Abstract
Background: Varieties identification of crop seeds is significant for breeders to screen out seeds with specific traits and for market regulators to detect seeds purity. Hyperspectral imaging technology provides a fast and non-destructive means for varieties identification. And deep learning algorithm is suitable for effective analysis of redundant spectral data. However, deep learning algorithms have serious big data dependency, while collecting high-quality large-scale samples was high-cost in many cases. This made it difficult to build an accurate identification model. This study aimed to explore a rapid and accurate method for varieties identification of different crop seeds under sample-limited condition based on hyperspectral imaging and deep transfer learning.Results: Three deep neural networks with typical structures were designed based on a samples-rich Pea dataset. Obtained the highest accuracy of 99.57 %, VGG-MODEL was transferred to classify four target datasets (Rice, Oat, Wheat, Cotton) with limited samples. The accuracies of deep transferred model achieved 95 %, 99 %, 80.8 %, and 83.86 % on the four datasets, respectively. Using training sets with different sizes, deep transferred model could always obtain higher performance than other traditional methods. Visualization of training process and classification results confirmed the portability of common features of seed spectra and provided an interpreted method for rapid and accurate varieties identification of crop seeds.Conclusions: This study combined hyperspectral imaging and deep transfer learning to identify varieties of different crop seeds, which was proved to be efficient under sample-limited condition. This facilitated crop variety screening process under the scenario of sample scarcity. It also provided a new idea for the detection of other qualities of crop seeds based on hyperspectral imaging under sample-limited condition.
Study on Radar Echo-Filling in an Occlusion Area by a Deep Learning Algorithm
