RESUMEN
Introduction: The use of chemical fertilizers in rice field management directly affects rice yield. Traditional rice cultivation often relies on the experience of farmers to develop fertilization plans, which cannot be adjusted according to the fertilizer requirements of rice. At present, agricultural drones are widely used for early monitoring of rice, but due to their lack of rationality, they cannot directly guide fertilization. How to accurately apply nitrogen fertilizer during the tillering stage to stabilize rice yield is an urgent problem to be solved in the current large-scale rice production process. Methods: WOFOST is a highly mechanistic crop growth model that can effectively simulate the effects of fertilization on rice growth and development. However, due to its lack of spatial heterogeneity, its ability to simulate crop growth at the field level is weak. This study is based on UAV remote sensing to obtain hyperspectral data of rice canopy and assimilation with the WOFOST crop growth model, to study the decision-making method of nitrogen fertilizer application during the rice tillering stage. Extracting hyperspectral features of rice canopy using Continuous Projection Algorithm and constructing a hyperspectral inversion model for rice biomass based on Extreme Learning Machine. By using two data assimilation methods, Ensemble Kalman Filter and Four-Dimensional Variational, the inverted biomass of the rice biomass hyperspectral inversion model and the localized WOFOST crop growth model were assimilated, and the simulation results of the WOFOST model were corrected. With the average yield as the goal, use the WOFOST model to formulate fertilization decisions and create a fertilization prescription map to achieve precise fertilization during the tillering stage of rice. Results: The research results indicate that the training set R2 and RMSE of the rice biomass hyperspectral inversion model are 0.953 and 0.076, respectively, while the testing set R2 and RMSE are 0.914 and 0.110, respectively. When obtaining the same yield, the fertilization strategy based on the ENKF assimilation method applied less fertilizer, reducing 5.9% compared to the standard fertilization scheme. Discussion: This study enhances the rationality of unmanned aerial vehicle remote sensing machines through data assimilation, providing a new theoretical basis for the decision-making of rice fertilization.
RESUMEN
Background and objective: The rapid diagnosis of rice nitrogen nutrition is of great significance to rice field management and precision fertilization. The nitrogen nutrition index (NNI) based on the standard nitrogen concentration curve is a common parameter for the quantitative diagnosis of rice nitrogen nutrition. However, the current NNI estimation methods based on hyperspectral techniques mainly focus on finding a better estimation model while ignoring the relationship between the critical nitrogen concentration curve and rice hyperspectral reflectance. Methods: This study obtained canopy spectral data using unmanned aerial vehicle (UAV) hyperspectral remote sensing and determined the rice critical nitrogen concentration curve and NNI. Taking the spectrum at critical nitrogen concentration as the standard spectrum, the original spectral reflectance and logarithmic spectral reflectance data were transformed by the difference method, and the features of the spectral data were extracted by a Autoencoder. Finally, the NNI inversion models of rice based on Extreme Learning Machine (ELM) and Bald Eagle Search-Extreme Learning Machine (BES-ELM) were constructed by taking the feature bands of four spectral extractions as input variables. Results: 1) from the feature extraction results of the self-encoder, simple logarithmic or difference transformation had little effect on NNI estimation, and logarithmic difference transformation effectively improved the NNI estimation results; 2) the estimation model based on the logarithmic difference spectrum and BES-ELM had the highest estimation accuracy, and the coefficient of determination (R2) values of the training set and verification set were 0.839 and 0.837, and the root mean square error (RMSE) values were 0.075 and 0.073, respectively; 3) according to the NNI, the samples were divided into a nitrogen-rich group (NNI ≥ 1) and nitrogen-deficient group (NNI < 1). Conclusion: The logarithmic difference transformation of the spectrum can effectively improve the estimation accuracy of the NNI estimation model, providing a new approach for improving NNI estimation methods based on hyperspectral technology.
