, computed both at the macro-average level and individually per disease class. The confusion matrix was generated to identify inter-class misclassification patterns. All metrics were computed using sklearn.metrics from scikit-learn 1.3.0.

•Mobile benchmarking: On-device inference performance was measured on a Xiaomi Redmi Note 11 (Qualcomm Snapdragon 680 processor, 4 GB RAM, Android 12). Inference latency was recorded as the mean duration over 100 consecutive inference calls using the device's System Clock. Elapsed Realtime timer. Memory consumption was measured via Android Profiler during active inference. Battery consumption was estimated using Android's Battery Manager API across a standardized 100-inference test sequence.

•LLM output expert validation: Two certified agronomists with a minimum of five years of experience in chili disease management independently evaluated 35 LLM-generated responses (five per disease class) using a structured 5-point Likert scale across three criteria: (1) factual accuracy of disease description and etiology, (2) relevance and practicability of management recommendations for Indonesian smallholder farming conditions, and (3) language clarity and comprehensibility for a non-expert audience. Inter-rater reliability was computed using Cohen's Kappa (κ).

•Functional black-box testing: Systematic functional testing was conducted to verify correct behavior of all application modules under normal, edge-case, and error conditions, including image input handling, classification and confidence display, LLM response generation, offline fallback activation, history record creation and retrieval, and guide module navigation.

This study developed, implemented, and evaluated an Android-based chili leaf disease classification system that integrates a CNN for visual disease classification with a LLM for contextual natural-language explanation and management recommendation generation. The system makes three original contributions to the field of intelligent agricultural decision-support systems. First, a lightweight CNN architecture was designed and converted to TFLite format for fully offline on-device inference on mid-range Android smartphones, achieving 99.10% overall classification accuracy across seven disease categories with an inference latency of 87 ms. Second, a structured prompt-engineering pipeline was developed to translate CNN classification outputs into semantically constrained LLM queries, incorporating disease labels, confidence scores, and explicit output format requirements to minimize hallucination risk and ensure agronomically relevant responses. Third, LLM output quality was systematically validated by two certified agronomists using a structured Likert evaluation protocol, yielding an overall mean score of 4.63/5.00 with Cohen's Kappa of 0.81, providing the first quantitative quality assurance evidence for an integrated CNN + LLM plant disease advisory system of this type.

Mobile benchmarking demonstrated that the system is practically deployable on mid-range Android devices widely accessible to Indonesian smallholder farmers, with a peak RAM footprint of 48 MB, battery consumption of 0.31 mAh per 100 inferences, and offline classification capability that eliminates dependency on network infrastructure. The LLM API response latency of 1,240 ms under 4G conditions is acceptable for advisory use and is supplemented by a static offline fallback for connectivity-limited environments. The complementary guide & education module and analysis history feature further enhance the system's value as a comprehensive field-deployable precision agriculture tool aligned with the Agriculture 4.0 transformation agenda in Indonesia.

Several limitations must be acknowledged to contextualize the reported findings appropriately. The CNN classification performance was evaluated exclusively on the Mendeley Chili Plant Leaf Disease dataset, which was collected under partially controlled conditions [23]. Real-world field deployment is expected to introduce accuracy-reducing factors including natural background complexity, variable illumination, motion blur, and mixed-symptom co-infection presentations, which are not represented in the current evaluation. The LLM component incurs API usage costs and requires internet connectivity for full functionality. The system's disease scope is currently limited to chili as a single crop species. The TFLite model file size of 98.3 MB may constrain deployment on devices with limited internal storage.

Future work should address these limitations across five research directions. First, a field evaluation campaign should be conducted to collect annotated leaf images from operational chili farms under natural conditions and use these to re-evaluate and retrain the model for real-world robustness. Second, Gradient-weighted Class Activation Mapping (Grad-CAM) should be integrated to generate saliency map overlays that visualize the CNN's decision-relevant image regions, providing both transparency for expert validation and educational value for farmer users. Third, post-training quantization (INT8) should be applied to reduce the TFLite model file size and improve inference speed while characterizing the accuracy trade-off. Fourth, the system should be extended to additional high-value horticultural crops in Indonesia, such as tomato, shallot, and potato, through multi-task or transfer learning frameworks. Fifth, the feasibility of replacing the cloud-based LLM with a locally deployed lightweight open-source language model (e.g., Phi-3-mini or TinyLlama) should be investigated to eliminate API dependency and operational cost.