Bormudoi et al. (2026) Disentangling Complexity and Performance: A Comparative Study of Deep Learning and Random Forest Models for Cropland Vulnerability Assessment in Bangladesh

Identification

• Journal: Land
• Year: 2026
• Date: 2026-01-16
• Authors: Arnob Bormudoi, Masahiko Nagai
• DOI: 10.3390/land15010174

Research Groups

• Center for Research and Application of Satellite Remote Sensing (YUCARS), Yamaguchi University, Ube, Yamaguchi, Japan
• New Space Intelligence, Ube, Yamaguchi, Japan
• Faculty of Engineering, Assam downtown University, Panikhaiti, Guwahati, India

Short Summary

This study compared deep learning and Random Forest models for cropland vulnerability assessment in Bangladesh using Earth Observation data and climate variables. The Random Forest model significantly outperformed the deep learning architecture, explaining 70% of cropland stress variance and identifying key biophysical drivers for early warning systems.

Objective

• To compare the predictive performance of spatially aware machine learning models (Deep Learning and Random Forest) for 2023 cropland Normalized Difference Vegetation Index (NDVI) anomalies in Bangladesh.
• To identify key biophysical drivers of cropland vulnerability through interpretability analysis of the best-performing model.

Study Configuration

• Spatial Scale: National-scale assessment of Bangladesh croplands at 250 meter (m) spatial resolution.
• Temporal Scale: 22-year time series (2001–2023) used for training and predicting 2023 NDVI anomalies.

Methodology and Data

• Models used: Dual-stream deep learning architecture (incorporating Gated Recurrent Unit - GRU), Random Forest regressor.
• Data sources:
  • Satellite: MODIS Terra Vegetation Indices 16-Day Global 250 m product (MODIS/006/MOD13Q1) for monthly NDVI.
  • Climate/Reanalysis:
    • CHIRPS (UCSB-CHG/CHIRPS/PENTAD) for monthly precipitation.
    • ERA5 Daily (ECMWF/ERA5/DAILY) for mean 2 m air temperature and total precipitation.
    • ERA5-Land Daily Aggregated (ECMWF/ERA5LAND/DAILYAGGR) for surface solar radiation and volumetric soil water.
  • Static Environmental:
    • OpenLandMap (OpenLandMap/SOL) for soil clay, sand, and pH content.
    • SRTM 90 m (CGIAR/SRTM90_V4) for elevation (from which slope and aspect were derived).
    • 2021 ESA WorldCover land cover map for cropland mask.

Main Results

• The Random Forest (RF) model (R² = 0.70, Root Mean Squared Error (RMSE) = 197.03) substantially outperformed the dual-stream deep learning (DL) architecture (R² = 0.02, RMSE = 357.57) in predicting 2023 cropland NDVI anomalies.
• The RF model explained 70% of the variance in cropland stress.
• Feature importance analysis of the RF model identified March precipitation, February NDVI, and March vapor pressure deficit (VPD) from the prior year as the primary biophysical drivers of cropland vulnerability.
• Static environmental factors such as soil clay, soil pH, and elevation also contributed but were less influential than the temporal climate-vegetation variables.
• Spatial analysis revealed that Natore (+1.21 Z-score) and Magura (+0.89 Z-score) districts exhibited elevated stress consistent with 2023 drought conditions.
• Dinajpur (−2.27 Z-score) and Faridpur (−2.16 Z-score) districts showed conditions more favorable than their historical norms.
• A national-scale cropland vulnerability map for 2023 was generated at 250 m spatial resolution.

Contributions

• Demonstrated that simpler, interpretable machine learning models (Random Forest) can outperform complex deep learning architectures for agricultural vulnerability assessment, particularly when training data exhibit spatial structure and sample sizes are constrained relative to feature dimensionality.
• Developed a reproducible framework integrating freely available Earth Observation data (MODIS, ERA5, CHIRPS) with open-source machine learning tools for national-scale vulnerability assessment.
• Provided a 250 m resolution cropland vulnerability map for Bangladesh, enabling more precise targeting of climate adaptation interventions.
• Identified key biophysical drivers (March precipitation, February NDVI, March VPD) that offer a 2–3 month lead time for early warning systems and targeted adaptation strategies.
• Employed a rigorous spatially aware block cross-validation strategy to ensure robust model evaluation.

Funding

This research received no external funding.

Citation

@article{Bormudoi2026Disentangling,
  author = {Bormudoi, Arnob and Nagai, Masahiko},
  title = {Disentangling Complexity and Performance: A Comparative Study of Deep Learning and Random Forest Models for Cropland Vulnerability Assessment in Bangladesh},
  journal = {Land},
  year = {2026},
  doi = {10.3390/land15010174},
  url = {https://doi.org/10.3390/land15010174}
}