Remote Sensing Retrieval of Cloud Top Height Using Neural Networks and Data from Cloud-Aerosol Lidar with Orthogonal Polarization

Yinhe Cheng, Hongjian He, Qiangyu Xue, Jiaxuan Yang, Wei Zhong, Xinyu Zhu, Xiangyu Peng2024 · DOI

TL;DR

This paper presents a novel approach to improve the accuracy of Cloud Top Height (CTH) retrieval from MODIS satellite data by leveraging neural networks trained with data from the CALIPSO lidar. The study demonstrates that combining specific MODIS cloud parameters with calibrated radiance data yields the most accurate CTH estimates, outperforming existing MODIS products and offering a more detailed understanding of typhoon structures.

What Problem Does This Paper Address?

Accurately measuring Cloud Top Height (CTH) is crucial for understanding Earth's radiation budget, climate modeling, and weather forecasting. While passive sensors like MODIS provide extensive global coverage, their CTH retrieval can suffer from inaccuracies, especially with multi-layer clouds or thin cirrus. Active sensors like CALIPSO offer more precise CTH measurements but have limited spatial coverage. This paper aims to bridge this gap by developing a more accurate CTH retrieval method for MODIS data, using CALIPSO as a ground truth, thereby enhancing the utility of MODIS data for climate and weather applications.

Methods

The core of this research lies in applying neural networks, specifically backpropagation neural networks (BP NNs), to learn the complex relationship between MODIS observations and CALIPSO-derived CTH. The authors explored three main types of input strategies:

  1. MODIS L2 Cloud Product Data: Utilizing existing cloud parameters like cloud top pressure (CTP) and cloud top temperature (CTT).
  2. MODIS L1 Calibrated Radiance Data: Employing raw radiance values from various spectral bands.
  3. Combination of Cloud Parameters and Radiance: Integrating both approaches to leverage their respective strengths.

Seven different neural network models were trained and evaluated based on these input strategies. The models were trained using a large dataset of matched MODIS and CALIPSO data from 2008, with 2009 data used for validation. Model performance was assessed using metrics like Mean Absolute Error (MAE), Mean Squared Error (MSE), Root-Mean-Squared Error (RMSE), and the coefficient of determination (R²).

Datasets

  • MODIS (Moderate-Resolution Imaging Spectroradiometer) Data: Both L1 (calibrated radiance) and L2 (cloud product) data from the Aqua satellite were used. This includes spectral radiance measurements across 36 bands and derived cloud parameters.
  • CALIPSO (Cloud-Aerosol Lidar with Orthogonal Polarization) Data: CALIPSO L2 data products were used as the ground truth for CTH. This data provides high-vertical-resolution profiles of clouds and aerosols.

The study matched MODIS and CALIPSO data based on spatiotemporal proximity (within 3 minutes and 1 km) and cloud type consistency. A dataset from a specific MODIS image on August 9, 2019, was also used for a case study.

Key Insights and Ideas

  • CTP and CTT are Crucial: Cloud Top Pressure (CTP) and Cloud Top Temperature (CTT) were identified as the most critical parameters for accurate CTH retrieval from MODIS data. Models that included these parameters generally performed better.
  • Combination is Key: Models that combined MODIS cloud parameters (especially CTP and CTT) with calibrated radiance data (Type I and Type III models) significantly outperformed models relying solely on radiance or only on a limited set of cloud parameters.
  • Radiance Alone is Insufficient: Models using only calibrated radiance data (Type II models) showed poorer applicability and higher errors, suggesting that raw radiance alone is not enough for precise CTH retrieval without incorporating cloud physics information.
  • Neural Networks Excel: The study highlights the effectiveness of neural networks in capturing complex, non-linear relationships between satellite observations and cloud properties, leading to improved retrieval accuracy.
  • Typhoon Structure Revealed: The best-performing models were able to distinguish fine-scale structures within a typhoon, such as the eye, eyewall, and spiral arm regions, which is a significant improvement over standard MODIS products.

Key Results

  • Model 7 is Recommended: Model 7, which combined CTP, CTT, and specific radiance bands (20, 28, 31, 32, 33, 36), achieved the lowest RMSE and demonstrated the best overall performance, particularly in identifying typhoon structures and global CTH patterns.
  • Improved Accuracy: The best models (Type I and Type III) showed lower errors (MAE, RMSE) and higher goodness-of-fit (R²) and correlation (PCC) compared to the original MODIS CTH product. For instance, Model 7 achieved an R² of 0.97 and a PCC of 0.94.
  • Bias Reduction: The proposed models, especially Model 7, showed a significant reduction in bias for high and middle clouds compared to the MODIS product.
  • Global Consistency: Model 7's global CTH retrievals showed strong agreement with CALIPSO data, mirroring seasonal variations and latitudinal patterns, confirming its broad applicability.
  • Low Cloud Challenges Remain: While improvements were made, all models still showed some discrepancies for low clouds, indicating an area for future research.

Limitations and Future Work

  • Low Cloud Retrieval: The paper acknowledges that the accuracy of CTH retrieval for low clouds still leaves room for improvement across all models.
  • Neural Network Architecture: The study notes that the choice of neural network architecture (number of hidden layers and neurons) can significantly affect performance, suggesting further research into optimal architectures.
  • Empirical Models: Neural network models are empirical and require substantial training data. While generalization was tested, further validation across diverse conditions and regions would be beneficial.
  • Noise in Radiance Data: The presence of "stripe noise" in MODIS radiance data was identified as a potential issue, and future work could focus on more robust noise reduction techniques.
  • Further Exploration of Type II Models: The poor performance of models relying solely on radiance data warrants further investigation into why this approach is less effective and how it might be improved.

Summary updated

BibTeX
@misc{cheng2024remote,
  author = {Yinhe Cheng and Hongjian He and Qiangyu Xue and Jiaxuan Yang and Wei Zhong and Xinyu Zhu and Xiangyu Peng},
  title = {{Remote Sensing Retrieval of Cloud Top Height Using Neural Networks and Data from Cloud-Aerosol Lidar with Orthogonal Polarization}},
  year = {2024},
  doi = {10.3390/s24020541},
}
Figure