Classification Of Maize Moldy Seeds Based On Hyperspectral Imaging And SSA-RF

It is particularly important to increase the corn yield because our country is a big producer and seller of corn seeds.But improper seed storage is easy to produce mildew.Traditional chemical detection of moldy seeds will damage the structure and take a long time.In this paper,the degree of corn seed mildew was studied.In view of the difficulty of seed mildew detection,hyperspectral imaging technology(HSI)and sparrow search algorithm were used to optimize random forest(SSA-RF)to study the classification methods of corn seeds with different degrees of mildew.The main research contents and conclusions of this paper are as follows:Firstly,in view of the small corn seeds,the workload of scanning hyperspectral corn seed images per grain is large,and the amount of data processed by software is large and the traversal process is complex,a fast processing method of hyperspectral image is proposed to establish the hyperspectral data sets of corn seeds at different stages of mildew.The hyperspectral images of corn seeds were scanned,and the images of the bands with obvious brightness were selected.According to the specified structural elements,the regions of interest were obtained in batches and the spectral information of corn seeds was extracted based on the open and close operation of Otsu method and morphology.Feature importance screening,successive projection algorithm and competitive adaptive weighted sampling algorithm were used to screen feature wavelengths.Secondly,considering the large number of bands in maize seed hyperspectral data,the sparrow search algorithm(SSA)can effectively improve the search efficiency,but it is easy to fall into local optimality in the process of sparrow search optimization parameters.In order to expand the search range of the optimal solution,the dynamic threshold of elite population is given in the search process,and random probability is introduced as a parameter.Furthermore,the elite reverse sparrow search algorithm model(JYSSA)was established based on the initial population restriction.In the search process of hyperspectral data set,the internal search mechanism is prone to overreach phenomenon,so the simplex operation is simplified,only reflection,expansion and external search are adopted,and the Levy flight mechanism is used to supplement the search.A starting point and target point are randomly selected in the search space to simulate the flight behavior of sparrow,so as to achieve an effective search in the search space.The Elite simplified simplex sparrow search algorithm model(JYMSSA)will be established.Compared with particle swarm optimization algorithm(PSO)and whale optimization algorithm(WOA),the results show that JYSSA and JYMSSA models have a better ability to search for the optimal solution.Finally,JYMSSA and JYSSA were used to optimize the selected feature importance based band data set and the all-band data set as input of random forest classifier.In the verification set,the random forest,SSA optimized random forest,JYSSA optimized random forest and JYMSSA optimized random forest algorithm were compared.After feature screening,the band accuracy and classification accuracy of JYSSA were 96%,superior to other models.The accuracy of JYMSSA in the all-band data set is 95%,which is better than other models.In conclusion,in this paper,by scanning healthy corn seeds and corn seeds with different degrees of mildew by nondestructive testing,JYSSA optimized stochastic forest model has the best classification effect on wavelength after feature screening,and JYMSSA optimized stochastic forest model has the best classification effect on full band.The results of this study can provide valuable reference for classification of moldy seeds in nondestructive testing.