Enhancing canopy nitrogen estimation in Torreya Grandis based on advanced SLIC-EVI and HMT-seCNN methods using hyperspectral UAV data

Abstract

As UAV-based hyperspectral remote sensing technology becomes increasingly prevalent in agriculture and forestry, the estimation of plant nutrient content through hyperspectral data has become crucial for enhancing the efficiency of precision agricultural management. Hyperspectral imaging technology, capable of capturing the spectral characteristics of plants, shows significant potential for estimating nitrogen content. However, the preprocessing of hyperspectral images remains a challenge in practical applications, particularly with canopy images of Torreya Grandis, where mixed pixels lead to inaccuracies in extracting canopy pixel reflectance. To address this, we developed a novel Simple Linear Iterative Clustering- Enhanced Vegetation Index (SLIC-EVI) method specifically tailored for Torreya Grandis. Furthermore, we proposed a new approach that combines Hyperspectral Multiscale Transformation (HMT) with a squeeze-excitation convolutional neural network (seCNN). The HMT-seCNN model transforms one-dimensional spectral data into a three-dimensional map, optimizing the capabilities of CNN and the learning potential of the SE module, thereby significantly enhancing prediction performance. Experimental results demonstrate that the correlation between canopy pixel reflectance, as extracted by the SLIC-EVI method, and nitrogen content reached 0.744, substantially higher than the correlations achieved with NDVI (0.303) and EVI (0.551). This indicates that the SLIC-EVI method provides superior precision. In estimating nitrogen content, the HMT-seCNN model showed enhanced accuracy and generalization capabilities compared to traditional models that rely on local spectral features. In the test set, the HMT-seCNN model achieved a coefficient of determination (R2) of 0.765, a root mean square error (RMSE) of 2.281, and a relative prediction deviation (RPD) of 2.06, outperforming other methods. These findings underscore the benefits of integrating global spectral features with deep learning technology to enhance the accuracy of nitrogen content estimation. The main conclusions of this study include: (1) The SLIC-EVI method offers significant advantages over traditional methods in extracting reflectance; (2) The combination of global spectral features with the deep learning HMT-seCNN technique effectively enhances the predictive ability for Torreya Grandis nitrogen content; (3) The SE channel attention mechanism plays a crucial role in enhancing model performance.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.