基于深度学习结合高光谱技术的大豆种子活力检测方法

打开文本图片集

关键字：大豆；种子活力；检测；高光谱；深度学习；注意力机制中图分类号：S127 文献标识码：A 文章编号：1000-4440（2025）05-0927-10

Abstract：To achieve eficient，accurate，and non-destructive identification of soybean seed vigor，this study used seedsofthesoybeanvariety Williams 82asexperimentalmaterials.Alibraryof soybeanseeds withdiferentlevelsof vigor wasconstructedthroughartificialagingtreatments.HyperspectralimagesandRGBimagesof theseedswerethencollected to generate three image datasets （RGB dataset，SIQ dataset，and ENVIdataset）.Four deep learning models （Vg16Net， GoogLeNet，MobileV3Net，and ResNet-34）were employed to detectseed vigor，and theoptimal modelsand datasets were selected.Furthermore，thecoordinate atention（CA）mechanismandlabel smothing loss function were incorporated into theoptimal models toenhancetheirdetectionperformanceandrobustnessTheresultsdemonstratedthatusingtheSIQdata

set and ResNet-34 model，the recognition accuracy reached 97.6% and 96.8% on the training set and validation set，respectively.The detection performance was superiortoothercombinationsofmodelsand datasets.The CA-ResNet-34 model，which incorporated the CA mechanismand label smoothinglossfunction into the ResNet-34 model，achieved a detection accuracy of 98.5% for soy

bean seed vigor based onthe SIQdataset.Thisrepresentedanimprovementof1.7percentage points inaccuracycompared totheoriginal ResNet-34model.Theresultsofthis studycanprovideanew methodfortheaccurate，non-destructive，and efficient detection of soybean seed vigor.

Key words： soybean；seed vigor；detection；hyperspectral；deep learning；atention mechanism

大豆是中国重要的油料作物和经济作物，2023年中国大豆种植面积达 1.047×107hm2 ，产量约2.084×107 t。（剩余14180字）