Optimizing Temporal Window for Constructing Efficient Bare Soil Composites from Sentinel-2 Data for Estimating Soil Properties in Semi-Arid Tropics

Majeed, I and Sahoo, D P and Roy, S and Reddy, N N and Budama, N and Akuraju, V R and Singh, R and Das, B S and Garg, K K (2026) Optimizing Temporal Window for Constructing Efficient Bare Soil Composites from Sentinel-2 Data for Estimating Soil Properties in Semi-Arid Tropics. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. pp. 1-16. ISSN 1939-1404

Text - Accepted Version Available under License Creative Commons Attribution. Download (1MB)

Abstract

Bare soil composites (BSCs) derived from Sentinel-2 (S2) imagery offer a promising approach for estimating soil properties. However, the influence of temporal window length and spatial heterogeneity on predictive performance remains insufficiently understood. This study evaluated BSCs developed from short-term (one-year), medium-term (three-year), and long term (six-year) temporal windows for estimating soil properties in a semi-arid agricultural region of Maharashtra, India. A total of 443 surface soil samples were collected from ground reference locations distributed across five districts representing diverse soil conditions and land-use practices. Four regression models, namely support vector regression (SVR), partial least squares regression (PLSR), random forest (RF), and feature selection-based PLSR (PLSRFS), were evaluated using Sentinel-2 spectral data. Laboratory reflectance spectra were also compared with each BSC using the Spectral Angle Mapper (SAM) to identify the Sentinel-2 dataset with the highest spectral agreement. Moreover, the fuzzy c-means (FCM) membership function was explored as a novel measure of soil spectral heterogeneity for use as a predictor variable along with S2 data. Results showed that S2 data derived from the long-term temporal window (six-year BSC) had the highest spectral agreement with laboratory spectra (mean SAM = 0.09), with ∼50% of samples exhibiting SAM < 0.1. Yet, the coefficient of determination (R2) values in the test datasets developed for a given temporal window varied with the target soil property. For example, for sand, R² increased from 0.22 to 0.33, RMSE decreased from 13.0 to 12.1, and RPIQ increased from 1.69 to 1.88; for clay, R² increased from 0.24 to 0.37, RMSE (root mean square error) decreased from 11.1 to 10.2, and RPIQ increased from 1.82 to 1.96; and for SOC, R² increased from 0.15 to 0.28, RMSE decreased from 0.13 to 0.12, and ratio of performance to interquartile range (RPIQ) increased from 1.56 to 1.72, respectively. Among the evaluated soil properties, Mn showed the highest predictive performance (R² = 0.76, RMSE = 8.86, RPIQ = 2.13). Incorporating FCM membership values as an additional predictor consistently improved model performance, increasing R² from 0.33 to 0.56, 0.37 to 0.53, and 0.61 to 0.72, and RPIQ from 1.88 to 2.20, 1.96 to 2.23, and 1.41 to 1.80 for sand, clay, and available Fe, respectively. Moderate improvements were also observed for pH and SOC. Clearly, longer temporal windows for BSCs and a soil heterogeneity measure provide an improved pathway for high-resolution soil assessment in heterogeneous landscapes.

Item Type:	Article
Divisions:	Research Program - Resilient Farming Systems
CRP:	UNSPECIFIED
Uncontrolled Keywords:	Soil, Modeling, Soil properties, Windows, Reflectivity, Sentinel-2, Distance measurement, Magnesium, System-on-chip, Iron
Subjects:	Others > GIS Techniques/Remote Sensing Others > Semi-Arid Tropics Others > Soil
Depositing User:	Mr Nagaraju T
Date Deposited:	28 Oct 2026 09:49
Last Modified:	28 Oct 2026 09:49
URI:	http://oar.icrisat.org/id/eprint/13683
Official URL:	https://ieeexplore.ieee.org/document/11753463
Projects:	UNSPECIFIED
Funders:	UNSPECIFIED
Acknowledgement:	The authors acknowledge the support of the Groundwater Surveys and Development Agency (GSDA), Government of Maharashtra, under the Atal Bhujal Yojana for developing benchmark sites. The Multifunctional Landscape (MFL) Science Program of CGIAR is also acknowledged for partially supporting the time of the lead and corresponding authors.
Links:

View Statistics

Actions (login required)

View Item