Main content area

Transfer learning to localise a continental soil vis-NIR calibration model

Padarian, J., Minasny, B., McBratney, A.B.
Geoderma 2019 v.340 pp. 279-288
artificial intelligence, cation exchange capacity, clay fraction, data collection, databases, models, organic carbon, pH, spectral analysis, spectroscopy
The rapid development in NIR and information technologies saw the development of various initiatives that have generated large scale databases of soil spectroscopy globally. Models generated within a specific spectral or geographical domain should be carefully used in other contexts since they may lose their validity. This includes the application of a global, continental or national spectral libraries to local areas or regions. Both, global and local models are valuable and, ideally, we would like to transfer some of the rules learnt by the more general global models to a local domain. In machine learning, the process of sharing intra-domain information is known as transfer learning. This paper aims to describe and evaluate the effectiveness of transfer learning to “localise” a general soil spectral model. The transfer process consists in, first, training a model with a big volume of data covering a diverse group of cases. Second, some layers of the trained neural network are used to build a local model, which is fine-tuned by using a smaller amount of local data. We demonstrated this method using the LUCAS database, an European dataset, comprising spectral data from 21 countries. For each country, we generated three models: a) Global, with data from all except the country of interest; b) Local, with data from the country; and c) Transfer, pre-trained as the Global model and fine-tuned with data from the country. The results showed that the Transfer model can lower the error (expressed as RMSE) 91% of the cases, with a mean reduction of RMSE: 10.5, 11.8, 12.0 and 11.5% for organic carbon, cation exchange capacity, clay content and pH, respectively. This paper demonstrates the usefulness of transfer learning for soil spectroscopy, which will enhance the use of global spectral libraries for local application.