Deep transfer learning of global spectra for local soil carbon monitoring

Authors: SHEN, ZEFANG - Curtin University; RAMIREZ-LOPEZ, LEONARDO - Buchi Laboratories; BEHRENS, THORSTEN - Bern University Of Applied Sciences; CUI, LEI - Curtin University; ZHANG, MINGXI - Curtin University; WALDEN, LEWIS - Curtin University; WETTERLIND, JOHANNA - Swedish University Of Agricultural Sciences; SHI, ZHOU - Zhejiang University; Sudduth, Kenneth - Ken; BAUMANN, PHILIPP - Bern University Of Applied Sciences; SONG, YONGZE - Curtin University; CATAMBAY, KEVIN - Curtin University; VISCARRA ROSSEL, RAPHAEL - Curtin University

Submitted to: Journal of Photogrammetry and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/14/2022
Publication Date: 4/26/2022
Citation: Shen, Z., Ramirez-Lopez, L., Behrens, T., Cui, L., Zhang, M., Walden, L., Wetterlind, J., Shi, Z., Sudduth, K.A., Baumann, P., Song, Y., Catambay, K., Viscarra Rossel, R.A. 2022. Deep transfer learning of global spectra for local soil carbon monitoring. Journal of Photogrammetry and Remote Sensing. 188:190-200. https://doi.org/10.1016/j.isprsjprs.2022.04.009.
DOI: https://doi.org/10.1016/j.isprsjprs.2022.04.009

Interpretive Summary: Interest in soil carbon sequestration as a way to mitigate climate change has led to the need to measure and monitor soil organic carbon (SOC) levels. One rapid and non-destructive approach to SOC measurement is visible and near infrared (vis-NIR) spectroscopy. Various modeling approaches have been used to relate vis-NIR soil spectral reflectance to SOC, 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 SOC. In this research, we applied machine learning methods called "deep transfer learning" (DTL) to a global soil spectral library and developed local SOC estimates for individual farms or fields in four countries. The DTL methods successfully extracted information that was most directly applicable for estimating SOC in the test fields and provided improved results compared to other methods. This new DTL approach will enhance the value of existing soil spectral libraries for estimating SOC at local scales as needed for accurate soil carbon monitoring.

Technical Abstract: There is global interest in spectroscopy and the development of There is global interest in spectroscopy and the development of large and diverse soil spectral libraries (SSL) to model soil organic carbon (SOC) and measure, monitor, report, and verify its changes. The reason is that increasing SOC can improve food production and mitigate climate change. However, 'global' modelling of SOC with such diverse and hyperdimensional SSLs do not generalise well locally, e.g. at a field scale. To address this challenge, we propose deep transfer learning (DTL). DTL can leverage large SSLs by transferring instances and representations from convolutional neural networks trained on the SSLs. We used one global and four country-specific SSLs with DTL to improve the modelling and localise the SOC estimates in individual fields or farms in each country. Transferring instances effectively used information from the global SSL to most accurately estimate SOC in each site. Our results highlight the effectiveness of DTL and the value of diverse, global SSLs for accurate local predictions.