YOLOv8-Based Deep Learning for Multi-Class Farmland Intrusion Detection: Architecture, Transfer Learning, and Wavelet Coefficient Shrinkage Preprocessing

The computational vision problem of object detection and classification in outdoor agricultural settings is particularly difficult due to the combination of highly variable illumination, cluttered backgrounds, multi-scale target sizes, class imbalance, and sensor noise from low-cost IoT camera nodes. In this paper, a YOLOv8-based Convolutional Neural Network (CNN) pipeline for real-time multi-class agricultural incursion detection is designed, implemented, and rigorously evaluated. Images are categorized by the system into three security-related groups: No-Intrusion, Animal (cows, goats, and livestock), and Human Intruder. While domain-specific fine-tuning on a bespoke 1,850-image field dataset adjusts the model to agricultural circumstances, transfer learning from the Microsoft COCO dataset (330,000+ annotated images, 80 object categories) offers fundamental feature representations. Daubechies-1 (db1) wavelets with two-level decomposition are used in a specific wavelet-based picture preprocessing pipeline and soft coefficient thresholding maintains important object borders and textures while reducing sensor and ambient noise. The system obtained a mAP@50 of 98.58% on the held-out test set after APO-based hyperparameter optimization, with per-class precision above 97% in all three categories. While the Animal class showed a recall of 94.51%, which represents a 29.11 percentage point gain over the 65.4% baseline recall and directly addresses the most important safety gap in previous farmland security systems, the Human Intruder class attained perfect memory (100%). Confusion matrix study confirms operational suitability for production IoT deployment with a remarkably low false positive rate of 0.72%.