Real-time monitoring of oyster mushroom cultivation using CCTV and attention-enhanced ShuffleNet-based explainable AI techniques

Timely identification of oyster mushroom maturity is critical for optimizing harvest schedules, improving quality, and reducing post-harvest losses. This paper presents a real-time monitoring framework that integrates lightweight deep learning with explainable artificial intelligence (XAI). Using a curated dataset of 20,000 annotated CCTV images collected under diverse cultiva- tion conditions, we evaluate multiple architectures including EfficientNet, DenseNet, VGG, and Vision Transformers. Among these, a ShuffleNetV2 model enhanced with a self-attention mechanism achieved superior perfor- mance, reaching 99.70 % accuracy through 10-fold cross-validation while main- taining a compact size of 1.3 M parameters. This efficiency enables deploy- ment on edge devices suitable for smart farms. To ensure transparency, XAI techniques such as Grad-CAM, Grad-CAM++, EigenCAM, and LIME were applied, revealing biologically meaningful regions used in decision-making. The system was further integrated with real-time CCTV streams, demon- strating feasibility for on-farm maturity classification with low latency. By combining accuracy, efficiency, and interpretability, the proposed framework offers a scalable solution for precision agriculture and contributes a reusable model for intelligent mushroom cultivation systems.