EnhancedTinyCNN: A lightweight embedded AI framework for Real-Time pest detection in smart farming

This paper presents EnhancedTinyCNN, a lightweight embedded AI framework for real-time pest detection on resource-constrained microcontrollers, targeting smart farming within Integrated Pest Management (IPM) workflows. The framework combines a compact convolutional neural network designed for integer-only inference on an ESP32-S3 microcontroller, a curated image dataset with domain-adapted preprocessing, and a hardware-software platform with environmental sensing, power management, and alerting capabilities. Post-training 8-bit integer (INT8) quantization reduces model size by 75% while preserving accuracy on a three-class task (grasshopper, locust, other): 93.98% on the held-out test set (best single split) and 93.10 ± 0.56% under 5-fold cross-validation, with a 95% bootstrap confidence interval of [89.9%–96.8%]. A learning curve analysis demonstrates stable convergence. An ablation study across architecture, preprocessing, quantization, and training dimensions identifies batch normalization, data augmentation, and dropout as the most influential components. Comparative evaluation against MobileNetV3-Small and EfficientNet-B0 confirms that, trained from scratch without ImageNet pretraining—the only regime feasible on MCUs—EnhancedTinyCNN achieves the highest accuracy (93.2 ± 0.6% mean over 5 seeds) within the ESP32-S3 memory budget (834 KB INT8). The three-class taxonomy distinguishing migratory locusts from grasshoppers and background is explicitly aligned with IPM action thresholds, enabling differential alert strategies that reduce unnecessary pesticide applications in resource-constrained smallholder farming contexts. We benchmark the framework against five existing TinyML systems, demonstrating superior accuracy, a compact footprint (834 KB INT8), and complete deployment integration. All hardware specifications, software, and evaluation materials are provided to support reproducibility.