An Edge AI and Mobile Sensing Framework for Real-Time Multi-Crop Disease Detection and Geospatial Surveillance in Smallholder Agricultural Systems

Technical Summary: An Edge AI and Mobile Sensing Framework for Real-Time Multi-Crop Disease Detection

Problem Statement
Maize and cassava are critical staple crops for food security in Nigeria and sub-Saharan Africa, yet their productivity is severely compromised by viral and fungal diseases such as Cassava Mosaic Disease (CMD), Cassava Brown Streak Disease (CBSD), Maize Leaf Blight, Maize Rust, and Maize Streak Virus. Current diagnostic methods rely heavily on manual visual inspection by experts or resource-intensive laboratory analyses, both of which are often inaccessible to smallholder farmers in rural areas due to cost, time, and lack of expertise. While deep learning models have shown promise in plant disease detection, existing solutions are often computationally expensive, require cloud connectivity, and are trained on controlled laboratory datasets, limiting their applicability in real-time, resource-constrained field conditions.

Methodology
The study proposes a smartphone-based Edge AI framework designed for offline, real-time disease detection in smallholder agricultural systems.

• Data Collection: A dataset of 10,800 leaf images was collected during the 2025 planting season from a 10-acre experimental farm in Ibadan, Nigeria. The farm was divided into three plots (mixed cassava-maize, maize-only, and cassava-only) to capture spatial variability. The dataset comprises seven classes: healthy cassava, CMD, CBSD, healthy maize, maize leaf blight, maize rust, and maize streak virus.
• Preprocessing and Augmentation: Images were resized to 224×224 pixels and normalized. To enhance generalization across field conditions, data augmentation techniques were applied, including random rotation (±30°), horizontal/vertical flipping, brightness adjustment (±20%), and random cropping (80–100%).
• Model Architecture: Four lightweight Convolutional Neural Network (CNN) architectures were evaluated: MobileNetV2, EfficientNet-Lite, ShuffleNet, and a Custom Lightweight CNN. Transfer learning was utilized where appropriate. The models were trained using categorical cross-entropy loss and the Adam optimizer, with an 80:20 train-test split.
• Deployment: Trained models were converted to TensorFlow Lite and PyTorch Mobile formats for deployment on Android smartphones. The system integrates GPS for geospatial surveillance, linking disease images to specific field coordinates.
• Mathematical Framework: The study formulated a mathematical model for disease classification, feature extraction, and a "Mobile Deployment Index" (MDI) to evaluate the trade-off between classification accuracy, inference time, and memory utilization.

Key Results

• Classification Performance: Among the tested models, EfficientNet-Lite achieved the highest overall performance with an accuracy of 94.6%, precision of 0.95, recall of 0.94, F1-score of 0.94, and an ROC-AUC of 0.97. MobileNetV2 followed with 93.8% accuracy, while ShuffleNet achieved 91.7%.
• Class-wise Accuracy: The model performed best on Maize Leaf Blight (95.2%) and Cassava Mosaic Disease (94.1%). Lower accuracy was observed for Maize Rust (92.6%) and Cassava Brown Streak Disease, attributed to visual similarities between certain disease symptoms and natural leaf discoloration.
• Mobile Deployment Metrics:
  • Inference Speed: ShuffleNet was the fastest at 65 ms, while EfficientNet-Lite required 92 ms.
  • Resource Usage: EfficientNet-Lite utilized 135 MB of memory and 41% CPU, whereas ShuffleNet used 95 MB and 32% CPU.
  • Offline Capability: All models successfully operated offline on Android devices, providing instant classification and management advice without internet connectivity.
• Deployment Index: While ShuffleNet offered superior computational efficiency (speed and memory), EfficientNet-Lite provided the optimal balance between high diagnostic accuracy and acceptable mobile performance, yielding the highest overall deployment score for practical field use.

Significance and Claims
The paper claims that this framework offers a viable, low-cost solution for enhancing crop management and food security among Nigerian smallholder farmers. By leveraging ubiquitous smartphone sensing capabilities and lightweight CNNs, the system addresses three critical barriers: lack of access to extension officers, delayed disease diagnosis, and poor internet connectivity in rural areas.

The study emphasizes that the integration of geospatial data (GPS) with image analysis allows for real-time disease surveillance and the mapping of disease spread across specific field coordinates. The authors assert that their approach moves beyond standard accuracy metrics by introducing a deployment-oriented evaluation (MDI), ensuring that selected models are not only accurate but also practical for edge devices in resource-poor environments. The findings suggest that such AI-enabled tools can empower farmers to make timely, data-driven decisions to manage crop health independently, thereby contributing to sustainable agriculture in peri-urban and rural settings.