Estimating soil organic carbon of sown biodiverse permanent pastures in Portugal using near infrared spectral data and artificial neural networks

• Published in: Geoderma Vol. 404; p. 115387
• Main Authors: Morais, Tiago G., Tufik, Camila, Rato, Ana E., Rodrigues, Nuno R., Gama, Ivo, Jongen, Marjan, Serrano, João, Fangueiro, David, Domingos, Tiago, Teixeira, Ricardo F.M.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.12.2021
• Subjects: Grassland; Machine learning; Sentinel-2; Soil organic matter; Spectroscopy; Grassland; Soil organic matter; Spectroscopy; Machine learning; Sentinel-2
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/j.geoderma.2021.115387

•Soil organic matter is an important indicator to assess soil quality.•SOM content was estimated for Portuguese sown biodiverse pastures.•A combined approach of spectral data with artificial neural networks was used.•Cross-validation was performed with an 8-fold leave-one-out approach.•ANN model was able to estimate SOM contents with low error. Grasslands in Portugal are key managed ecosystems, supporting and providing a diverse number of ecosystem services. Here, we developed a procedure for rapid estimation of soil organic carbon (SOC) in soil samples of sown biodiverse permanent pastures rich in legumes (SBP) in Portugal. We combined laboratory NIR spectral data analysis with artificial neural networks (ANN) to estimate the SOC content of SBP soil samples. To train and test the ANN, we used more than 340 soil samples collected in the 0–20 cm topsoil layer from three farms in 2018 and 2019 and two other farms in 2019 only. The number of bands of the spectra (800–2778 nm) was reduced using two different approaches: (a) aggregation to Sentinel-2 (S2) bands using the average reflectance within each bandwidths; and (b) principal component analysis (PCA). For the S2 approach, we considered the six S2 bands that overlap with the spectral range of the instrument used. For the PCA approach, we considered the five first principal components. Additional covariates were used for prediction, including weather and terrain attributes, e.g. accumulated precipitation, average temperature, elevation, and slope. To test for transferability of the models to different farms, we used an eight-fold leave-one-out cross-validation approach to calculate estimation errors. Each fold is a unique combination of farm and year and is used to assess the model's performance calibrated from the seven other folds. The ANN was able to estimate both low and high SOC contents without systematic errors and with similar estimation errors for both full and reduced spectral data approaches. The average root mean squared error (RMSE) for the S2 approach was 1.95 g kg−1 (0.45 – 2.33 g kg−1 depending on the hold-out fold) and for the PCA approach was 1.81 g kg−1 (0.74 – 2.42 g kg−1) (compared to the average SOC content of 12 g kg−1). These RMSE values were similar to the RMSE obtained using the full spectra, suggesting that the original spectral resolution could be reduced without losing information. These results suggest the potential for using remotely sensed data to estimate the variation of SOC content for SBP. They are a first step towards developing algorithms that can alleviate the cost and time of soil sampling and chemical SOC laboratory analysis through indirect estimation.