Abstract:

Objective: This paper proposed a prediction method of total iron content (TIC) in Dalbergia odorifera leaves based on vegetation index and multi-spectral texture features, in order to provide an alternative approach for the diagnosis of heavy metal nutrition of precious tree species. Method: In this paper, D. odorifera saplings were subjected to four levels (CK, F1, F2, F3) of iron treatment. At the end of treatment, the leaves were collected, and multi-spectral images were obtained, from which vegetation indexes and texture parameters, such as texture parameters mean value (TFMV) and texture parameters variance (TFV), were extracted and calculated, and the relationship between the variables and TIC was analyzed. The variables significantly correlated with TFC at 0.05 and 0.01 levels were screened out by significance test. Then, correlation analysis (CA), principal component analysis (PCA), average impact value (MIV) and genetic algorithm (GA) were used for secondary screening. The results were used as input variables of particle swarm optimization-back feedback neural network (PSO-BPNN) to analyze and compare the prediction values. Results: 1) In the CK ~F2 gradient range, the growth amount of tree height, crown width and stem increased with the increase of iron application, while under gradient F3, the growth amount of tree height and crown width decreased and growth of stem increased significantly. 2) With the increase of leaf TIC, band B first decreased and then increased; band G was contrary to band B; band R first decreased and then increased, and then remained stable; bands RE and NIR showed a trend of rise all the time; With the increase of TIC, different changes occurred in different bands. In band B, the spectral reflectance in level CK, F1 and F2 decreased but that in level F3 increased. In band G, the changes were contrary to band B. In band R, the spectral reflectance in level CK and F1 decreased, while it in level F2 increased, and maintained stability in level F3. bands RE and NIR showed an upward trend in all levels. 3) Most VI and TIC were correlated at 0.01 and 0.05 levels. TFMV and TFV were also able to reflect leaf TIC, but TFV was superior to TFMV in terms of correlation level and number. From viewpoint of band, the correlation between bands RE and NIR and TIC was better than other bands in vegetation index and texture features. 4) The results obtained by different screening methods were different, among which CA and GA had the best and similar evaluation indicators. But the prediction values of GA in the range of 150-300 mg/kg were lower than the measured values, which is not suitable for field fertilization guidance. (5) The prediction results obtained by vegetation index alone were inaccurate, and the goodness of fit and prediction accuracy were improved by adding texture parameters. By comparison, the variance of texture parameters had a greater impact on the accuracy of the model. It was shown that the degree of dispersion of leaf texture could be used as a good auxiliary information for predicting TIC. Conclusion: Gradient fertilization of F1 and F2 can promote the growth and biomass accumulation of Dalbergia odorifera. According to the experimental data, the optimum total iron content in leaves is 150-300 mg/kg. In addition to vegetation index, MPV as an auxiliary factor can improve the goodness of fit and prediction accuracy of the model. The CA-PSO-BPNN method can be effectively applied to field fertilization guidance and provide more accurate prediction for monitoring the heavy metal content of precious tree species.

Key words: Dalbergia odorifera, multi-spectral image, vegetation index, texture feature, BPNN