Ahmed et al. (2025) Attention-Based Ensemble Learning for Crop Classification Using Landsat 8–9 Fusion

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-12-12
• Authors: Nisar Ahmed, Zeeshan Ramzan, Qurat-ul-Ain Akram, Shahzad Asif, Muhammad Shahbaz, Rabin Chakrabortty, Ahmed F. Elaksher
• DOI: 10.1007/s41748-025-00968-6

Research Groups

• Department of Informatics and Systems, University of Management and Technology, Lahore, Pakistan
• Department of Computer Science (New Campus), University of Engineering and Technology, Lahore, Pakistan
• Department of Computer Engineering, University of Engineering and Technology, Lahore, Pakistan
• Department of Civil Engineering, College of Engineering, American University of Sharjah, Sharjah, United Arab Emirates
• Department of Engineering Technology and Survey Engineering, New Mexico State University, Las Cruces, NM, USA

Short Summary

This study developed an Attention-guided Stacked Ensemble Network (ASEN) leveraging fused Landsat 8–9 imagery and SHAP-based feature selection for crop classification, achieving 98.43% overall accuracy and an 89.29% F1-score for six major crops in Central Punjab, Pakistan.

Objective

• To develop and evaluate an Attention-guided Stacked Ensemble Network (ASEN) that integrates Landsat 8–9 spectral fusion, SHAP-based embedded feature selection, and an attention mechanism to improve crop classification accuracy and interpretability, addressing challenges like spectral redundancy and class imbalance.

Study Configuration

• Spatial Scale: Central Punjab region of Pakistan, specifically four districts: Bahawalpur, Jhang, Chiniot, and Sargodha.
• Temporal Scale: Single-date Landsat 8–9 imagery acquired during January–February 2023, synchronized with field surveys.

Methodology and Data

• Models used:
  • Attention-guided Stacked Ensemble Network (ASEN) as the proposed model, utilizing an ensemble of Multilayer Perceptrons (MLPs) as base learners.
  • Adam optimizer and categorical cross-entropy loss for ASEN training.
  • Baseline classifiers: Support Vector Machine (SVM), Logistic Regression, Random Forest, and Gradient Boosting.
• Data sources:
  • Field-verified samples: 50,835 geocoded samples for six major crops (wheat, sugarcane, maize, potato, mustard, cotton) collected via field surveys in Central Punjab, Pakistan.
  • Satellite imagery: Fused Landsat 8 and Landsat 9 imagery (single-date, cloud-free acquisitions from January–February 2023), preprocessed using Google Earth Engine.
  • Derived features: 10 vegetation indices (NDVI, EVI, SAVI, GNDVI, RENDVI, MSAVI, NDWI, NDRE, SR, PRI) and raw spectral bands from Landsat 8–9.
  • Feature selection: SHAP (SHapley Additive exPlanations) values for embedded feature selection.

Main Results

• The Attention-guided Stacked Ensemble Network (ASEN) achieved an overall classification accuracy of 98.43% and an F1-score of 89.29%.
• ASEN consistently outperformed conventional machine-learning baselines (SVM, Logistic Regression, Random Forest, Gradient Boosting) across all evaluation metrics (Accuracy, Precision, Recall, F1-Score, AUC-ROC).
• The model demonstrated reduced misclassification rates for spectrally similar crop pairs (e.g., Mustard vs. Potato, Wheat vs. Maize) compared to baselines.
• Near-Infrared (NIR) and Shortwave Infrared (SWIR) bands, along with vegetation indices like NDVI and EVI, were identified as the most significant features for crop classification.
• SHAP-based feature selection improved ASEN's classification accuracy, F1-score, and AUC, enhancing generalization and reducing training complexity.

Contributions

• Introduction of the Attention-guided Stacked Ensemble Network (ASEN), integrating Landsat 8–9 spectral fusion with an attention-driven deep ensemble of multilayer perceptrons.
• Implementation of SHAP-based embedded feature selection to identify the most informative spectral bands and vegetation indices, reducing redundancy and enhancing model interpretability.
• Provision of an extensive comparative analysis against traditional classifiers (SVM, Logistic Regression, Random Forest, Gradient Boosting) and recent deep learning approaches.
• Offering practical insights for precision agriculture applications, while explicitly acknowledging limitations related to single-season and single-region sampling.

Funding

The study has not obtained any specific funding.

Citation

@article{Ahmed2025AttentionBased,
  author = {Ahmed, Nisar and Ramzan, Zeeshan and Akram, Qurat-ul-Ain and Asif, Shahzad and Shahbaz, Muhammad and Chakrabortty, Rabin and Elaksher, Ahmed F.},
  title = {Attention-Based Ensemble Learning for Crop Classification Using Landsat 8–9 Fusion},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00968-6},
  url = {https://doi.org/10.1007/s41748-025-00968-6}
}