Scale transfer learning of hyperspectral prediction model of heavy metal content in maize: From laboratory to satellite

ABSTRACT

For sustainable development of the environment, there has been a growing interest in using remote sensing technology to monitor the content of heavy metals in crops. Satellite hyperspectral data, while capable of providing valuable information, is often subject to environmental factors. On the other hand, laboratory spectra are collected under controlled conditions, enabling more accurate and stable models to be built. As a result, this study used the laboratory maize spectra measured using an SVC spectrometer to establish the heavy metal content prediction model and applied it to hyperspectral satellite images (Gaofen-5 and Zhuhai-1) to achieve large-scale and rapid monitoring of heavy metal content in farmland maize. However, due to the difference between the laboratory spectra (source domain) and satellite spectra (target domain), the direct transfer of laboratory models to the satellite scale has shown poor results. To address this challenge, multiple factors, including spectral feature selection, domain adaptation, regression methods, and correction models were taken into account to construct the multi-factor combined model. The research has shown that modelling after spectral feature selection can increase the accuracy of the model while reducing its complexity. Additionally, the domain adaptation method can narrow the gap between the source domain and the target domain, improving the transfer learning performance of different domains to a certain extent. Furthermore, the addition of a constructed correction model generally improves the accuracy of the transfer model. The optimal transferable model is finally obtained, which can effectively apply the laboratory heavy metal content prediction model to satellite images, and achieve large-scale and rapid monitoring of heavy metal content in farmland maize.

KEYWORDS:

• Heavy metal content
• Scale transfer learning
• Domain adaptation
• Uninformative variable elimination
• Correction model

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

Research data are not shared.

Additional information

Funding

The authors thank the reviewers and the editor for providing valuable suggestions to improve the manuscript. This work has been financially supported by the Science & Technology Fundamental Resources Investigation Program (2022FY101905), the National Natural Science Foundation of China (41971401) and the Fundamental Research Funds for the Central Universities (2022JCCXDC01, 2022YJSDC22). The authors would like to thank Zhuhai Orbita Aerospace Science & Technology Co., Ltd for providing the Zhuhai-1 hyperspectral image