Scholarlink

<p>Prediction of soil properties at different depth levels and scales requires the use of multimodal datasets to represent the characteristics of the SCORPAN factors within the digital soil mapping (DSM) technique. With spectral variables representing the dynamic nature of soil property estimation, evaluating their efficiency can enhance the applicability of the models for further standardization. The current study aimed to compare the utility of different spectral modalities (i.e., Sentinel 1A, Sentinel 2A, Landsat 8, and PRISMA) for digital soil mapping of soil infiltration rates (cm&#xa0;h<sup>−1</sup>) and textural classes in the Thiruparankundram block of Madurai district. Further, the utility of the spectral subsets, both independently and in combination with other SCORPAN variables, was studied to perceive the efficiency of the framework and to determine the integration potential of spatial variables from multiple modalities. A total of 224 soil observations were derived from the legacy soil maps using a stratified random sampling procedure. The extracted soil and the covariate information were then subjected to the random forest algorithm for the mapping process. To mitigate the redundancy effect of the PRISMA spectral variables, supervised band selection techniques—viz., Recursive Feature Elimination (RFE), Boruta, Variable Selection Using Random Forest (VSURF), and Genetic Algorithm (GA)—were utilized to derive the optimal bands for the prediction process. The efficiency of the models trained for each subset was determined using data partitioning and repeated <i>k</i>-fold cross-validation techniques (10 folds; 10 repetitions) with several validation metrics opted for continuous and categorical variables. Among the spectral subsets utilized, the PRISMA bands selected through embedded selection methods (Boruta and VSURF) and the Landsat 8 spectral datasets under both validation strategies had the highest efficiency in predicting the soil infiltration and textural classes, respectively. Further, the variable importance measure determined for both continuous and categorical predictions through the permutation feature importance (PFI) indicated that Physiography, Geomorphology, Land Use and Land Cover, Multiresolution Ridge Top Flatness (MRRTF), and Multiresolution Valley Bottom Flatness (MRVB), among others, made the highest contribution towards the entire modelling process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative assessment of spectral covariates from Sentinel 1 A, Sentinel 2 A, Landsat 8, and PRISMA for digital soil mapping of infiltration rate and textural classes

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Nivas Raj Moorthi,
  • Kumaraperumal Ramalingam,
  • Pazhanivelan Sellaperumal,
  • Muthumanickam Dhanaraju,
  • Sivasubramanian K,
  • Ragunath Kaliaperumal,
  • Prabu Padanillay Chidambaram

摘要

Prediction of soil properties at different depth levels and scales requires the use of multimodal datasets to represent the characteristics of the SCORPAN factors within the digital soil mapping (DSM) technique. With spectral variables representing the dynamic nature of soil property estimation, evaluating their efficiency can enhance the applicability of the models for further standardization. The current study aimed to compare the utility of different spectral modalities (i.e., Sentinel 1A, Sentinel 2A, Landsat 8, and PRISMA) for digital soil mapping of soil infiltration rates (cm h−1) and textural classes in the Thiruparankundram block of Madurai district. Further, the utility of the spectral subsets, both independently and in combination with other SCORPAN variables, was studied to perceive the efficiency of the framework and to determine the integration potential of spatial variables from multiple modalities. A total of 224 soil observations were derived from the legacy soil maps using a stratified random sampling procedure. The extracted soil and the covariate information were then subjected to the random forest algorithm for the mapping process. To mitigate the redundancy effect of the PRISMA spectral variables, supervised band selection techniques—viz., Recursive Feature Elimination (RFE), Boruta, Variable Selection Using Random Forest (VSURF), and Genetic Algorithm (GA)—were utilized to derive the optimal bands for the prediction process. The efficiency of the models trained for each subset was determined using data partitioning and repeated k-fold cross-validation techniques (10 folds; 10 repetitions) with several validation metrics opted for continuous and categorical variables. Among the spectral subsets utilized, the PRISMA bands selected through embedded selection methods (Boruta and VSURF) and the Landsat 8 spectral datasets under both validation strategies had the highest efficiency in predicting the soil infiltration and textural classes, respectively. Further, the variable importance measure determined for both continuous and categorical predictions through the permutation feature importance (PFI) indicated that Physiography, Geomorphology, Land Use and Land Cover, Multiresolution Ridge Top Flatness (MRRTF), and Multiresolution Valley Bottom Flatness (MRVB), among others, made the highest contribution towards the entire modelling process.