Scholarlink

<p>Accurate crop classification is essential for agricultural monitoring and food security. Although remote sensing provides rich spectral information, traditional classifiers often struggle with spectral redundancy, class imbalance, and overlapping crop signatures. This study introduces an Attention-guided Stacked Ensemble Network (ASEN) that integrates fused Landsat 8–9 imagery with an attention-based ensemble of multilayer perceptrons. An embedded feature selection strategy based on SHAP values identifies the most informative spectral bands and vegetation indices, reducing redundancy and enhancing interpretability. The framework was evaluated on 50,835 field-verified samples from six major crops in Central Punjab and achieved an overall accuracy of 98.43% and an F1-score of 89.29%. Results were averaged over five independent runs with corresponding standard deviations to ensure robustness. ASEN consistently outperformed conventional machine-learning baselines, demonstrating the value of combining spectral fusion, attention mechanisms, and embedded feature selection for reliable crop classification. Key limitations include the use of single-season imagery and a region-specific dataset. Future work will incorporate multi-temporal and multi-regional imagery, as well as complementary SAR and hyperspectral data, to improve scalability and phenological discrimination.</p> Graphical Abstract <p> This study presents a remote sensing-based framework for crop type classification in the irrigated regions of Central Punjab using Landsat 8-9 imagery and field-surveyed GPS data. A total of six major crops were geocoded and labeled through field visits conducted in early 2023. Satellite images were preprocessed through radiometric and atmospheric corrections, followed by image fusion and vegetation index extraction (NDVI, SAVO, RECI, NDRE). A comprehensive dataset of 50,835 points was used to train conventional classifiers, ensemble models, and artificial neural networks. Feature selection techniques were employed to optimize classification performance. The final output highlights the accurate identification of crop types, demonstrating the effectiveness of combining field-based data, remote sensing, and machine learning for agricultural monitoring and decision support.</p> <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Attention-Based Ensemble Learning for Crop Classification Using Landsat 8–9 Fusion

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Nisar Ahmed,
  • Zeeshan Ramzan,
  • Qurat-ul-Ain Akram,
  • Shahzad Asif,
  • Muhammad Shahbaz,
  • Rabin Chakrabortty,
  • Ahmed F. Elaksher

摘要

Accurate crop classification is essential for agricultural monitoring and food security. Although remote sensing provides rich spectral information, traditional classifiers often struggle with spectral redundancy, class imbalance, and overlapping crop signatures. This study introduces an Attention-guided Stacked Ensemble Network (ASEN) that integrates fused Landsat 8–9 imagery with an attention-based ensemble of multilayer perceptrons. An embedded feature selection strategy based on SHAP values identifies the most informative spectral bands and vegetation indices, reducing redundancy and enhancing interpretability. The framework was evaluated on 50,835 field-verified samples from six major crops in Central Punjab and achieved an overall accuracy of 98.43% and an F1-score of 89.29%. Results were averaged over five independent runs with corresponding standard deviations to ensure robustness. ASEN consistently outperformed conventional machine-learning baselines, demonstrating the value of combining spectral fusion, attention mechanisms, and embedded feature selection for reliable crop classification. Key limitations include the use of single-season imagery and a region-specific dataset. Future work will incorporate multi-temporal and multi-regional imagery, as well as complementary SAR and hyperspectral data, to improve scalability and phenological discrimination.

Graphical Abstract

This study presents a remote sensing-based framework for crop type classification in the irrigated regions of Central Punjab using Landsat 8-9 imagery and field-surveyed GPS data. A total of six major crops were geocoded and labeled through field visits conducted in early 2023. Satellite images were preprocessed through radiometric and atmospheric corrections, followed by image fusion and vegetation index extraction (NDVI, SAVO, RECI, NDRE). A comprehensive dataset of 50,835 points was used to train conventional classifiers, ensemble models, and artificial neural networks. Feature selection techniques were employed to optimize classification performance. The final output highlights the accurate identification of crop types, demonstrating the effectiveness of combining field-based data, remote sensing, and machine learning for agricultural monitoring and decision support.