Leveraging Temporal Distribution Shifts for Rice Disease Detection

Md. Radoan Islam1,*, Sayantan Chakraborty2,*, Ushashi Bhattacharjee3, Tirtho Roy3,†

1 Patuakhali Science and Technology University, Bangladesh

2 University of Dhaka, Bangladesh

3 Iowa State University, USA

*These authors contributed equally

†Corresponding author: tirtho@iastate.edu

Accepted 3rd IEEE International Conference on Computing, Applications and Systems 2026 — IEEE Xplore

IEEE Xplore (soon) Code

The cross-year evaluation framework

Cross-year evaluation framework: rice disease images published 2021 to 2025 form the training and validation pool, models are trained under full and partial fine-tuning, and evaluation happens only on a held-out 2026 partition

Train on rice-leaf images released 2021–2025, test only on 2026. Accepted at the 3rd IEEE International Conference on Computing, Applications and Systems (2026); full paper on IEEE Xplore after the conference.

In one paragraph

Does a rice-disease classifier that looks good this season still work next season? We check by holding out an entire year: four backbones (three CNNs and DINOv2), each with the backbone frozen or unfrozen, ten seeds each, trained on 2021–2025 and scored on 2026. Every setting loses more than half of its validation score on the new year, the loss is visible in the pixels themselves, and the fine-tuning choice that works best flips between the CNNs and the transformer. Numbers, tables, and code are in the paper and the repository below.

BibTeX

Status: accepted for presentation at the 3rd IEEE International Conference on Computing, Applications and Systems 2026 (9–10 October 2026); inclusion in IEEE Xplore follows presentation. The BibTeX below will be updated with page numbers and the DOI once the Xplore record is published.

@inproceedings{islam2026temporal,
  title     = {Leveraging Temporal Distribution Shifts for Rice Disease Detection},
  author    = {Islam, Md. Radoan and Chakraborty, Sayantan and
               Bhattacharjee, Ushashi and Roy, Tirtho},
  booktitle = {Proceedings of the 3rd IEEE International Conference on
               Computing, Applications and Systems},
  year      = {2026},
  publisher = {IEEE}
}

Team

Md. Radoan Islam Md. Radoan Islam

Equal contribution

Sayantan Chakraborty Sayantan Chakraborty

Equal contribution

Ushashi Bhattacharjee Ushashi Bhattacharjee

Tirtho Roy Tirtho Roy

Corresponding author