Better inversion of rice nitrogen nutrition index at early panicle initiation stage using spectral features, texture features, and wavelet features based on UAV multispectral imagery

The early panicle initiation stage plays a pivotal role in rice yield formation and nitrogen use efficiency. Rapid and accurate estimation of the Nitrogen Nutrition Index (NNI) during this stage is essential for guiding precise fertilization in high-yield rice cultivation. Although discrete wavelet transform (DWT) serves as an effective feature extraction tool, its application to crop NNI estimation remains unexplored. In this study, three-year field experiments involving ten rice varieties and five nitrogen application levels were conducted in Jiangsu Province, China. NNI data at the early panicle initiation stage and multispectral Unmanned Aerial Vehicle (UAV) imagery were collected. The sets of vegetation indices (VIs), texture indices (TIs), and DWT feature variables were extracted and fused from the imagery. Three feature selection methods were each combined with four machine learning algorithms to build distinct NNI estimation models, followed by an assessment of model accuracy. The results indicated that the overall estimation accuracy of models developed from different feature sets followed this order: VIs+TIs+DWT > VIs+DWT > VIs+TIs > VIs > TIs+DWT. RFECV-RF models constructed with the VIs+TIs+DWT and VIs+DWT feature sets both exhibited significantly higher estimation accuracy than the two existing methods using VIs and VIs+TIs, reaching a level suitable for precise quantitative analysis. The ratio of performance to deviation (RPD) of the model built with the VIs+TIs+DWT feature set was significantly higher than that of the model using the VIs+DWT feature set. Integrating DWT with VIs and TIs has significantly enhanced the accuracy of remotely sensed NNI estimation during the early panicle initiation stage, providing a method for precise nitrogen status diagnosis in rice at this critical growth phase.