Scholarlink

<p>Accurate crop mapping is essential for effective agricultural management, food security, and environmental monitoring. While optical satellite imagery is widely employed for crop mapping, the presence of cloud cover often diminishes classification accuracy. This study presents an integrated approach to generate cloud gap filled Sentinel-2 images for improved in-season crop type classification using machine learning techniques. The study utilizes the blue, green, red, near infra-red, and red-edge spectral bands. The study is structured into three sequential phases: (1) cloud detection, (2) cloud reconstruction, and (3) crop classification. For cloud detection, three supervised machine learning classifiers – Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boost (XGBoost) were evaluated, with SVM achieving the highest accuracy (96.03%). The resulting cloud mask from cloud detection model was then used to generate cloud-masked images for subsequent processing. Next, cloud reconstruction was performed using two regression models (RF and XGBoost), with XGBoost selected as the optimal approach based on performance metrics. Finally, the reconstructed cloud-free imagery was utilized for crop classification. Crop classification utilized an XGBoost classification model integrating Normalized Difference Vegetation Index time series and multispectral data. Classification results demonstrated a significant improvement in accuracy when using reconstructed imagery compared to the original cloud-affected images. Specifically, the paddy crop map derived from reconstructed images achieved an overall accuracy of 82.53% (kappa = 0.653), outperforming the map generated from original images (75.99% accuracy, kappa = 0.523). This 6.54 percentage-point increase in accuracy highlights the value of the cloud reconstruction process for agricultural monitoring applications. This approach has the potential to enhance agricultural decision-making, resource management, and environmental analysis, supporting various applications such as crop insurance, market forecasting and supply-chain logistics.</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

Generating Cloud-Free Sentinel-2 Images for Enhanced In-Season Crop Type Classification: a Case Study

  • 期刊论文

  • 发表日期：

    2025年11月8日

查看全文
  • Shikha Gupta,
  • Prabhakar Alok Verma

摘要

Accurate crop mapping is essential for effective agricultural management, food security, and environmental monitoring. While optical satellite imagery is widely employed for crop mapping, the presence of cloud cover often diminishes classification accuracy. This study presents an integrated approach to generate cloud gap filled Sentinel-2 images for improved in-season crop type classification using machine learning techniques. The study utilizes the blue, green, red, near infra-red, and red-edge spectral bands. The study is structured into three sequential phases: (1) cloud detection, (2) cloud reconstruction, and (3) crop classification. For cloud detection, three supervised machine learning classifiers – Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boost (XGBoost) were evaluated, with SVM achieving the highest accuracy (96.03%). The resulting cloud mask from cloud detection model was then used to generate cloud-masked images for subsequent processing. Next, cloud reconstruction was performed using two regression models (RF and XGBoost), with XGBoost selected as the optimal approach based on performance metrics. Finally, the reconstructed cloud-free imagery was utilized for crop classification. Crop classification utilized an XGBoost classification model integrating Normalized Difference Vegetation Index time series and multispectral data. Classification results demonstrated a significant improvement in accuracy when using reconstructed imagery compared to the original cloud-affected images. Specifically, the paddy crop map derived from reconstructed images achieved an overall accuracy of 82.53% (kappa = 0.653), outperforming the map generated from original images (75.99% accuracy, kappa = 0.523). This 6.54 percentage-point increase in accuracy highlights the value of the cloud reconstruction process for agricultural monitoring applications. This approach has the potential to enhance agricultural decision-making, resource management, and environmental analysis, supporting various applications such as crop insurance, market forecasting and supply-chain logistics.