9 December 2024 in IEEE Transactions on Geoscience and Remote Sensing
Weak atmospheric signals received by radar wind profilers (RWPs) often overlap with the clutter, making it difficult to identify pure atmospheric spectra. A solution with great potential is to develop RWPs with adaptive dynamic identification capabilities. We therefore propose a multi-feature dynamic weighting (MFDW) method and specifically define a score function that includes spectral statistical features and their weights. This score function first calculates multiple spectral features and constructs a prediction function that maps the features to prediction values. Then, the weights are determined by Gaussian curve fitting and contributed to the predicted score calculation to extract pure atmospheric spectra. After that, the multi-parameter cost function (MPCF) method is first improved by Doppler window settings, and then utilized to estimate accurate spectral moments and profiles. MFDW method is validated using RWP data from different regions and frequencies, and the extracted profiles are compared with GPS radiosonde (GPS RS) data. Research results confirm that dynamic weighting is adaptable to various overlap cases. Furthermore, Doppler wind profile estimation with this proposed method has a good match with GPS RS data. Both the detection accuracy and the running time excel many currently adopted methods. In brief, the chief contributions in this paper lies in our innovative MFDW method and feature- and weights-involved score function.
13 May 2026 in IEEE Transactions on Geoscience and Remote Sensing
Accurate wind profile estimation from 915- and 1290-MHz boundary layer radar wind profilers (RWP) remains particularly challenging under complex atmospheric conditions with low signal-to-noise ratios (SNR), where conventional methods often lack generalizability due to their limited adaptability to spectral contamination and signal variability. To address these limitations, we propose a novel hybrid deep learning framework that innovatively combines graph structural representation with sequential modeling. The methodology begins with a heterogeneous graph construction from a Range-Doppler spectrum, where potential signal peaks across different range gates are treated as nodes characterized by ten physical features, with two distinct edge types defined to enhance node connectivity. A Graph Attention Network (GAT) is then employed to extract deep features, enhanced by an intra-layer aggregation mechanism that effectively captures the coherence among signal peaks at different locations. A Transformer model subsequently processes these extracted features to obtain the wind profile, incorporating a redesigned height-aware position encoding to obtain sequential dependencies. Furthermore, a magnitude-weighted loss is designed to address the label imbalance, and a semi-supervised strategy is implemented to leverage the abundant unlabeled data. Crucially, physical constraints including wind shear thresholds, opposing beam parallel training and wind vector synthesis principles are embedded to guide the model toward physically consistent solutions. Validated on Atmospheric Radiation Measurement (ARM) dataset, the proposed model demonstrates accurate tracking of wind profile trends and shows strong agreement with radiosonde data, achieving correlation coefficients of 0.928 and 0.922 for zonal and meridional winds across all sites, respectively. It maintains exceptional accuracy across various SNR conditions, outperforming existing algorithms. Comparisons with other sequence prediction models further confirm its superior generalization capability and robustness. In summary, the GAT-Transformer hybrid model provides a more reliable solution for atmospheric remote sensing applications.
1 December 2025 in IEEE Transactions on Geoscience and Remote Sensing
The planetary boundary layer (PBL) height (PBLH), as a critical meteorological parameter, plays a pivotal role in atmospheric dynamic, climate, and pollution diffusion studies. Radar wind profiler (RWP) serves as one of the instruments for PBLH detection. However, conventional RWP-based algorithms struggle to adapt to diverse complex atmospheric conditions caused by temporal and geographic variability. This study proposes a hybrid CNN-LGBM machine learning (ML) model for PBLH estimation using RWP observational data. The model employs a convolutional neural network (CNN) to extract depth-resolved features from six vertical profiles of RWP measurements, while constructing extra temporal-height-spatial (THS) features. These features are fused into a high-dimensional representation to enhance PBLH characterization. A Light gradient-boosting machine (LGBM) regressor, as another contribution, is then applied to train the fused features, with Label Distribution Smoothing (LDS) and Feature Distribution Smoothing (FDS) strategies integrated to mitigate data imbalance. Evaluated on RWP datasets from ten atmospheric radiation measurement (ARM) sites, CNN-LGBM demonstrates robust performance across varied PBL conditions, totally achieving a mean absolute error (MAE) of 0.23, root mean square error (RMSE) of 0.35, and correlation coefficient (R) of 0.82, with good consistency against radiosonde (RS) measurements. Selection and analysis of different features reveal the positive contribution and interpretability of THS features to regression outcomes. Comparative experiments confirm the model’s superior stability and generalizability over existing RWP-based remote sensing algorithms in diverse seasonal and regional conditions, outperforming other PBLH retrieval instruments in accuracy under different cloud cover conditions. The LGBM framework exhibits optimal regression efficiency and stability compared to other advanced ML models, while LDS and FDS strategies partially alleviate label distribution skewness. In summary, the proposed CNN-LGBM hybrid model improves RWP-based PBLH estimation accuracy, advancing the application of ML in boundary layer meteorology.