Rapid and Non-Destructive Detection of Soybean Seed Viability Using Temporally Enhanced Speckle Features Coupled with Lightweight Attention Network

Soybean seed viability is critical for field emergence, stand uniformity, and yield stability, whereas conventional assessment methods are often destructive, time-consuming, and labor-intensive. This study developed a rapid and non-destructive soybean seed viability identification framework centered on a temporally enhanced GraphAVD representation, together with a lightweight channel-aware classifier for downstream discrimination. A class- and cultivar-balanced dataset of 3000 soybean seeds with unambiguous binary labels was constructed, with each of the five cultivars contributing 300 germinated untreated viable seeds and 300 non-germinated heat-treated non-viable seeds. The dataset was divided into a development set (n = 2500) and a fully isolated independent test set (n = 500). Dynamic speckle sequences were converted into feature maps using a temporally enhanced GraphAVD method (TD-GraphAVD), which incorporates multi-step temporal differencing, exponentially decaying weighting, and overlapping sliding-window aggregation to more systematically exploit temporal dynamic information in laser speckle sequences. Under the same ResNet-50 framework, TD-GraphAVD achieved the best feature-extraction performance among the compared methods, with a best validation Accuracy of 84.80% and a best validation Macro-F1 of 84.80%. Under TD-GraphAVD feature input, EfficientNet-B0 enhanced with an Efficient Channel Attention module achieved the strongest downstream classification performance, with a best validation Accuracy and Macro-F1 of 95.87% on the fixed internal validation split. Across the 10-fold cross-validation procedure on the development set, the model achieved mean validation Accuracy and Macro-F1 of 94.85% and 94.85%, respectively. On the independent test set, the mean Accuracy and Macro-F1 of the fold-trained sub-models remained 92.26% and 92.25%, respectively. These results indicate that the proposed framework is effective for rapid and non-destructive soybean seed viability discrimination under the present heat-inactivation-based laboratory-controlled binary setting, while its applicability to naturally aged, partially deteriorated, or otherwise heterogeneous non-viable seed populations remains to be further verified.