Snowpack Estimation in Key Mountainous Water Basins from Openly-Available, Multimodal Data Sources

Malachy Moran, Kayla Woputz, Derrick Hee, Manuela Girotto, Paolo D'Odorico, Ritwik Gupta, Daniel Feldman, Puya Vahabi, Alberto Todeschini, Colorado J. ReedarXiv 2022 · DOI · OpenReview

TL;DR

This paper presents a novel multimodal deep learning model that fuses openly available satellite and weather data to estimate snow water equivalent (SWE) in mountainous regions. By overcoming the limitations of expensive, sparse, and biased traditional measurement methods, their approach achieves a significant improvement in accuracy, outperforming single-source estimations and even sparse in-situ measurements.

What Problem Does This Paper Address?

Accurate estimation of snowpack, specifically Snow Water Equivalent (SWE), in mountainous regions is crucial for water resource management, impacting everything from agriculture and power generation to ecosystems and public policy. However, current methods for obtaining this data are problematic. In-situ measurements (like SNOTEL stations) are sparse, expensive to maintain, and biased towards accessible areas, failing to capture the variability in remote, high-altitude regions. Airborne LiDAR surveys (like ASO) offer more comprehensive coverage but are prohibitively expensive, infrequent, and geographically limited. This creates a critical gap in our ability to monitor and predict water availability from snowmelt, a problem exacerbated by climate change.

Methods

The authors propose a multimodal deep learning architecture that integrates various data sources. The core idea is to leverage the strengths of different data types by processing them with specialized neural network components before fusing their representations.

  • Satellite Imagery: Images from Copernicus, Sentinel 1, Sentinel 2, and MODIS are fed into Convolutional Neural Networks (CNNs). A smaller "Mixnet" is used for the static Copernicus DEM, while larger "EfficientNet_B2s" are employed for the time-varying satellite data.
  • Time Series Weather Data: GridMET weather data (precipitation, temperature, wind) is processed by a Long Short-Term Memory (LSTM) network to capture temporal dependencies.
  • Fusion: The encoded representations from the CNNs and LSTM are concatenated into a single vector. This fused vector is then passed through a 3-layer Multi-Layer Perceptron (MLP) to directly predict the SWE for each 1km² cell.
  • Post-processing: A Gaussian smoothing filter is applied to the predicted output to reduce errors along cell boundaries, a common issue in gridded predictions.

Datasets

  • Training Data:
    • Satellite Imagery: Copernicus DEM GLO 30, Sentinel-1 SAR GRD, Sentinel-2 MSI, MODIS Terra/Aqua Snow Cover Daily Global 500m.
  • Weather Data: gridMET Gridded Surface Meteorological Dataset.
  • Ground Truth (Target Variable): Airborne Snow Observatory (ASO) Snow Water Equivalent (SWE) measurements.
  • Evaluation: The model was trained on data from 2016-2019 and tested on recent ASO flyovers from March-April 2022.

The study focuses on nine basins in the Sierra Nevada region, covering approximately 28,750 km². The authors do not introduce a new benchmark dataset but rather leverage existing, openly available sources and ASO data as ground truth.

Key Insights and Ideas

  • Multimodal Fusion is Key: The paper strongly suggests that combining diverse data sources (satellite imagery, weather data) is essential for accurate SWE estimation, significantly outperforming any single source.
  • Addressing Data Sparsity and Bias: The model effectively overcomes the limitations of sparse and biased in-situ measurements by leveraging the broader coverage of satellite data.
  • Deep Learning for Complex Terrain: The use of CNNs and LSTMs allows the model to learn complex spatial and temporal patterns inherent in mountainous terrain, which are difficult for traditional physical models to capture.
  • Trade-offs in Resolution: The authors discuss the challenges of integrating data at different resolutions, particularly with MODIS, and how they adapted their approach to tabular data to manage computational resources.
  • Predictive Bias: The model exhibits a tendency to underpredict SWE for high values and overpredict for low values, suggesting areas for model refinement.

Key Results

The multimodal fused model achieved a mean RMSE of 7.5 inches across the studied Sierra Nevada basins, weighted by basin area.

  • This represents a significant improvement over single-source estimations, with the multisource model outperforming individual sources by an average of 5.0 inches RMSE.
  • It also outperforms sparse in-situ measurements by 1.2 inches RMSE.
  • The RMSE varied considerably by basin, ranging from 2.7 inches in the Feather basin to 17.5 inches in Kings Canyon, correlating with the model's predictive bias.

Limitations and Future Work

  • Resolution: The current model predicts SWE at a 1km² resolution, which is coarser than the original 50m ASO data. Future work aims to explore higher resolutions.
  • Input Features: The authors suggest incorporating more sophisticated snow albedo inputs and leveraging data from sensors like GEDI for more accurate vegetation coverage and canopy height information.
  • Wind Metrics: The current model treats wind direction and aspect separately. Future work will explore how to better integrate these features, as wind is known to be a significant factor in snow accumulation.
  • Temporal Coverage: Expanding temporal coverage and better capturing seasonal variations are also identified as future research directions.
  • Model Bias: The observed under/overprediction bias needs further investigation and mitigation.

Summary updated

Abstract

Accurately estimating the snowpack in key mountainous basins is critical for water resource managers to make decisions that impact local and global economies, wildlife, and public policy. Currently, this estimation requires multiple LiDAR-equipped plane flights or in situ measurements, both of which are expensive, sparse, and biased towards accessible regions. In this paper, we demonstrate that fusing spatial and temporal information from multiple, openly-available satellite and weather data sources enables estimation of snowpack in key mountainous regions. Our multisource model outperforms single-source estimation by 5.0 inches RMSE, as well as outperforms sparse in situ measurements by 1.2 inches RMSE.
BibTeX
@misc{moran2022snowpack,
  author = {Malachy Moran and Kayla Woputz and Derrick Hee and Manuela Girotto and Paolo D'Odorico and Ritwik Gupta and Daniel Feldman and Puya Vahabi and Alberto Todeschini and Colorado J. Reed},
  title = {{Snowpack Estimation in Key Mountainous Water Basins from Openly-Available, Multimodal Data Sources}},
  year = {2022},
  eprint = {2208.04246},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.04246},
}
Figure