What question did this study set out to answer?

The aim is to create a dataset for identifying primitive Indian rice panicle varieties using deep learning techniques.

April 16, 2026Open Access

A Benchmark Dataset of Primitive Indian Paddy Panicle Images and Identification via Deep Residual Transfer Learning

Key Points

The aim is to create a dataset for identifying primitive Indian rice panicle varieties using deep learning techniques.
Developed a benchmark dataset of 22 rice panicle varieties.
Applied deep residual transfer learning with ResNet-50 model.
Implemented image augmentation techniques for better training results.
Conducted an 80/10/10 split for train, validation, and test sets.
Achieved 100.0% validation accuracy on the model.
Test accuracy reached 98.74% on the held-out set.
Per-class AUC scores were perfect (1.000) for all 22 classes during validation.
Test AUCs ranged from 0.9924 to 1.000, with a mean of approximately 0.999.

Abstract

We introduce "Primitive Indian Paddy Panicle Images," a benchmark image dataset of 22 primitive Indian rice panicle varieties (Sethy, Prabira; Pamerelli, Ranjith, 2026; Mendeley Data, V1, 10.17632/khfd7pzskd.1) and present an identification approach based on deep residual transfer learning. Using a transfer-learned ResNet-50 with image augmentation and an 80/10/10 train/validation/test split, the model attains 100.0% validation accuracy and 98.74% accuracy on the held-out test set. Per-class one-vs-rest AUCs on validation are 1.000 for all 22 classes; test AUCs range from 0.9924 to 1.000 (mean ≈ 0.999), with separate confusion matrices and ROC curves provided for validation and test partitions. These results demonstrate that deep residual transfer learning can robustly discriminate closely related panicle morphotypes when trained on a carefully curated dataset. We release the dataset to support reproducible research in germplasm identification, varietal purity assessment, and automated phenotyping.

A Benchmark Dataset of Primitive Indian Paddy Panicle Images and Identification via Deep Residual Transfer Learning

Key Points

Abstract

Cite This Study