Field-condition robustness under natural acquisition conditions

Establish whether AgroVisNet maintains its reported disease-classification accuracy under field illumination, cluttered backgrounds, and partial occlusion by evaluating a genuine field-condition subset of plant images.

Background

BD-PlantDX was collected under controlled illumination and against a uniform background, whereas deployment is expected to involve natural illumination, cluttered scenes, and partially occluded leaves or roots. The paper reports synthetic corruption experiments as a proxy, finding sensitivity to brightness changes and additive noise, but these experiments do not replace evaluation on naturally acquired field imagery.

The authors explicitly identify a field-condition subset as the outstanding test needed to determine whether the high accuracy observed on the curated benchmark survives realistic deployment conditions.

References

All BD-PlantDX imagery was acquired under a controlled illumination arrangement against a uniform background, and accuracy under field illumination, cluttered backgrounds and partial occlusion is not yet measured; the synthetic corruption sweep of Table~\ref{tab:corruption} is a proxy that shows tolerance to compression and blur but sensitivity to additive noise and brightness shift, and a genuine field-condition subset remains the outstanding test.

— AgroVisNet: A lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification  (2609.10469 - Mandal et al., 9 Sep 2026) in Section 7, Discussion and Limitations; Section 8, Conclusion