Automated Plant Disease and Pest Detection System Using Hybrid Lightweight CNN-MobileViT Models for Diagnosis of Indigenous Crops

AI summary

Overview

Research area: Applied computer vision for agriculture — lightweight deep learning for plant disease and pest detection on edge devices, with a focus on an indigenous crop in post-conflict Ethiopia.

Technical level: Intermediate to Advanced. The paper assumes familiarity with convolutional neural networks, Vision Transformers, self-attention, quantization, and mobile deployment runtimes, though the conceptual arguments are explained in accessible terms.

Scope (one sentence): The paper benchmarks a custom lightweight CNN, EfficientNet-Lite1, and the MobileViT-XS hybrid on a newly curated 3,587-image dataset of cactus-fig (Opuntia ficus-indica) field imagery, then deploys the models inside an offline, Tigrigna- and Amharic-localized Android application for Cortex-A53-class devices.

What This Paper Is About

Farmers in Ethiopia's Tigray region grow cactus-fig ("Beles") as a buffer crop against seasonal food scarcity, but the invasive cochineal insect and fungal rots produce visually ambiguous symptoms — white waxy pest clusters versus mildew-like fungal textures — that are hard to tell apart without expert help, and the region's damaged infrastructure has cut off that expert help. The paper asks whether an attention-based hybrid model can resolve this ambiguity better than pure convolutional models, and whether either can run fully offline on the low-end Android hardware farmers actually own. To answer this, the authors built a field-image dataset for the crop and compared three mobile-efficient architectures on accuracy, size, latency, and failure modes.

Key Contributions

1. Indigenous dataset construction. A dataset of 3,587 annotated Opuntia ficus-indica images collected at Mekelle University and Adigrat University experimental sites, split into three classes — Affected (1,500), Healthy (1,500), and No Cactus (587) — captured in uncontrolled field conditions with dust, shadows, and mixed infections. The dataset is open-sourced via the project repository and Kaggle.

2. Hybrid architecture benchmarking. A controlled comparison of inductive bias versus self-attention — a ~1.2M-parameter custom CNN, EfficientNet-Lite1 (4.6M), and MobileViT-XS (~2.3M) — measuring accuracy, F1, parameter count, model size, and latency, without cross-crop knowledge leakage.

3. Failure-mode isolation. An explicit analysis of why CNNs fail on this crop, showing that the "scarring vs. lesion" texture ambiguity accounts for 68% of the custom CNN's errors, and that the MobileViT-XS's global self-attention reduces false positives by 94% relative to the CNN baselines.

4. Offline-first deployment. Float16 post-training quantization and export to the TensorFlow Lite runtime, wrapped in a Tigrigna- and Amharic-localized Flutter Android application that performs inference on ARM Cortex-A53-class devices with no internet connectivity.

Main Findings

• Accuracy ranking. MobileViT-XS reached 97.3% accuracy with F1 0.98; EfficientNet-Lite1 reached 90.7% with F1 0.90; the custom CNN reached 89.5% with F1 0.89. The abstract and conclusion describe the 97.3% figure as mean cross-validation accuracy, while Table I lists it as test-set accuracy on the held-out set of N = 1,195.

• Efficiency ranking. The custom CNN had the most favorable deployment profile at 42 ms latency and 4.8 MB model size with 1.2M parameters. EfficientNet-Lite1 measured 55 ms and 19.0 MB with 4.6M parameters. MobileViT-XS measured 68 ms and 9.3 MB with 2.3M parameters.

• Failure mode of convolutional models. Both the custom CNN and EfficientNet-Lite1 misclassified older, scarred Healthy cladodes as Affected. This "scarring vs. lesion" ambiguity accounted for 68% of the custom CNN's errors, because local convolutional filters cannot reliably distinguish a physical scar's rough texture from necrotic fungal rot.

• Global context resolves the ambiguity. MobileViT-XS achieved precision of 0.97 for the Affected class. Its multi-head self-attention learned that cochineal infestations appear as clusters of white powder, whereas scarring appears as random, isolated patches. This let it reject 94% of the false positives produced by the CNN baselines.

• Early-stage detection. Under Scenario A, MobileViT detected minute cochineal specks covering less than 5% of surface area that EfficientNet missed.

• Background rejection. Under Scenario B, the No Cactus background class (soil, sky, weeds) was classified with 99% precision by all models. MobileViT was more robust in mixed frames where a healthy cactus was partially occluded by weeds.

• Lighting invariance. Under Scenario C, harsh midday sunlight and specular reflection on waxy pads caused the custom CNN to occasionally predict false lesions from glare, while MobileViT's representations remained invariant to these local intensity shifts.

• Pareto trade-off. The results support a tiered deployment strategy: the custom CNN as a default real-time scanner on legacy hardware, and MobileViT-XS as a selective high-precision verifier for visually ambiguous cases.

Methodology in Plain English

The authors first assembled a background dataset of 26,394 images from public repositories such as PlantVillage, covering tomato, potato, maize, and apple, to pre-train the early feature-extraction layers before fine-tuning on the target crop. The core cactus-fig dataset was then collected in the field and labeled into three classes, including a deliberate "No Cactus" rejection class so the system would not fire on soil, sky, or weeds.

