Frontiers in Remote Sensing (Sep 2024)
Prediction of soil texture using remote sensing data. A systematic review
Abstract
Soil particle size fractions play a critical role in determining soil health attributes, including soil aeration, water infiltration and retention capacity, nutrients, and organic matter dynamics. Traditional soil mapping methods rely predominantly on ground-based surveys and laboratory analysis which are reported to be time-consuming and expensive. To address these challenges, there has been a global shift towards digital soil mapping (DSM) techniques that utilize remote sensing data. This review, conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guideline, aims to provide a comprehensive synthesis of the current state of soil texture prediction using remote sensing data. In particular, the review extract and synthesizes the satellite images used, identify the derived environmental covariates and their relative importance, and assesses the prediction models/algorithms used in the prediction of soil texture. Synthesis and analysis of 70 articles show that clay content is the most predicted of the three soil particle fractions accounting for 37% of the reviewed studies predominantly from topsoil layer (74.29%). Sentinel 2 and Landsat 8 are reported as the most frequently used satellite images. Among the covariates derived from these images, NDVI (80.4%) and SAVI (60.8%) are by far the most derived band ratios (indices). Red (37.3%), NIR (35.3%), Green (33.3%), Blue (33.3%), and SW2 (29.4%) bands were the five most incorporated as covariates for soil texture prediction amongst individual satellite bands. Regarding the DSM algorithms, Random Forest (RF) appeared in most reviewed articles followed by Support Vector Machines (SVM), and Quantile Regression Forest (QRF). The comparative model performance analysis showed that RF and Artificial neural network (ANN) had a good trade-off across validation metrics indicating their best performance in the prediction of both clay, sand, and silt. The RF performance showed a decreasing trend with increasing depth interval for clay and sand prediction and inconsistent for silt prediction.
Keywords
