Digital soil mapping using remote sensing indices, terrain attributes, and vegetation features in the rangelands of northeastern Iran

Abstract

Digital soil mapping has been introduced as a viable alternative to the traditional mapping methods due to being fast and cost-effective. The objective of the present study was to investigate the capability of the vegetation features and spectral indices as auxiliary variables in digital soil mapping models to predict soil properties. A region with an area of 1225 ha located in Bajgiran rangelands, Khorasan Razavi province, northeastern Iran, was chosen. A total of 137 sampling sites, each containing 3-5 plots with 10-m interval distance along a transect established based on randomized-systematic method, were investigated. In each plot, plant species names and numbers as well as vegetation cover percentage (VCP) were recorded, and finally one composite soil sample was taken from each transect at each site (137 soil samples in total). Terrain attributes were derived from a digital elevation model, different bands and spectral indices were obtained from the Landsat7 ETM+ images, and vegetation features were calculated in the plots, all of which were used as auxiliary variables to predict soil properties using artificial neural network, gene expression programming, and multivariate linear regression models. According to R RMSE and MBE values, artificial neutral network was obtained as the most accurate soil properties prediction function used in scorpan model. Vegetation features and indices were more effective than remotely sensed data and terrain attributes in predicting soil properties including calcium carbonate equivalent, clay, bulk density, total nitrogen, carbon, sand, silt, and saturated moisture capacity. It was also shown that vegetation indices including NDVI, SAVI, MSAVI, SARVI, RDVI, and DVI were more effective in estimating the majority of soil properties compared to separate bands and even some soil spectral indices.

Selecting appropriate machine learning methods for digital soil mapping

“…Using ANN, Mahmoudabadi et al [56] predicted total nitrogen with only slightly higher accuracy than Kalambukattu et al [48]; R 2 of independent validation was 0.67 and 0.62, respectively. This difference between the R 2 values was negligible considering Mahmoudabadi et al [56] used nearly twice as many sample points and included digital terrain derivatives in addition to remote sensing in the covariate pool. Although our experience has been that results can be improved by manually calibrating hyperparameters, using caret to calibrate the hyperparameters helped to standardize the time taken to complete this step.…”

“…Even though Angelini et al [28] covered a broader selection of covariates by also including climate in the covariate pool, the R 2 of CV for their study was only 0.28, compared to 0.50 for Zeraatpisheh et al [14]. Using ANN, Mahmoudabadi et al [56] predicted total nitrogen with only slightly higher accuracy than Kalambukattu et al [48]; R 2 of independent validation was 0.67 and 0.62, respectively. This difference between the R 2 values was negligible considering Mahmoudabadi et al [56] used nearly twice as many sample points and included digital terrain derivatives in addition to remote sensing in the covariate pool.…”

“…Even though sample size is not the only factor affecting R 2 when comparing different studies, a notable pattern can be observed. For example, comparing two studies predicting calcium carbonate equivalent (CCE) using MLR, the R 2 of independent validation improved from 0.06 with 120 observations [55] to 0.51 with 137 observation [56]. In those same studies, the R 2 for independent validation increased from 0.05 to 0.40 for predicting sand content.…”

Machine learning for digital soil mapping: applications, challenges and suggested solutions

“…To ensure that models fit the existing knowledge, they must be opened and understood in their functioning. Opening the "black box" is then necessary but not straightforward (see next section on interpretability), and is often reduced to the analysis of which environmental covariates are the most often used by the model to make a prediction (see for example Mahmoudabadi et al (2017) that ML algorithm should not be used for obtaining new soil knowledge because the ML algorithm aims at predicting a pattern rather than finding causal relationships. Wadoux et al (2019c) suggest to use calibrated ML models as a "hypothesis discovery" tool, in which the mechanisms conveyed by the calibrated ML model are supplied to the researcher for possible explanations of the soil process, which can then be confronted to experiments and principles of soil genesis.…”

Combining Object-Based Image Analysis with Topographic Data for Landform Mapping: A Case Study in the Semi-Arid Chaco Ecosystem, Argentina

“…Most of the studies based on DSM techniques combine multispectral satellite data with topographic data to improve geoform classifications, especially in complex landscapes [9][10][11], whereas traditional image analysis techniques use pixel-based classification approaches. However, when medium spatial resolution imagery is used in large areas, especially in land characterized by high intra-class spectral variability, analyzing pixels individually can produce misclassifications.…”