Deep transfer learning of global spectra for local soil carbon monitoring

There is global interest in spectroscopy and the development of large and diverse soil spectral libraries (SSL) to model soil organic carbon (SOC) and monitor, report, and verify (MRV) 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) to leverage useful information from large-scale SSLs to assist local modelling. We used one global, three country-specific SSLs and data from three local sites with DTL to improve the modelling and localise the SOC estimates in individual fields or farms in each country. With DTL, we transferred instances from the SSLs, representations from one-dimensional convolutional neural networks (1D-CNNs) trained on the SSLs, and both instances and representations to improve local modelling. Transferring instances effectively used information from the global SSL to most accurately estimate SOC in each site, reducing the root mean square error (RMSE) by 25.8% on average compared with local modelling. Our results highlight the effectiveness of DTL and the value of diverse, global SSLs for accurate local SOC predictions. Applying DTL with a global SSL one could estimate SOC anywhere in the world more accurately, rapidly, and cost-effectively, enabling MRV protocols to monitor SOC changes.