Remote sensing factors are important environmental covariates in digital soil mapping, but existing studies often rely on single-source imagery or single-type factor, limiting the detailed representation of soil spatial distribution. To explore the potential of multi-source remote sensing imagery and multi-type factors for soil classification, this study constructed polarization features, vegetation indices, and texture features based on Sentinel-1 SAR and Sentinel-2A data, and formed 7 environmental variable combinations combined with topographic factors. Recursive feature elimination (RFE) was applied to select variables, and soil parent material information was fused. An XGBoost-based soil-landscape relationship model was established to realize the inference and mapping of soil types in the study area, and accuracy verification was conducted using field sampling points. The results showed that among single-factor mapping, vegetation indices exhibited the highest importance, although the mapping accuracy reached 65.25% when only texture features were used. When vegetation indices, polarization features, and texture features were applied for mapping, the overall mapping accuracy significantly increased from 58.16% to 66.67% compared with using only polarization features. Further analysis of individual soil types revealed that the multi-source factor combination substantially improved user’s and producer’s accuracies for sandy loam, forest loam, and silty clay paddy soils, demonstrating the advantages of multi-source fusion in complex terrain and transitional areas. This study confirmed the feasibility of combining multi-source remote sensing data with XGBoost for soil classification and provided a new technical pathway for accurately capturing soil spatial distribution and improving mapping precision.
MOOREI D, GESSLERP E, NIELSENG A, et al. Soil attribute prediction using terrain analysis[J]. Soil Science Society of America Journal, 1993, 57(2): 443-452.
[15]
ODEHAI O A, MCBRATNEYA B, CHITTLEBOROUGHD J. Spatial prediction of soil properties from landform attributes derived from a digital elevation model[J]. Geoderma, 1994, 63(3-4): 197-214.
WEREK, BUID T, DICKØ B, et al. A comparative assessment of support vector regression, artificial neural networks, and random forests for predicting and mapping soil organic carbon stocks across an Afromontane landscape[J]. Ecological Indicators, 2015, 52: 394-403.
[19]
ZERAATPISHEHM, AYOUBIS, JAFARIA, et al. Digital mapping of soil properties using multiple machine learning in a semi-arid region, central Iran[J]. Geoderma, 2019, 338: 445-452.
[20]
CHENS C, LIANGZ Z, WEBSTERR, et al. A high-resolution map of soil pH in China made by hybrid modelling of sparse soil data and environmental covariates and its implications for pollution[J]. Science of the Total Environment, 2019, 655: 273-283.
FREEMANE A, MOISENG G, COULSTONJ W, et al. Random forests and stochastic gradient boosting for predicting tree canopy cover: Comparing tuning processes and model performance[J]. Canadian Journal of Forest Research, 2016, 46(3): 323-339.
[25]
MAXWELLA E, WARNERT A, FANGF. Implementation of machine-learning classification in remote sensing: An applied review[J]. International Journal of Remote Sensing, 2018, 39(9): 2784-2817.
CHENT, GUESTRINC. Xgboost: A scalable tree boosting system[C]//Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 2016: 785-794
[30]
HUANGD M, XUS J, SUNJ Q, et al. Accuracy assessment model for classification result of remote sensing image based on spatial sampling[J]. Journal of Applied Remote Sensing, 2017, 11(4): 046023.
[31]
PONTIUSR G Jr, MILLONESM. Death to Kappa: Birth of quantity disagreement and allocation disagreement for accuracy assessment[J]. International Journal of Remote Sensing, 2011, 32(15): 4407-4429.
[32]
WARRENSM J. Properties of the quantity disagreement and the allocation disagreement[J]. International Journal of Remote Sensing, 2015, 36(5): 1439-1446.