Due to the limitations of satellite orbit coverage and soil moisture retrieval modes, satellite based daily soil moisture products inevitably have the disadvantage of low global land coverage. The SGD-SM 2.0 dataset uses three sensors, namely AMSR-E, AMSR2, and WindSat. The global daily precipitation product is integrated into the proposed reconstruction model. An integrated Long Short Term Memory Convolutional Neural Network (LSTM-CNN) was proposed to fill in the gaps and missing areas in daily soil moisture products. In situ validation and time series validation have demonstrated the reconstruction accuracy and usability of SGD-SM 2.0 (R: 0.672, RMSE: 0.096, MAE: 0.078). The time series curve of the improved SGD-SM 2.0 is consistent with the original daily time series distribution of soil moisture and precipitation. Compared with SGD-SM 1.0, the improved SGD-SM 2.0 has significant advantages in reconstruction accuracy and time series consistency.
In this dataset, satellite based soil moisture products and precipitation products were simultaneously integrated to generate the SGD-SM 2.0 dataset. The reconstruction accuracy of SGD-SM 2.0 was validated using in-situ soil moisture sites. These in-situ data were downloaded from the International Soil Moisture Network (ISMN).
From 2002 to 2022, AMSR-E, AMSR2, and WindSat utilized global daily soil moisture products.
The comprehensive multi satellite E inversion product of GPM (IMERG) global daily precipitation V6 was used from 2002 to 2022. These precipitation products come from multiple satellite passive microwave sensors related to precipitation.
The spatial resolution is represented as 0.1∘grid (approximately 10 km), which collects global in-situ surface data at IMERG Level 3 Global Day ISMN. These data have been widely used for hydrological and soil moisture validation. We selected 124 sites from ISMN from 2002 to 2022 and matched them with corresponding soil moisture products in SGD-SM 2.0.
This dataset uses Python 3.7 language, PyCharm platform, and Windows 10 environment to generate seamless global daily soil moisture products.
In terms of hardware configuration, we use NVIDIA Titan X (Pascal) GPU, Inter E5-2609v3 CPU, and 16 GB DDR4 RAM to execute the proposed LSTM-CNN model. Integrate global daily precipitation products and global daily soil moisture products into SGD-SM 2.0 simultaneously. We have developed an integrated Long Short Term Memory Convolutional Neural Network (LSTM-CNN) reconstruction model to fill in the gaps and missing areas in daily soil moisture products worldwide.
Finally, we recursively generate seamless daily soil moisture products in the SGD-SM 2.0 dataset.
From a spatial perspective, the proposed SGD-SM 2.0 dataset exhibits both global soil moisture uniformity and local soil moisture heterogeneity. It ensures spatial consistency, especially for the gap areas with adjacent soil moisture zones. In addition, the reconstructed regions in SGD-SM 2.0 do not reflect significant plaque or boundary effects. This also demonstrates the powerful ability of some CNNs in the proposed framework to effectively eliminate invalid information in areas with gaps or missing soil moisture.
From a temporal perspective, the proposed SGD-SM 2.0 dataset utilizes complementary and continuous time series soil moisture information. By integrating global daily precipitation, SGD-SM 2.0 can consider sporadic extreme weather conditions on a single day. In addition, through the LSTM model, consistent time information can be restored and saved.
This work is licensed under a
CC BY 4.0.
CSTR:11738.11.NCDC.ZENODO.DB6460.2024
10.5281/zenodo.6041561
