A benchmark dataset for primitive Indian paddy field images with deep learning based classification.

We introduce a primitive Indian paddy field image benchmark dataset and deep-learning-based classification baseline for automated variety identification task. The dataset (sethy, PRABIRA; Pamirelli, Ranjith (2026), "Primitive Indian Paddy On-Field Images", Mendeley Data, V1, https://doi.org/10.17632/jthrm4mj5r.1) includes 33 primitive rice varieties each consisting of 100 field images (3400 images in total). The images were captured from Regional Research and Technology Transfer Station (RRTTS), Chiplima field margins and native habitats. RRTTS-Chiplima is one of the regional centres under Orissa University of Agriculture & Technology (OUAT), Sambalpur, Odisha, India. We utilize deep features learnt from pretrained NASNet-Large model (input image size: 224 × 224) and train one-vs-all linear Error Correcting Output Codes (ECOC) classifier using features from prediction layer of the network. We split the data stratified with 80% training images and 20% test images. Results include accuracy of 0.9424, sensitivity of 0.9424, specificity of 0.9982, precision of 0.9536, false positive rate of 0.0018, F1-score of 0.9439, Matthews correlation coefficient (MCC) of 0.9443 and Cohen's kappa of 0.9203. The total computation time for this pipeline is ∼82,811 s. We hope that this benchmark dataset and the baseline results reported here will help the researchers working towards crop monitoring, crop varietal identification and precision agriculture research.