Hyperspectral remote sensing technology has become increasingly crucial in agricultural and environmental monitoring. Targeting the high-dimensional spatial characteristics of hyperspectral remote sensing data, we propose an advanced machine learning method to extract and validate the characteristic bands associated with soil nutrient contents. By applying this method, characteristic bands of various soil nutrients can be identified, and their contents efficiently assessed. This research provides a solid theoretical foundation for large-scale, rapid soil nutrient monitoring using drone-based hyperspectral remote sensing. It presents a reliable solution for the development and application of hyperspectral remote sensing technology, offering significant value for agricultural production and environmental protection.

Spectral experiments are conducted using drones equipped with hyperspectral sensors, collecting soil reflectance data across 176 spectral bands within the range of 398?1003 nm. The extraction of soil nutrient characteristic bands involves two main steps. The importance of the 176 spectral bands is ranked using a combination of random forest (RF) and differential evolution (DE) algorithms. The random forest method evaluates the importance of each spectral band, while the differential evolution algorithm refines the selection of spectral features, ensuring that the most informative bands are retained. This process results in a subset of spectral features indicative of soil nutrient content. The analytic hierarchy process (AHP) is employed to determine the relative importance of the spectral features in the subset. By ranking and applying weight thresholds, characteristic bands of different soil nutrients are successfully identified from the hyperspectral data. Finally, a quantitative inversion model for soil nutrient content is developed using a back-propagation neural network (BPNN).

Using available potassium as an example, the extracted characteristic bands are identified as 469.0, 501.6, 581.2, 697.2, 791.8, 795.4, 802.5, and 954.8 nm. Among these, the three most significant bands, 469.0, 501.6, and 954.8 nm, show the highest correlation with soil potassium content, making them critical for accurate nutrient assessment. The back-propagation neural network model trained with these characteristic bands achieves remarkable results. In the Training set, the coefficient of determination (R²) is 0.954 (Fig. 8), the ratio of performance to deviation (RPD) is 4.78, and the root mean square error (RMSE) is 14.32 mg/kg (Fig. 11). In the validation set, the model achieves R² of 0.848, RPD of 2.21, and RMSE of 16.71 mg/kg (Fig. 11). These results significantly outperform those obtained using the traditional first-order derivative mathematical transformation method, which yields R² values of 0.729 and 0.521, RPDs of 1.81 and 1.13, and RMSEs of 36.02 mg/kg and 191.05 mg/kg in the modeling and validation sets, respectively (Table 2).

The findings of this paper demonstrate the effectiveness and feasibility of machine learning methods for extracting hyperspectral soil nutrient characteristic bands. By integrating advanced algorithms such as random forest, differential evolution, and analytic hierarchy process, we offer a robust solution for the application of hyperspectral remote sensing technology. It enables more accurate and efficient soil nutrient assessments, significantly reducing the time and cost associated with traditional soil sampling and analysis. The successful extraction of characteristic bands and the development of a reliable predictive model underscore the potential of drone-based hyperspectral remote sensing technology for large-scale, rapid soil nutrient monitoring. This approach not only improves the precision of soil nutrient assessments, but also supports informed decision-making in agricultural production and environmental management. The outcomes contribute to enhanced crop yields, better resource allocation, and more sustainable agricultural practices.