Abstract:

Aboveground nitrogen concentration is an important indicator for accurately diagnosing crop nitrogen abundance and deficiency and thus evaluating its growth status. By constructing a hyperspectral-based nitrogen concentration estimation model for aboveground nitrogen concentration in drip-irrigated cotton, the real-time, non-destructive and accurate acquisition of nitrogen content in cotton can be realized, providing theoretical basis and technical support for precise fertilization. Six N application treatments (0, 120, 240, 348, 360, 480 kg/hm² of pure N) were set up with the main cotton cultivars ‘Xinluzao 45’ and ‘Xinluzao 53’ in northern Xinjiang as the test varieties. Measuring the hyperspectral information of the cotton canopy and removing redundancy using functional transformations, the first- and second-order derivative screening results were similar, and the logarithmic screening of the inverse was more scattered. Machine learning weight ordering was used for feature screening, and a total of seven feature bands, including 359, 371, 751, 752, 746, 739, and 755 nm, were selected. We also traversed the wavelength combinations to optimize the vegetation indices associated with high nitrogen in previous studies, and selected seven spectral indices including RVI′_810,460, NDVI′_811,856, NDVI′_750,705, RVI′_740,720, RVI′_851,852, DVI′_359,360, NDVI′_851,852. The filtered feature bands and vegetation indices were used to establish a nutrient estimation model with cotton nitrogen using ridge regression, decision tree, bootstrap clustering, and enhancement learning algorithms, respectively. Finally, the aboveground nitrogen concentration estimation model of drip-irrigated cotton established by the Adaboost iterative algorithm was the most effective, with a model accuracy of R² reaching 0.911 and an RMSE of 1.362. Spectral information can be used to effectively invert the nitrogen nutritional status of cotton, and the estimation accuracy of the model constructed based on the vegetation index is more stable than that of the feature band; the optimization of the existing feature band of the vegetation index can effectively improve the estimation accuracy of the model; comparing and analyzing the accuracy of the model under different modeling methods, the integrated learning is more advantageous than the single-machine learning in the nitrogen nutritional estimation of cotton.

Key words: cotton canopy, hyperspectral, aboveground nitrogen concentration, vegetation index, machine learning