Prediction Of Total Nitrogen And Total Iron Content In Young Sandalwood Based On Visible-Near Infrared Images

Sandalwood is a kind of typical precious tree species with high economic and medicinal value which originated in Australia,Indonesia and other countries.In resent years,sandalwood has been widely planted in Hainan Province of China.However,due to extensive management mode and the lack of real-time nutrition diagnosis technology,the survival rate and growth rate of sandalwood are very low.In this study,a method for predicting total nitrogen and total iron content in young sandalwood based on visible-near infrared images was proposed.The effects coursed by nitrogen and iron stress on growth of stem,tree height and crown width of young sandalwood during the experiment were analyzed in this paper.Prediction models of total nitrogen and total iron content on two plant and leaves scales were discussed in detail.The images of sandalwood plants under natural conditions were obtained by ordinary SLR camera,and the images of leaves in vitro were obtained by multispectral camera.At the plant-scale,a method of image segmentation based on visible light is proposed.We transformed the original image into L~*a~*b~*color model and using Otsu method and morphological operation to extract the foreground.On the basis of this method,a new method of judging old and new leaves is proposed,which combines canopy width data.The main purpose is to highlight the characteristics of the regional effects of iron stress on leaves.Finally,in the construction of plant-scale models of total nitrogen and total iron,the effects of different BP neural network optimization methods on the prediction results were discussed.At the same time,the effects of color values of different regions as independent variables on the prediction results were compared.At the leaf-scale,the effects of different stress level of nitrogen and iron on leaf multispectal reflectance were analyzed.Then,vegetation indexes,mean texture parameters and variance features of texture parameters were extracted.Pearson correlation coefficients between nutrient content and those features were used to compare the indicative degree of nitrogen and iron stress.Finally,BP neural network optimized by genetic algorithm was selected as a prediction model,then the influence of different variables selection methods on the results of the prediction models is discussed.The main conclusions are as follows:(1)Nitrogen and iron stress has certain but different effects on the growth of young sandalwood.In the nitrogen stress test,the tree height and crown width grow rapidly while the growth of the ground stem is relatively slow with the increase of nitrogen application rate.The low nitrogen application level has the greatest effects on height,but the crown width and stem growth under high nitrogen application level are more obvious.In the iron stress test,the growth of tree height,stem and crown increased continuously under the low and medium iron application level.But when iron fertilizer was excessive,the growth of tree height and crown slowed down while the growth of stem continued to increase.(2)L~*a~*b~*color model get better results in sandalwood images segmentation in complex background.By comparing the original RGB,HSI and L~*a~*b~*color models,we found that the segmentation effect of each channel under L~*a~*b~*color system is better than other two models.The advantages of L~*channel and b~*channel are complementary after using Otsu method.By combining it with morphological operations,the background can be effectively removed.(3)When predicting total iron content at plant-scale,the model fitting and validation accuracy obtained by color ratio of new leaves to old leaves as independent variables is the highest.While the results obtained by using the color features of new leaves or whole plants have the similar poor indication.Also on the plant-scale,the prediction accuracy of total nitrogen and total iron content obtained by optimizing BPNN prediction model with GA algorithm is the highest.By comparing the predicted results of different optimization algorithm,it was found that GA and PSO algorithm have similar optimization effect on BPNN,but PSO-BPNN predicted value is obviously larger than GA-BPNN,which has poor indication of nitrogen and iron deficiency.In other words,is easy to cause misjudgment when using PSO-BPNN prediction model which is not suitable for guiding the actual field fertilization work.(4)Vegetation index can be used to indicate the degree of nitrogen stress,and texture parameters can be used to indicate the degree of iron stress.In nitrogen stress,vegetation index calculated from RE and NIR bands had the greatest correlation with total nitrogen content.G and NIR bands also have the same indicating ability but with lower Pearson correlation coefficient.In iron stress,the mean value of texture parameters and variance of texture parameters are highly correlated with total iron content in leaves,but the variance of texture parameters is slightly better than the mean value of texture parameters.The optimal bands of both are B and R bands.The best parameters are variance and contrast.(5)On the leaf-scale,the prediction accuracy of total nitrogen and total iron is higher when the independent variables were selected by GA algorithm.CA,GA and MIV methods are used to select the variables as the input layer of BP neural network,and then used GA algorithm to optimize them.The results showed that although the evaluation indexes obtained by GA were not all the best,the results were closer to the measured values in the critical range of nutrient content.So GA screening method can reduce the probability of insufficient and excessive topdressing.