Multi-Class Pest Recognition in Agricultural Environment: From Efficient Training to Deployment

Agriculture sustains the livelihoods of millions across Sierra Leone and sub-Saharan Africa, yet insect pests cause substantial crop losses each year. Traditional pest identification depends on informal farmer experience, delaying effective intervention. Although deep learning has demonstrated strong potential for automated pest recognition, most studies focus narrowly on classification accuracy and overlook the practical challenge of deploying models in resource-limited settings. This study proposes an integrated pipeline combining model training, comparative architectural evaluation, compiler-based deployment optimization, and end-user interface development, designed explicitly for the hardware constraints of sub-Saharan agricultural environments. Two architectures MobileNetV2 (a lightweight convolutional neural network) and Vision Transformer (ViT-B/16) were trained and evaluated under identical conditions on the IP102 benchmark dataset, which contains over 75,000 images across 102 pest categories. Classification performance was assessed using accuracy, precision, recall, and F1-score, alongside deployment metrics including parameter count, model size, and inference latency. Trained models were compiled using Apache TVM and executed on CPU-only hardware to assess deployment feasibility. Numerical verification confirmed that compiled models retained high consistency with PyTorch outputs (cosine similarity ≈ 0.9963). Vision Transformer achieved 69.17% test accuracy, outperforming MobileNetV2 (53.8%) by 15.37 percentage points a statistically significant difference (z ≈ 30.1, p < 0.0001) while MobileNetV2 offered substantially lower computational cost (4.2M parameters, 16 MB versus 87.3M parameters, 349 MB). A web-based application achieved average inference latency of 241 ms under CPU-only conditions. Field validation on locally collected Sierra Leonean pest images remains an important avenue for future work. These findings provide empirical guidance for selecting between accuracy-optimized and efficiency-optimized architectures and demonstrate that compiler-based deployment is a scientifically viable pathway for agricultural AI in low-resource settings.