Akademik Veri Yönetim Sistemi
APPLIED FRUIT SCIENCE, cilt.68, sa.1, ss.1-14, 2026 (SCI-Expanded, Scopus)
In real-world agricultural environments, the most accurate deep learning model is often not the most deployable one due to strict constraints on memory, computational capacity, and real-time inference requirements. This study introduces a deployment-oriented tiny machine learning (TinyML) framework for strawberry ripeness classification under limited data and resource-constrained conditions. To enable robust evaluation, the Strawberry-DS dataset was restructured into a leakage-free four-class classification set (Green, White, Turning, and Red) using object-level cropping and scene-level partitioning. A stratified five-fold cross-validation protocol was adopted to obtain statistically reliable performance estimates. In addition, a resolution ablation study was conducted across three input scales (64 × 64, 96 × 96, and 128 × 128) to investigate the trade-off between visual detail and computational efficiency. Three lightweight architectures—MobileNetV2 with transfer learning, ShuffleNetV2, and a parameter-efficient tiny modular convolutional neural network (TM-CNN)—were evaluated under a unified experimental framework. Generalization performance was further supported through regularization strategies and knowledge distillation. Experimental results show that MobileNetV2 achieves the highest predictive performance, whereas the proposed TM-CNN provides a more favorable balance between computational efficiency, memory footprint, and inference latency. These findings highlight that, for TinyML-based agricultural applications, model selection should not be driven solely by accuracy but by a balanced consideration of deployability constraints. The proposed framework provides a practical methodology for designing real-time, resource-aware perception systems in precision horticulture.