Abstract The reduction in soybean yield and quality due to drought makes rapid and non-destructive grading at the seed level critical. In this study, five genotypes × four irrigation treatments (20 classes) were discriminated using RGB images in a scalable DL pipeline. ResNet-18, ConvNeXt-Tiny/Base, DenseNet-121, and EfficientNet-B1 were trained using transfer learning. HPO was performed with Optuna, and the accuracy, F1 score, ROC-AUC, and resource consumption were reported. On a held-out test set that was disjoint from model selection, EfficientNet-B1 achieved the best accuracy of 94.8% (AUC = 0.9990), followed by DenseNet-121 96.2% (AUC = 0.9993), ResNet-18 92.9% (AUC = 0.9985), ConvNeXt-Base 92.1% (AUC = 0.9981) and ConvNeXt-Tiny 94.0% (AUC = 0.9985). In terms of efficiency, ResNet-18 was balanced at 2232 s and 46.9 MB, with the lowest inference latency (15.5 ms). Grad-CAM foci pointed to the hilum, edge contours, and shell texture. The results showed that 20-class genotype × irrigation discrimination is possible with high accuracy using RGB images, while Grad-CAM provided qualitative support that the models generally focused on seed-related regions rather than background areas. EfficientNet-B1 is recommended when the highest accuracy with a small model footprint is required, whereas ResNet-18 provides the lowest latency on the tested workstation and may be a suitable candidate for future edge-device benchmarking. These capabilities point to potential future directions, including rapid pre-screening prototypes, human-in-the-loop active learning, multi-location validation, on-device inference tests, and RGB+NIR fusion, which remain to be demonstrated. Highlights <UnorderedList Mark="Bullet"> <ItemContent> 20 classes (5 genotypes × 4 irrigation) were successfully discriminated using RGB images. </ItemContent> <ItemContent> DenseNet-121 and EfficientNet-B1 achieved the highest performance results. </ItemContent> <ItemContent> The most confusion occurred between medium- to high-stress levels. </ItemContent> <ItemContent> Grad-CAM showed focus on the hilum, edge, and cortex tissue. </ItemContent> <ItemContent> Edge deployment is possible; lightweight models have reduced latency. </ItemContent> </UnorderedList>

Comparative Analysis of CNN Architectures with Optuna-Assisted Hyperparameter Optimization for Classifying Drought-Stressed Soybean Seeds

期刊论文
发表日期：

2026年7月29日

查看全文

Rustem Ustun,
Banu Ulu,
Muhammad Amjid,
Burak Ulu,
Bulent Uzun,
Necati Çetin

摘要

Abstract

The reduction in soybean yield and quality due to drought makes rapid and non-destructive grading at the seed level critical. In this study, five genotypes × four irrigation treatments (20 classes) were discriminated using RGB images in a scalable DL pipeline. ResNet-18, ConvNeXt-Tiny/Base, DenseNet-121, and EfficientNet-B1 were trained using transfer learning. HPO was performed with Optuna, and the accuracy, F1 score, ROC-AUC, and resource consumption were reported. On a held-out test set that was disjoint from model selection, EfficientNet-B1 achieved the best accuracy of 94.8% (AUC = 0.9990), followed by DenseNet-121 96.2% (AUC = 0.9993), ResNet-18 92.9% (AUC = 0.9985), ConvNeXt-Base 92.1% (AUC = 0.9981) and ConvNeXt-Tiny 94.0% (AUC = 0.9985). In terms of efficiency, ResNet-18 was balanced at 2232 s and 46.9 MB, with the lowest inference latency (15.5 ms). Grad-CAM foci pointed to the hilum, edge contours, and shell texture. The results showed that 20-class genotype × irrigation discrimination is possible with high accuracy using RGB images, while Grad-CAM provided qualitative support that the models generally focused on seed-related regions rather than background areas. EfficientNet-B1 is recommended when the highest accuracy with a small model footprint is required, whereas ResNet-18 provides the lowest latency on the tested workstation and may be a suitable candidate for future edge-device benchmarking. These capabilities point to potential future directions, including rapid pre-screening prototypes, human-in-the-loop active learning, multi-location validation, on-device inference tests, and RGB+NIR fusion, which remain to be demonstrated.

Highlights

20 classes (5 genotypes × 4 irrigation) were successfully discriminated using RGB images.

DenseNet-121 and EfficientNet-B1 achieved the highest performance results.

The most confusion occurred between medium- to high-stress levels.

Grad-CAM showed focus on the hilum, edge, and cortex tissue.

Edge deployment is possible; lightweight models have reduced latency.