Skip to main content
ARS Home » Midwest Area » Columbia, Missouri » Cropping Systems and Water Quality Research » Research » Publications at this Location » Publication #406502
Research Project: Sustainable Intensification of Cropping Systems on Spatially Variable Landscapes and Soils

Location: Cropping Systems and Water Quality Research

Title: An imperative for soil spectroscopic modelling is to think global but fit local with transfer learning

Author
item VISCARRA ROSSEL, RAPHAEL - Curtin University
item SHEN, ZEFANG - Curtin University
item RAMIREZ-LOPEZ, LEONARDO - Buchi Laboratories
item BEHRENS, THORSTEN - Bern University Of Applied Sciences
item SHI, ZHOU - Zhejiang University
item WETTERLIND, JOHANNA - Swedish University Of Agricultural Sciences
item Sudduth, Kenneth - Ken
item STENBERG, BO - Swedish University Of Agricultural Sciences
item GUERRERO, CESAR - Miguel Hernandez University
item GHOLIZADEH, ASA - Czech University Of Life Sciences Prague
item BEN-DOR, EYAL - Tel Aviv University
item ST. LUCE, MERVIN - Agriculture And Agri-Food Canada
item ORELLANO, CLAUDIO - Buchi Laboratories

Submitted to: Earth-Science Reviews
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/27/2024
Publication Date: 5/8/2024
Citation: Viscarra Rossel, R.A., Shen, Z., Ramirez-Lopez, L., Behrens, T., Shi, Z., Wetterlind, J., Sudduth, K.A., Stenberg, B., Guerrero, C., Gholizadeh, A., Ben-Dor, E., St. Luce, M., Orellano, C. 2024. An imperative for soil spectroscopic modelling is to think global but fit local with transfer learning. Earth-Science Reviews. 254. Article 104797. https://doi.org/10.1016/j.earscirev.2024.104797
DOI: https://doi.org/10.1016/j.earscirev.2024.104797

Interpretive Summary: Various modeling approaches have been used to relate visible and near infrared (vis-NIR) soil spectral reflectance to soil properties, including models based on soil spectral libraries. These spectral libraries generally consist of data from samples collected over national, continental, or global scales. Because of scale differences, it is often difficult to obtain good local (farm or field) estimates of soil properties. In this research, we applied machine learning methods called "transfer learning" (TL) to a global soil spectral library and developed local soil property estimates for 12 individual farms or fields from 10 countries in the seven continents. The TL methods successfully extracted information that was most directly applicable for estimating soil properties in the test fields and provided improved results compared to other methods in 10 of the 12 cases. This new approach will enhance the value of existing soil spectral libraries for estimating soil properties at local scales.

Technical Abstract: Diffuse reflectance spectroscopy in the visible--near infrared (vis--NIR; 400 - 2,500 nm) range can be used to characterise various soil properties. Large soil spectral libraries (SSLs) and machine learning (ML) have been investigated to derive accurate soil properties predictions. However, how to effectively use the SSLs for accurate predictions at local scale, i.e., localisation or localising spectroscopic modelling, is still challenging due to the different characteristics in the SSLs and local soil, and therefore lacks understanding of what information could be transferred and how. We argued that localisation is a typical transfer learning (TL) problem and reviewed and classified existing localisation methods under the TL framework. We then used a promising method to transfer instances (samples) from a global SSL to 12 local sites from 10 countries in the seven continents and interpreted the transfer by inspecting the data, model, soil and environmental factors related to the soil samples. Instance-based TL with the RS-LOCAL 2.0 data-driven heuristic search method improved the local modelling for 10 out 12 local sites.