An End-to-End Deep Learning Framework for Crop Type and Phenological Stage Identification
IEEE Access, vol.14, pp.97546-97558, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 14
- Publication Date: 2026
- Doi Number: 10.1109/access.2026.3707086
- Journal Name: IEEE Access
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Page Numbers: pp.97546-97558
- Keywords: crop monitoring, deep learning, growth stage detection, Labeled crop image dataset, plant phenotyping, precision agriculture, RGB-E architecture
- Bursa Uludag University Affiliated: Yes
Abstract
In precision agriculture, crop monitoring is imperative for determining optimal harvest periods. Detecting early growth stages is non-trivial for crops with small seeds, as standard image resizing can remove crucial pixel-level details. Furthermore, an insufficiency of data in precision agriculture limits the effectiveness of AI tools. To address this data limitation, a new labeled dataset was introduced to the literature comprising four growth stages - seed, bud, flower, and harvestable fruit - of okra (Abelmoschus esculentus), pepper (Capsicum annuum), and eggplant (Solanum melongena). The images for the growth stages were collected from agricultural fields in Mugla, Türkiye, and supplemented with diverse open-source data to enhance model generalization. To organize plant growth stage detection within a single pipeline, we introduced a comprehensive framework comprising data curation, preprocessing, and deep learning architectures, including YOLO variants, Faster R-CNN, and RT-DETR-L. Dynamic Canny edge-density maps were also integrated into the deep learning models to preserve the representation of small seeds. To address the severe class imbalance in the dataset, a custom Class-Balanced Focal Loss mechanism was integrated. To ensure the validity and generalization of the developed models, a five-step ablation study and cross-distribution validation were carried out. The results showed that the developed YOLO framework can accurately achieve the 12-class prediction of crop stages, with a superior localization precision consisting of a mAP@0.5:0.95 of 72.81% and a peak mean IoU of 0.8876 for true positives. Ultimately, the primary contribution of this work is a highly valuable labeled dataset, supported by a robust detection pipeline for future plant phenology identification tasks.