DDM-Former: Transformer networks for GNSS reflectometry global ocean wind speed estimation

Zhao, Daixin; Heidler, Konrad; Asgarimehr, Milad; Arnold, Caroline; Xiao, Tianqi; Wickert, Jens; Zhu, Xiao Xiang; Mou, Lichao

doi:10.1016/j.rse.2023.113629

Cited by 10 publications

(3 citation statements)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It must be noted that the specific hyperparameters in each network, including the number of hidden layers and neurons in these layers, the unit size, and the dropout rate, were determined through trial and error, based on achieving the best performance on the validation dataset. This was the only viable approach and is also common in deep learning model development studies, including those for GNSS-R wind speed retrievals [11][12][13][14][15].…”

Section: Model Architecturesmentioning

confidence: 99%

See 1 more Smart Citation

Hybrid CNN-LSTM Deep Learning for Track-Wise GNSS-R Ocean Wind Speed Retrieval

Arabi,

Asgarimehr,

Kada

et al. 2023

Remote Sensing

Self Cite

View full text Add to dashboard Cite

The NASA Cyclone GNSS (CYGNSS) mission provides one Delay Doppler Map (DDM) per second along observational tracks. To account for spatiotemporal correlations within adjacent DDMs in a track, a deep hybrid CNN-LSTM model is proposed for wind speed prediction. The model combines convolutional and pooling layers to extract features from DDMs in one track, which are then processed by LSTM as a sequence of data. This method leads to a test RMSE of 1.84 m/s. The track-wise processing approach outperforms the architectures that process the DMMs individually, namely based on Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), and a network based solely on fully connected (FC) layers, as well as the official retrieval algorithm of the CYGNSS mission with RMSEs of 1.92 m/s, 1.92 m/s, 1.93 m/s, and 1.90 m/s respectively. Expanding on the CNN-LSTM model, the CNN-LSTM+ model is proposed with additional FC layers parallel with convolutional and pooling layers to process ancillary data. It achieves a notable reduction in test RMSE to 1.49 m/s, demonstrating successful implementation. This highlights the significant potential of track-wise processing of GNSS-R data, capturing spatiotemporal correlations between DDMs along a track. The hybrid deep learning model processing the data sequentially in one track learns these dependencies effectively, improving the accuracy of wind speed predictions.

show abstract

Section: Model Architecturesmentioning

confidence: 99%

“…Another study reported a CNN-based model with an RMSE of 1.53 m/s [14]. Recently, CYGNSS DDMs have been processed based on transformer networks, resulting an RMSE of 1.43 m/s, and the models offered explainability through attention maps [15].…”

Section: Introductionmentioning

confidence: 99%

Hybrid CNN-LSTM Deep Learning for Track-Wise GNSS-R Ocean Wind Speed Retrieval

Arabi,

Asgarimehr,

Kada

et al. 2023

Remote Sensing

Self Cite

View full text Add to dashboard Cite

show abstract

“…This results in more than 99% of the data available from the delay-Doppler map (DDM) shown in Figure 2 being discarded due to the difficulty in incorporating them into traditional retrieval algorithms. There have been studies that propose to use the full DDM to extract additional information using deep learning models for soil moisture estimation [35] and wind speed retrieval [36]. Before moving further, it is worth providing more details and characteristics, motivating the use of the full DDM for biomass estimation.…”

Section: Introductionmentioning

confidence: 99%

Biomass Estimation with GNSS Reflectometry Using a Deep Learning Retrieval Model

Pilikos,

Clarizia,

Floury

2024

Remote Sensing

View full text Add to dashboard Cite

GNSS Reflectometry (GNSS-R) is an emerging technique for the remote sensing of the environment. Traditional GNSS-R bio-geophysical parameter retrieval algorithms and deep learning models utilize observables derived from only the peak power of the delay-Doppler maps (DDMs), discarding the rest. This reduces the data available, which potentially hinders estimation accuracy. In addition, reflections from water bodies dominate the signal amplitude, and using only the peak power in those areas is challenging. Motivated by all the above, we propose a novel deep learning retrieval model for biomass estimation that uses the full DDM of surface reflectivity. Experiments using CYGNSS data have illustrated the improvements achieved when using the full DDM of surface reflectivity. Our proposed model was able to estimate biomass, trained using the ESA Climate Change Initiative (CCI) biomass map, outperforming the model that used peak reflectivity. Global and regional analysis is provided along with an illustration of how biomass estimation is achieved when using the full DDM around water bodies. GNSS-R could become an efficient method for biomass monitoring with fast revisit times. However, an elaborate calibration is necessary for the retrieval models, to associate GNSS-R data with bio-geophysical parameters on the ground. To achieve this, further developments with improved training data are required, as well as work using in situ validation data. Nevertheless, using GNSS-R and deep learning retrieval models has the potential to enable fast and persistent biomass monitoring and help us better understand our changing climate.

show abstract