Hybrid Deep Learning Model for Open-Set Seed Classification Using Hyperspectral and RGB Imaging

The classification of seed varieties is important to maintain the productivity of agriculture and check fraudulence and the control of quality in seed businesses. The earlier approaches for the seed classification conventional techniques are tedious and rudimentary, besides being to an extent, arbitrary and though permissible they have low scalability, especially to big data or even as real-time solutions. This paper is a suggestion of using hybrid deep learning framework open to hyperspectral imaging with HSI and RGB Imaging. To this end, the newly proposed Dual Deep Support Vector Data Description model and a vision transformer-based architecture are used for our classification task of known seed types along with the rejection of unknown seeds. The framework uses incremental learning techniques, and spatial-spectral features as well as attention-based bands to enhance the scalability of the technique and its performance. Comparing the results of the present work with previous techniques, it can be concluded that the proposed methods of hybrid imaging and improved deep learning model may help to solve the problems of classic, as well as open-set recognition in agriculture. Future studies could apply this approach to other crops with the goal to pursue precision agriculture and improve world’s food safety.