Selection of spectral indices and comparison of estimation models for soil total nitrogen content in honey pomelo orchard

【Objectives】Nitrogen is an essential nutrient for the growth of honey pomelo trees, and it plays an irreplaceable role in regulating cellular metabolism, improving fruit quality and increasing yield. Currently, in the monitoring of soil nitrogen in honey pomelo orchards, traditional laboratory methods are commonly used to determine soil total nitrogen content. Although this method has high accuracy, its experimental steps are cumbersome and its data exhibits poor timeliness. Hyperspectral technology offers multiple advantages, including rapid analysis, no requirement for chemical reagents, simple operation, and a high degree of automation. It enables single-plant nutrition diagnosis, as well as the monitoring and analysis of a large number of samples, which makes precise variable fertilization feasible for large-scale bases. Therefore, rapid diagnosis of soil nitrogen content is critical for rational fertilization and improving the yield and quality of orchard produce. Developing a soil nitrogen spectral model for honey pomelo orchards can provide theoretical support for the rapid diagnosis and monitoring of soil nitrogen.【Methods】We extracted and analyzed the original and first-order derivative spectral characteristic bands, along with three types of spectral indices (DSI, RSI and NDSI), for soil sam-ples collected from the honey pomelo orchard. We constructed univariate estimation models, partial least squares estimation models (PLS), backpropagation neural network estimation models (BPNN), random forest estimation models (RF), and support vector machine estimation models (SVM) for soil total nitrogen content in honey pomelo orchards, then evaluated and validated the optimal spectral estimation model for soil total nitrogen content in the study area.【Results】There was a significant multi-band correlation between the original and first derivative soil spectra and nitrogen content in honey pomelo orchards. Based on the correlation coefficients of the original spectrum and first derivative spectrum, the maximum correlation wavelengths were 1441 nm, 1443 nm and 448 nm, 449 nm, respectively, and the corresponding correlation coefficients r were 0.50, 0.50, 0.52, and 0.52, respectively. Based on the original spectrum, the maximum negative correlation coefficients between the spectral index and soil total nitrogen content in the honey pomelo orchard were - 0.57 (DSI638,641), - 0.56 (RSI638,641) and - 0.56 (NDSI638,641), respectively (P＜0.001); and the maximum positive correlation coefficients were 0.49 (DSI2222,2267), 0.47 (RSI518,599), and 0.47 (NDSI518,599), respectively. Based on first-order derivative spectroscopy, the maximum negative correlation coefficients between the spectral index and soil total nitrogen content were -0.72 (DSI′1718,1736), -0.74 (RSI′2315,2414) and -0.74 (NDSI′2315,2414), respectively. The maximum positive correlation r values were 0.73 (DSI′797,1313), 0.71 (NDSI′671,2348), and 0.70 (RSI′2318,2411), respectively. The overall correlation between soil total nitrogen content and first- order derivative spectra was stronger than that between soil total nitrogen content and the original spectral index. The polynomial estimation model constructed with spectral indices including DSI′797,1313, RSI′2318,2411, RSI′2315,2414, NDSI′2315,2414, DSI′1718,1736, NDSI′671,2348, and DSI638,641 as independent variables yielded a high coefficient of determination R2 (R2 ＞0.40). The univariate model based on the spectral index only considered the influence of a single variable and yielded low correlation. Therefore, we selected variables with good correlation from the spectral parameters, and established and evaluated hyperspectral estimation models for soil total nitrogen content using the four methods mentioned above. Due to differences in their calculation approaches, models constructed via different regression methods produced varying prediction performances for soil total nitrogen content in honey pomelo orchards, and the resulting prediction models differ in R2 and RMSE values. The prediction model constructed by the RF method had the highest R2 and the lowest RMSE from the 95% confidence interval, it can be seen that the data concentration of the confidence interval of the RF method is greater than that of the SVM, BPNN, and PLS methods, and the measured and predicted values are more concentrated. Its modeling R2 , RMSE, and RE were 0.83, 0.10, and 8.83% , respectively, indicating the highest modeling accuracy. The BPNN method had modeling R2 , RMSE, and RE of 0.67, 0.12, and 10.46%, respectively, and it ranked second in modeling accuracy. The R2 , RMSE, and RE of SVM modeling were 0.64, 0.13, and 8.41%, respectively. The R2 , RMSE and RE of PLS modeling were 0.62%, 0.13% and 10.67%, respectively. Based on the model’s overall accuracy evaluation criteria, the R2 value ranged from 0.67 to 0.89, indicating that the model has sound prediction accuracy and predictive capability. It can be seen that the predictions of RF and BPNN were consistent. The validation R2 values for the PLS, BPNN, RF, and SVM methods were 0.57, 0.63, 0.80, and 0.70, respectively. The RMSE values were 0.10, 0.10, 0.09 and 0.09, respectively; the RE values were 7.00%, 7.39%, 6.58% and 4.15%, respectively. Compared with other models, the RF validation model had a higher R2 , lower RMSE, and lower RE, indicating that the RF method could more accurately estimate the soil total nitrogen content in honey pomelo orchards, followed by the SVM method. The accuracy of the validation models ranked as RF ＞ SVM ＞ BPNN ＞ PLS.【Conclusions】In the accuracy comparison of soil total nitrogen content prediction models for (honey pomelo orchards, the RF model achieves the best predictive performance. These results provide technical support for promoting the timely monitoring of total soil nitrogen content in honey pomelo orchards and guiding rational fertilization. However, the applicability of the model proposed in this article under different environmental conditions still requires further verification.