Harmonizing ground and UAV hyperspectral data: A novel spectral correction method for maximizing estimation models and datasets of ground hyperspectral

Abstract

The accurate and effective monitoring of rice nitrogen status using hyperspectral datasets and estimation models is important for precision agriculture and intelligent breeding. Ground-based non-imaging hyperspectral datasets have provided high-throughput solutions, but developed nitrogen estimation models through it and accumulated hyperspectral datasets are difficult to generalize for practical production applications. Hyperspectral imaging technique from unmanned aerial vehicles (UAV) offers opportunities for real-time and large-area monitoring of rice nitrogen status, but their datasets and developed estimation models are not as rich compared to non-imaging hyperspectral. However, we found that non-imaging hyperspectral datasets and estimation models are difficult to adapt to hyperspectral images. Thus, this study aims to harmonize non-imaging hyperspectral databases and nitrogen estimation models to accommodate UAV hyperspectral images. We proposed a spectral correction method for harmonizing non-imaging hyperspectral and hyperspectral image data. Estimation models of canopy nitrogen content (CNC) were developed using machine learning algorithms with non-imaging hyperspectral, hyperspectral images, and corrected hyperspectral datasets. Importantly, the applicability and effectiveness of the CNC estimation models and the dataset of corrected hyperspectral were explored for hyperspectral images. The results showed that the corrected hyperspectral provided a modeling accuracy comparable to that of the non-imaging hyperspectral and hyperspectral images when estimating the CNC. The applicability of the corrected hyperspectral estimation models is superior to non-imaging hyperspectral when applied to hyperspectral images. Similarly, the corrected hyperspectral dataset outperformed the non-imaging hyperspectral dataset in estimating nitrogen content from the estimation models of hyperspectral images. To conclude, corrected hyperspectral estimation models and datasets can be effectively transferred to hyperspectral images to estimate CNC and address the problem of heterogeneity between non-imaging and hyperspectral images. This study provides a new approach to maximize the use of estimation models and databases developed by non-imaging hyperspectral in precision agriculture.