研究目的
To classify three varieties of soybeans (Zhonghuang37, Zhonghuang41, and Zhonghuang55) using near-infrared hyperspectral imaging and convolutional neural networks (CNN) to explore the possibility of achieving decent identification results with few samples.
研究成果
Hyperspectral imaging coupled with a CNN model showed great potential for soybean variety classification. Pixel-wise CNN models using a few samples achieved good results, illustrating the possibility of discriminating soybean varieties using few samples by acquiring pixel-wise spectra. The results would also help to identify other samples with a similar situation.
研究不足
The study focused on three specific varieties of soybeans, and the generalizability to other varieties or conditions was not explored. The computational time for pixel-wise CNN models was longer than that for object-wise CNN models.
1：Experimental Design and Method Selection:
Near-infrared hyperspectral imaging was applied to classify soybean varieties. Pixel-wise spectra were extracted and preprocessed, and average spectra were obtained. CNN models using the average spectra and pixel-wise spectra of different numbers of soybeans were built.
2：Sample Selection and Data Sources:
Three varieties of soybeans, including Zhonghuang37, Zhonghuang41, and Zhonghuang55, were purchased from a local seed company in Changzhou, Hebei province, China. For each variety, 1890 intact and healthy soybeans were prepared.
3：List of Experimental Equipment and Materials:
A near-infrared hyperspectral imaging system covering the spectral range of 874–1734 nm was used to acquire hyperspectral images of soybeans. Main components included an imaging spectrograph, camera, lens, light sources, and a mobile platform.
4：Experimental Procedures and Operational Workflow:
Single soybean kernels were placed separately in a black sampling plate for hyperspectral image acquisition. Hyperspectral images were corrected as the reflectance hyperspectral images. Spectral data were extracted and preprocessed.
5：Data Analysis Methods:
PCA was applied for qualitative analysis. Object-wise and pixel-wise CNN models were established for classification.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