Images were resized to 256 × 256 pixels and normalized to the range [0, 1]. To avoid overfitting on a small indigenous dataset, the authors used what they call a "Sanitized Validation Protocol": geometric augmentations — rotation of ±20 degrees, zoom of ±15%, and horizontal flipping — were applied only to the training split, while validation used clean, static images.

They then trained and compared three models. The custom lightweight CNN uses three sequential convolutional blocks with ReLU activations and pooling, deliberately keeping a small receptive field to capture local texture anomalies cheaply. EfficientNet-Lite1 serves as a mid-range baseline. MobileViT-XS is a hybrid that inserts multi-head self-attention blocks inside a lightweight CNN backbone, so the model can attend to the whole image at once — which is what lets it tell scattered pest clusters apart from isolated scars. MobileViT-XS was fine-tuned with the AdamW optimizer and a cosine decay learning rate schedule.

Training used an NVIDIA Tesla P100 GPU with 16 GB VRAM, PyTorch, and the timm library, with 3-fold stratified cross-validation, 50 epochs, AdamW at a learning rate of 2e-4, weight decay of 0.05, and a cosine decay scheduler. Results were reported on a held-out test set of 1,195 images.

For deployment, models were exported to TensorFlow Lite and compressed with Float16 post-training quantization, which maps 32-bit floating-point weights to lower precision using a scaling factor and zero-point. The custom CNN runs as the real-time scanner; MobileViT-XS is invoked selectively for ambiguous cases. LIME saliency maps were generated to show that the model attends to pest or lesion regions rather than background artifacts.

Why This Matters

Impact on research. The paper targets a documented gap: agricultural AI is heavily biased toward globally commercial broad-leaf crops, and prior Transformer work on cactus pathology (CactiViT) targeted high-end smartphones. This work benchmarks indigenous cactus-fig field imagery with explicit failure-mode isolation and transfer-leakage prevention on hardware representative of post-conflict zones. It also frames the CNN-versus-attention trade-off not as a simple accuracy contest but as a deployment decision with measurable cost.

Real-world applications:

• Field diagnosis for smallholder cactus-fig farmers in Tigray and comparable highland agro-systems, delivered in Tigrigna and Amharic with no internet requirement.
• Early warning for cochineal infestation, which the paper links to large-scale yield devastation if not caught early.
• Background-rejection screening that avoids false alarms from soil, sky, and weeds in real field photos.
• A tiered triage workflow: fast local screening on the device, with harder cases routed to the higher-precision model.

Industry relevance. The Pareto results — 42 ms and 4.8 MB versus 68 ms and 9.3 MB — are directly actionable for anyone shipping on-device vision on Cortex-A53-class hardware. The combination of Float16 post-training quantization, TensorFlow Lite, a Flutter client, and a Django backend (listed among the index terms) is a reproducible reference stack for low-resource edge AI. The paper also argues that explainability via LIME is a trust requirement, not a nice-to-have, for adoption in agrarian communities.

Future Directions

• Hybrid deployment. Implementing the tiered system fully: the lightweight CNN scans in real time while ambiguous cases are flagged for MobileViT analysis when connectivity allows.
• Dataset expansion. Extending the indigenous dataset to other regional staples such as Teff and Sorghum, and to additional pests such as the Fall Armyworm.
• UAV integration. Adapting the lightweight model for low-cost drones to monitor cactus plantations in inaccessible terrain at scale.
• Open question on evaluation consistency. The paper reports the 97.3% figure as mean cross-validation accuracy in the abstract and conclusion but as test accuracy in Table I, so a clearer separation of cross-validation and held-out test protocol would strengthen future comparisons.

Target Audience

Researchers and practitioners in agricultural computer vision and edge AI, especially those working on CNN-versus-Transformer trade-offs, domain shift in field imagery, and quantization for mobile deployment. It is also relevant to ICT4D and food-security researchers interested in localization and offline-first design, to engineers building on-device inference pipelines on ARM Cortex-A53-class hardware, and to agricultural extension programs in resource-constrained or post-conflict regions. The dataset and source code are publicly available at https://github.com/Tekleab15/Automated_plant_disease_and_pest_detection_system.

Authors’ abstract

Agriculture supports over 80% of the population in the Tigray region of Ethiopia, where infrastructural disruptions limit access to expert crop disease diagnosis. We present an offline-first detection system centered on a newly curated indigenous cactus-fig (Opuntia ficus-indica) dataset consisting of 3,587 field images across three core symptom classes. Given deployment constraints in post-conflict edge environments, we benchmark three mobile-efficient architectures: a custom lightweight CNN, EfficientNet-Lite1, and the CNN-Transformer hybrid MobileViT-XS. While the broader system contains independent modules for potato, apple, and corn, this study isolates cactus-fig model performance to evaluate attention sensitivity and inductive bias transfer on indigenous morphology alone. Results establish a clear Pareto trade-off: EfficientNet-Lite1 achieves 90.7% test accuracy, the lightweight CNN reaches 89.5% with the most favorable deployment profile (42 ms inference latency, 4.8 MB model size), and MobileViT-XS delivers 97.3% mean cross-validation accuracy, demonstrating that MHSA-based global reasoning disambiguates pest clusters from two dimensional fungal lesions more reliably than local texture CNN kernels. The ARM compatible models are deployed in a Tigrigna and Amharic localized Flutter application supporting fully offline inference on Cortex-A53 class devices, strengthening inclusivity for food security critical diagnostics.