Presentation Information
[U06-P04]Automated Multi-Temporal Bare-Soil Mapping in Agricultural Landscapes Using Harmonized Landsat–Sentinel Data and Machine Learning
*Miguel Conrado Valdez1, Chi-Farn Chen (1.324588)
Keywords:
Bare soil detection,Satellite remote sensing,Harmonized Landsat–Sentinel (HLS),Machine learning,Agricultural land surface
Bare-soil mapping is a fundamental component of agricultural monitoring, land-surface characterization, and environmental modelling, particularly in intensively managed regions where cropping cycles and field practices generate highly dynamic surface conditions. Reliable identification of exposed soil supports many downstream applications, including soil organic carbon (SOC) estimation, land-surface model calibration, evapotranspiration retrieval, and land degradation assessment. Despite its importance, operational bare-soil detection remains challenging due to spectral confusion with sparse vegetation, dry crop residues, and heterogeneous mineral surfaces, as well as seasonal variability and illumination effects. The recent availability of the Harmonized Landsat-Sentinel (HLS) dataset, providing 30 m spatial resolution and a 2-3 day revisit frequency, offers new opportunities for constructing detailed, multi-temporal bare-soil datasets. In this study, we develop an automated workflow that integrates spectral index thresholding and machine-learning techniques to map bare soil across agricultural regions using HLS surface reflectance imagery. A comprehensive suite of spectral indices is employed to capture vegetation, soil brightness, moisture, and disturbance signals, including the Normalized Difference Vegetation Index (NDVI), Bare Soil Index (BSI), Normalized Difference Moisture Index (NDMI), Soil Adjusted Vegetation Index (SAVI/GSAVI), Normalized Burn Ratio (NBR), Enhanced Vegetation Index (EVI), and the Dry Bare Soil Index (DBSI). High-confidence bare-soil pixels are initially identified using a conservative, rule-based classification with strict thresholds (e.g., NDVI < 0.2 and BSI > 0.4-0.7), minimizing commission errors and generating reliable presence samples for model calibration. To improve generalization across varying day-of-year (DOY) mosaics and reduce reliance on fixed thresholds, these reference samples are incorporated into a presence-background modelling framework using the MaxEnt algorithm implemented via the maxnet package. MaxEnt is well suited for remote-sensing applications due to its robustness to absence uncertainty, ability to capture complex interactions among predictors, and production of stable clog-log probability outputs. For each DOY, background samples are constrained to non-bare or vegetated areas (e.g., NDVI >= 0.3), ensuring plausible contrasts between bare and non-bare conditions. Model performance is evaluated using AUC, Kappa, sensitivity, specificity, and threshold-dependent accuracy metrics, while permutation importance and jackknife analyses quantify the contribution of individual spectral predictors. The resulting products include high-resolution bare-soil probability maps and binary masks for each DOY, which can be aggregated to derive bare-soil frequency maps, crop-cycle exposure indicators, and temporal soil-visibility profiles. These outputs provide valuable inputs for SOC modelling, tillage and erosion assessment, and agricultural land-use characterization. The proposed workflow is fully reproducible, scalable, and suitable for annual to multi-year monitoring of bare-soil dynamics across diverse agricultural landscapes.