- The paper introduces an LSTM approach that extends SMAP soil moisture data over the CONUS with high fidelity.
- It integrates atmospheric forcing, LSM-simulated moisture, and static attributes, achieving RMSE < 0.035 and correlation > 0.87 over 75% of the region.
- The study shows LSTM's superiority over conventional methods by significantly reducing bias and enhancing model generalization across diverse climates.
Prolongation of SMAP to Spatio-temporally Seamless Coverage of Continental US Using a Deep Learning Neural Network
This paper explores the use of Long Short-Term Memory (LSTM) networks in hydrological modeling to extend the Soil Moisture Active Passive (SMAP) mission's soil moisture data over the Continental United States (CONUS). SMAP, since its launch in 2015, has provided invaluable insights into soil moisture, a crucial variable influencing numerous hydrological and meteorological processes. However, it faces limitations due to its short operational timespan and irregular revisit pattern. To address these limitations, the authors develop a methodology leveraging LSTM, a robust deep learning architecture.
The paper successfully implements an LSTM network that uses atmospheric forcing, model-simulated soil moisture from land surface models (LSMs), and static physiographic attributes to predict SMAP soil moisture data. The results demonstrate a substantial reduction in bias and improved accuracy of soil moisture climatology predictions compared to traditional LSMs. Notably, the LSTM achieves a root-mean-squared error (RMSE) of less than 0.035 and a correlation coefficient greater than 0.87 for over 75% of the CONUS. These metrics underscore LSTM's effectiveness in generating high-fidelity soil moisture datasets compatible with SMAP observations.
The authors provide a detailed comparison between LSTM and several conventional algorithms, including linear regression (LR), auto-regressive models (AR), and a one-layer feedforward neural network (NN). Among these, LSTM shows the most favorable performance, characterized by a robust generalization capability across differing climatic and physiographic regions. This generalization stems from its recurrent structure, which allows the network to learn complex temporal dependencies, further supplemented by feature-rich data inputs.
LSTM's potential in this domain extends beyond merely mimicking SMAP data, as it also addresses the significant systematic biases inherent in LSM simulations. This capability is particularly illustrated in regions like the eastern United States, where traditional models often under-estimate soil moisture due to limitations in pedo-transfer functions and the representation of sub-surface hydrological processes.
The research implies vital practical applications in hydrology, allowing for improved long-term soil moisture hindcasting and possibly enhancing weather prediction models and strategies for data assimilation. Furthermore, the paper opens avenues for LSTM's application to other geophysical domains, given the model's ability to learn from data attributes not explicitly characterized in physics-based models.
Future work could focus on further refinement of LSTM models through the incorporation of additional static and dynamic attributes or through optimizing network architectures. The potential expansion of this methodology to global scales or its integration with other remote sensing datasets could also merit exploration. The current paper does not extend to evaluating the impact of LSTM's predictions during extreme hydrological events or its comparison with region-specific simpler models, avenues that could benefit from dedicated future research.
In conclusion, the paper presents compelling evidence for LSTM as a powerful tool in hydrological modeling, capable of generating high-fidelity soil moisture datasets that can supplement and extend the operational capabilities of satellite missions like SMAP.