Reanalysis Data: A Deep Dive into Soil Moisture and Atmospheric Predictions

Understanding Reanalysis Data

Reanalysis data plays a fundamental role in atmospheric sciences, providing insights into soil moisture, weather patterns, and climate variability. Our study focuses on root-zone soil moisture (RZSM), utilizing extensive datasets collected from innovative modeling techniques.

Soil Moisture Collection

Data Sources: We obtained daily soil moisture data spanning from December 1999 to February 2020, specifically for two depths: the surface (0–10 cm) and rooting zone (10–100 cm). Our primary resource was the Global Land Evapotranspiration Amsterdam Model (GLEAM v3.8a), which uses satellite observations to estimate soil moisture alongside terrestrial evaporation.

Calculating RZSM: For our analysis, we computed a weighted average for RZSM of the uppermost 100 cm of soil. This calculation is crucial as it reflects the moisture availability to plants. Using a predefined scaling factor by depth, the formula employed is:

[
RZSM_{weighted} = \frac{surface \cdot 10\,cm}{100\,cm} + \frac{rooting\ zone \cdot 90\,cm}{100\,cm}
]

GLEAM aggregates various algorithms aimed at estimating hydraulic fluxes using satellite data from the Soil Moisture and Ocean Salinity (SMOS) and the European Space Agency (ESA) Climate Change Initiative.

Data Processing: Raw data goes through several preprocessing steps, including bilinear interpolation for resolution refinement and seasonal anomaly calculation. Seasonal means from a reference period (2000–2015) are subtracted from daily observations to isolate anomalies, which can reveal patterns critical for forecasting.

Atmospheric Conditions and High-Resolution Forecasting

To complement soil moisture data, we also retrieved hourly atmospheric variables, including:

• 2 m dewpoint temperature
• 2 m maximum and minimum temperatures ((T{max}) and (T{min}))
• Precipitable water (pwat)
• Surface pressure
• 200 hPa geopotential height ((z_{200}))

This data originated from the fifth generation of European ReAnalysis (ERA5), spanning the same time frame, providing highly detailed atmospheric profiles.

Temperature Differences: We calculated the difference between maximum and minimum temperatures, termed (\Delta_{tmp}). This variable helps in understanding the effects of cloud cover and radiation processes on temperature variations, which could influence soil moisture dynamics.

Geographic Focus Areas

For clarity and relevance, the data subset focuses on specific regions such as the contiguous United States (CONUS), China, and Australia. Each region’s data was filtered according to specific latitude and longitude bounding boxes, ensuring localized analyses.

Retrospective Forecasting and Model Utilization

Dynamic Forecast Models: We utilized the Global Ensemble Forecast version 12 (GEFSv12) and subseasonal-to-seasonal (S2S) modeling from the ECMWF for retrospective forecasting, assessing weekly predictions up to a lead time of five weeks.

Model Characteristics: GEFSv12 is notable for its incorporation of a global 11-member ensemble, initialized with previously established conditions. The coupling of the Global Forecast System version 15.1 and the Noah Land Surface Model enables the system to produce high-accuracy predictions.

Anomaly Calculation: Similarly to the RZSM processing techniques, we computed seasonal anomalies from dynamic model forecasts. This ensures consistency in evaluating performance across both soil moisture and atmospheric inputs.

Data Preprocessing Procedures

Preprocessing involves systematic feature scaling to enhance model performance. Each observational variable is restructured from a time series into a multi-dimensional format, allowing for effective incorporation into machine learning models.

Normalization Techniques: We applied min-max normalization across various predictors to ensure comparability, yielding values bounded between 0 and 1. Areas with no recorded values, such as oceans, are masked with zeros, allowing for more accurate model training and validation.

Deep Learning Architecture: The RISE-UNet Model

Model Design: At the heart of our analytical approach is the RISE-UNet architecture, an advanced convolutional neural network structure tailored for weather and climate forecasting. This nested model with its unique skip connections preserves intricate information through various processing layers.

Inception and Squeeze-and-Excitation Mechanisms: The incorporation of multiple filters and advanced activation functions allows the model to better capture complex relationships across spatial variables, leading to improved forecasting accuracy.

Ensemble Forecasting and Statistical Evaluation

Loss Functions and Ensemble Skills: To quantify model performance, we utilize various metrics. The Continuous Ranked Probability Score (CRPS) serves as our primary loss function, measuring discrepancies between ensemble predictions and actual observations.

Evaluative Metrics: We also employ the Continuous Ranked Probability Skill Score (CRPSS) to determine the relative skill of our ensemble forecasts compared to raw predictions. The Anomaly Correlation Coefficient (ACC) further assesses correlation strength, aiding in identifying forecast value.

Conclusion

The intricate interplay of soil moisture dynamics and atmospheric conditions captured through advanced modeling techniques provides a holistic view of climate forecasting capabilities. This analysis, driven by robust data processing and innovative machine learning architectures, stands at the forefront of enhancing predictability in response to changing climatic conditions. As we advance, the continued integration of reanalysis data and dynamic forecasting models can yield critical insights for agriculture, hydrology, and broader environmental sciences.