The integration of Internet of Things (IoT) technologies in agriculture enables precision farming but introduces significant cybersecurity vulnerabilities. This paper presents FICNet (Feature Integrated Convolutional Network), a lightweight deep learning architecture for intrusion detection in agricultural IoT environments. Evaluated on the Farm-Flow AG-IoT security dataset, FICNet achieves 100% binary classification accuracy and 81.25% multiclass accuracy (macro F1: 80.43%, precision: 91.26%, ROC-AUC: 96.78%) across 8 traffic categories. A multi-dimensional component analysis confirms the contribution of each architectural component: multi-scale convolutions provide 5.3% noise robustness advantage, squeeze-and-excitation attention controls per-class detection trade-offs, and the full architecture achieves 8% data efficiency advantage over traditional baselines. Interpretability analysis via Integrated Gradients identifies header size, byte counts, and directional ratios as primary discriminative features. Comparative evaluation against 15 baselines including Transformer and GNN architectures validates FICNet’s effectiveness with only 147,092 parameters (1.81 MB), demonstrating suitability for resource-constrained agricultural IoT deployments.