Brickner et al. (2025) Field crop mapping using machine learning and multi-sensor satellite fusion: toward dynamic agricultural monitoring

Identification

• Journal: Smart Agricultural Technology
• Year: 2025
• Date: 2025-11-20
• Authors: Nechama Z. Brickner, Lior Fine, Offer Rozenstein, Tarin Paz‐Kagan
• DOI: 10.1016/j.atech.2025.101650

Research Groups

• French Associates Institute for Agriculture and Biotechnology of Dryland, The Jacob Blaustein Institutes for Desert Research, Ben-Gurion University of the Negev, Sede Boqer Campus, Israel
• Institute of Soil, Water and Environmental Sciences, Agricultural Research Organization– Volcani Institute, Rishon LeZion, Israel

Short Summary

This study developed a novel hierarchical machine learning framework integrating Sentinel-1 SAR and Sentinel-2 multispectral data to generate high-resolution, multi-season crop type maps in dryland agricultural regions. The framework achieved high accuracy in classifying agricultural land cover and specific crop types, enabling dynamic monitoring of crop rotation, land-use intensity, and climate-driven phenological gradients.

Objective

• To develop a multi-sensor, automated crop hierarchical classification approach for mapping agricultural land cover, field crops, and wheat, with a particular focus on the western Negev, Israel.
• To identify and characterize changes in agricultural land use, including shifts in crop types, phenological deviations, or temporary field abandonment related to environmental conditions.
• To examine intra- and inter-annual variability in wheat phenology along an aridity gradient using vegetation index time-series (VI-TS) analysis, with the aim of disentangling the effects of climatic factors.

Study Configuration

• Spatial Scale: Approximately 1700 square kilometers (km²) in the western Negev, Israel, with over 1000 km² used for agriculture.
• Temporal Scale: Agricultural years from October 2017 to September 2024 (seven agricultural years).

Methodology and Data

• Models used:
  • Random Forest (RF) for hierarchical classification.
  • Boruta algorithm for feature selection.
  • Shapley Additive exPlanations (SHAP) for feature importance assessment.
  • Mann-Kendall (MK) trend test (original and seasonal) for temporal trend analysis.
  • One-way analysis of variance (ANOVA) and Tukey’s Honest Significant Difference (HSD) test for phenological parameter comparison.
• Data sources:
  • Sentinel-1 (SAR) and Sentinel-2 (multispectral optical) satellite imagery from the Copernicus program (ESA), accessed via Google Earth Engine (GEE).
  • Vector data: Official GIS layers from the Israeli Ministry of Agriculture (MOAG), dedicated field surveys (2024), and commercial agricultural field mapping datasets (2018–2024).
  • High-resolution aerial imagery for labeling agricultural plots.
  • Climate data: Daily mean temperature (°C) from 10 monitoring stations and daily precipitation (mm) from 45 stations, provided by the Israel Meteorological Service (IMS) (2018–2024).
  • Aridity Index (AI) raster dataset.

Main Results

• Hierarchical Classification Accuracy:
  • Model 1 (Agricultural Land Cover: orchards, covered areas, bare fields, row crops): Overall Accuracy (OA) of 94.87 %, F1 Score of 94.87 %, Kappa Score of 91.21. Maximum NDVI and annual mean Plastic Greenhouse Index (PGHI) were the most influential features.
  • Model 2 (Wheat vs. Other Crops): OA of 95.17 %, F1 score of 95.18 %, Kappa score of 90.25. Wheat was correctly classified in 96.78 % of cases. Start of Season Day of Year (SOS-DOY), VH-Ascending (VH-ASC), and Max of Season Day of Year (MOS-DOY) were most influential.
  • Model 3 (11 Specific Crop Types + "Other"): OA of 81.15 %, F1 Score of 79.43 %, Kappa Score of 78.25 %. High accuracy for cotton (98.17 %), peanut (93.33 %), and watermelon (97.5 %). Lower accuracy for melon (10.64 %), pumpkin (29.41 %), and tomato (56.0 %). MOS-DOY, End of Season Day of Year (EOS-DOY), and SOS-DOY were most influential.
• Land Cover Change Detection: The most frequent transitions occurred between bare fields and row crops, representing 11 % to 15 % of plots annually. Orchards and covered plots showed high temporal stability.
• Trend Analysis: Significant increasing trends were observed for orchards and corn (annual scale), and for covered plots and orchards (monthly scale). Significant decreasing trends were found for chickpeas and watermelons (monthly scale). Among consistently classified wheat plots, 51.01 % showed a significant increasing trend in NDVI, 34.95 % a decreasing trend, and 14.02 % no significant trend.
• Wheat Spatiotemporal Patterns: A clear north–south aridity gradient influenced wheat productivity. Northern zones exhibited significantly higher productivity (longer Length of Season (LOS), higher MOS NDVI values, and greater Area Under the Curve (AUC)) and later SOS dates compared to southern zones.

Contributions

• Developed a novel hierarchical Random Forest classification framework integrating multi-sensor (Sentinel-1 SAR and Sentinel-2 optical) satellite data for high-resolution, multi-season crop type mapping in dryland agricultural systems.
• Introduced a MOS-based cross-index and MOS-aligned SAR feature selection approach that synchronizes optical and radar inputs at phenological peaks, enhancing classification across crop types.
• Employed an expert-driven feature selection strategy informed by agronomic knowledge, identifying biologically meaningful predictors beyond purely statistical rankings.
• Demonstrated robust multi-year generalization of the hierarchical workflow across seven years (2018–2024) in a fragmented dryland environment, providing a transferable framework for operational crop monitoring at regional to national scales.
• Provided fine-scale, plot-level monitoring of land cover transitions and productivity trends, revealing spatial heterogeneity and localized changes often overlooked by traditional methods.
• Established a new spatial link between phenological productivity metrics and aridity gradients, offering a scalable framework for identifying future vulnerable agricultural areas.

Funding

• Ministry of Agriculture and Food Security (Grant No. 304081400)

Citation

@article{Brickner2025Field,
  author = {Brickner, Nechama Z. and Fine, Lior and Rozenstein, Offer and Paz‐Kagan, Tarin},
  title = {Field crop mapping using machine learning and multi-sensor satellite fusion: toward dynamic agricultural monitoring},
  journal = {Smart Agricultural Technology},
  year = {2025},
  doi = {10.1016/j.atech.2025.101650},
  url = {https://doi.org/10.1016/j.atech.2025.101650}
}