Identification of Soybean Varieties Using Hyperspectral Imaging Coupled with Convolutional Neural Network.

Soybean variety is connected to stress resistance ability, as well as nutritional and commercial value. Near-infrared hyperspectral imaging was applied to classify three varieties of soybeans (Zhonghuang37, Zhonghuang41, and Zhonghuang55). Pixel-wise spectra were extracted and preprocessed, and average spectra were also obtained. Convolutional neural networks (CNN) using the average spectra and pixel-wise spectra of different numbers of soybeans were built. Pixel-wise CNN models obtained good performance predicting pixel-wise spectra and average spectra. With the increase of soybean numbers, performances were improved, with the classification accuracy of each variety over 90%. Traditionally, the number of samples used for modeling is large. It is time-consuming and requires labor to obtain hyperspectral data from large batches of samples. To explore the possibility of achieving decent identification results with few samples, a majority vote was also applied to the pixel-wise CNN models to identify a single soybean variety. Prediction maps were obtained to present the classification results intuitively. Models using pixel-wise spectra of 60 soybeans showed equivalent performance to those using the average spectra of 810 soybeans, illustrating the possibility of discriminating soybean varieties using few samples by acquiring pixel-wise spectra.

Machine-learning-facilitated prediction of heavy metal contamination in distiller's dried grains with solubles.

Excessive heavy metal contamination often occurs in feed due to natural or anthropogenic activity, leading to poisoning and other health problems in animals. In this study, a visible/near-infrared hyperspectral imaging system (Vis/NIR HIS) was used to reveal the different characteristics of spectral reflectance of Distillers Dried Grains with Solubles (DDGS) doped with various heavy metals and to effectively predict metal concentrations. Two types of sample treatment were used, namely tablet and bulk. Three quantitative analysis models were constructed based on the full wavelength, and through comparison the support vector regression (SVR) model was found to show the best performance. As typical heavy metal contaminants, copper (Cu) and zinc (Zn) were used for modeling and prediction. The prediction set accuracy of the tablet samples doped with Cu and Zn was 94.9% and 86.2%, respectively. In addition, a novel characteristic wavelength selection model based on SVR (SVR-CWS) was proposed to filter characteristic wavelengths, which improved the detection performance. The regression accuracy of the SVR model on the prediction set of tableted samples with different Cu and Zn concentrations was 94.7% and 85.9%, respectively. The accuracy of bulk samples with different Cu and Zn concentrations was 81.3% and 80.3%, respectively, which indicated that the detection method can reduce the pretreatment steps and verify its practicability. The overall results suggested the potential of Vis/NIR-HIS in the detection of feed safety and quality.

Hyperspectral imaging with shallow convolutional neural networks (SCNN) predicts the early herbicide stress in wheat cultivars.

The toxicity impacts of herbicides on crop, animals, and human are big problems global wide. The rapid and non-invasive ways for assessing herbicide-responsible effects on crop growth regarding types and levels still remain unexplored. In this study, visible/near infrared hyperspectral imaging (Vis/NIR HSI) coupled with SCNN was used to reveal the different characteristics in the spectral reflectance of 2 varieties of wheat seedling leaves that were subjected to 4 stress levels of 3 herbicide types during 4 stress durations and make early herbicide stress prediction. The first-order derivative results showed the spectral reflectance exhibited obvious differences at 518-531 nm, 637-675 nm and the red-edge. A SCNN model with attention mechanism (SCNN-ATT) was proposed for herbicide type and level classification of different stress durations. Further, a SCNN-based feature selection model (SCNN-FS) was proposed to screen out the characteristic wavelengths. The proposed methods achieved 96% accuracy of herbicide type classification and around 80% accuracy of stress level classification for both wheat varieties after 48 h. Overall, this study illustrated the potential of using Vis/NIR HSI to rapidly distinguish different herbicide types and serial levels in wheat at an early stage, which held great value for developing on-line herbicide stress recognizing methods in the field.