Predicting regional spatial distribution of soil texture in floodplains using remote sensing data: A case of southeastern Iran

• Published in: Catena (Giessen) 2019-11, Vol.182, p.104149, Article 104149
• Main Authors: Shahriari, Mohammad, Delbari, Masoomeh, Afrasiab, Peyman, Pahlavan-Rad, Mohammad Reza
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Landsat 8; Random forest; Regression kriging; Soil texture; Spatial estimation
• ISSN: 0341-8162; 1872-6887
• DOI: 10.1016/j.catena.2019.104149

Soil texture is a physical property of soil and knowledge on its spatial distribution is essential for many agricultural and environmental activities especially in alluvial plains. This study aimed to predict the spatial distribution of soil fractions (i.e. percentages of sand, silt, and clay) in Sistan floodplain on a regional scale (area: 1500 km2). Random forest (RF), regression kriging-neural network residual kriging (RKNNRK), neural network residual kriging (NNRK), regression kriging (RK), and cokriging (COK) methods were used to map soil texture components over the region. Soil fractions were measured for 160 soil samples taken from the surface layer (0–30 cm) of various soil series in agriculture land of Sistan floodplain. The additive log-ratio (alr) transformation was applied to transform texture components prior to prediction. Remote sensing data including Landsat 8′ Bands (1–8), Band 4 to Band 8 ratio, Band 4 to Band 3 ratio, NDVI, GSI, Brightness Index, and Clay Index were used as auxiliary variables for interpolation of soil fractions. 80% of actual data was used for prediction and 20% of data was used for validation. The performance of methods used was evaluated using RMSE, ME and MAE criteria. The results showed that the RKNNRK model has the highest accuracy for prediction of sand (RMSE = 15.04%) and clay (RMSE = 8.77%) contents while the most accurate model for predicting silt content is NNRK (RMSE = 12.68%). RK and COK performed worse than NNRK and RKNNRK. A relatively high value of RMSE obtained for sand (17.89%), silt (14.15%), and clay (8.89%) contents by the RF model. This could be due to the low relief of study area, regional scale effects, low density of sampling points and high heterogeneity of soil texture components in floodplains. Our findings revealed that RKNNRK and NNRK models, when combined with remote sensing data, produce more accurate results and therefore can be used for appropriate mapping of soil fractions in floodplains and on a regional scale. •Hybrid geostatistical models properly estimated soil texture fractions over Sistan floodplain.•RMSE values were relatively high mainly due to heterogeneity of soil texture fractions across the region.•Cokriging and random forest were not able to accurately estimate soil fractions over the study region.