Name: PptxGenJS Presentation
Uploaded: 2025-04-28
Duration: 4 min 14 s
Description: Getnet Demil https://getnetdemil.github.io PhDing in Use of Image Processing for Snow Water Characteristics Estimation Under DIWA project.

Getnet Demil https://getnetdemil.github.io PhDing in Use of Image Processing for Snow Water Characteristics Estimation Under DIWA project. 

Scene 1

Getnet Demil https://getnetdemil.github.io
 PhDing in Use of Image Processing for Snow Water Characteristics Estimation
 Under DIWA project

Image Processing for Snow Water Characteristics Estimation. 

Scene 2

Image Processing for Snow Water Characteristics Estimation

Seasonal snowpack serves as a natural reservoir, providing billions of people worldwide with essential spring runoff for irrigation, hydropower, and drinking water. It significantly influences river flows and flood risk, while also impacting climate by reflecting sunlight and affecting building safety through its weight.

The Challenge of Snow Monitoring. Point Data Limitations. 

Scene 3

Traditional measurements like manual snow stakes or ultrasonic sensors provide only localized data points, failing to reveal variations across slopes or regions.

Manual surveying is labor-intensive and slow, providing only snapshots rather than continuous data.

Networks like SNOTEL provide valuable time series but are spatially sparse, leaving large uncertainties in mountainous regions and over broad basins.

Multi-Platform Approach. Satellite Imagery. Covers continent-scale areas with frequent revisits (daily to weekly) but at coarser resolution.. 

Scene 4

Covers continent-scale areas with frequent revisits (daily to weekly) but at coarser resolution.

Small UAVs flying a few hundred meters high can map snow with centimeter accuracy over limited areas.

Planes carrying cameras or LiDAR produce high-resolution snow depth maps at centimeter to decimeter scale.

Our approach leverages remote sensing from multiple platforms to capture snow information over large areas at high resolution, addressing the classic trade-off between spatial coverage and resolution.

Data Processing Pipeline. Data Curation. Clean noise, co-register maps, and standardize different datasets into common coordinates and units.. 

Scene 5

Clean noise, co-register maps, and standardize different datasets into common coordinates and units.

Apply deep-learning-based super-resolution algorithms to up-sample coarse satellite data to finer scales.

Deploy different neural networks to estimate snow variables from enhanced imagery.

Use Ensemble Kalman Filtering to merge image-derived estimates with snow model forecasts.

Super-Resolution Techniques. Single-Image Super-Resolution. 

Scene 6

SISR networks are trained to recover high-resolution images from low-resolution inputs, effectively sharpening satellite imagery by learning from high-res and low-res pairs.

These convolutional neural networks and generative models (e.g., S2GAN) have been shown to effectively enhance Sentinel-2 or Landsat images.

By applying these techniques, we can generate enhanced snow cover and depth maps with improved spatial detail, overcoming the resolution limitations of satellite data.

Low Resolution                               High Resolution

The utilization of different satellite data such as Sentinel 1 and Sentinel 2 data, to leverage the advantage of both modalities

Deep Learning for Snow Variable Estimation. Snow Cover Extent. 

Scene 7

Deep Learning for Snow Variable Estimation

Deep learning architectures segment snow vs. non-snow

Regression networks estimate continuous depth fields

Our pipeline is sequential: first detecting snow-covered areas, then estimating depth within that extent, and finally converting depth to SWE using density models. Recent work by Duan et al. demonstrated that deep learning models can achieve higher accuracy than traditional empirical models by learning physical relationships from precipitation and temperature.

Data Assimilation Framework. Image Acquisition. Collect multi-sensor imagery from satellites and UAVs. 

Scene 8

Collect multi-sensor imagery from satellites and UAVs

Generate snow forecasts from hydrological models

Optimally combine observations with model predictions

We plan to use Ensemble Kalman Filtering to merge our image-derived estimates with snow model forecasts. This approach has been successful in snow hydrology, as demonstrated by Lee et al. who used a local EnKF to optimally combine satellite SWE retrievals with land-surface model predictions.

Anticipated Outcomes. Super-Resolution Algorithms. 

Scene 9

Novel up-sampling tools for snow image data, improving spatial/temporal detail beyond raw satellite resolutions.

A systematic software framework and datasets for AI-driven snow mapping, made publicly available for researchers and practitioners.

More accurate, high-resolution snow maps allow reservoir operators and planners to better prepare for spring runoff and extreme events.

New insights into how surface texture or snowpack dynamics influence depth predictions through explainable AI analyses.

Conclusion and Future Work. Publication. Advances in image-based estimation of snow variables: A systematic literature review on recent studies https://doi.org/10.1016/j.jhydrol.2025.132855. 

Scene 10

Advances in image-based estimation of snow variables: A systematic literature review on recent studies https://doi.org/10.1016/j.jhydrol.2025.132855

Apply techniques to broader water management challenges

By combining cutting-edge image processing and data assimilation, this project aims to transform snow measurement. Our approach addresses the limitations of conventional methods by providing comprehensive, high-resolution snow information critical for water resources planning, flood forecasting, and climate impact assessment.

In the near future, we will publish our methods and release all code and data as open-source, enabling wider adoption and further advancement of these techniques.

Scene 11

PptxGenJS Presentation

 Getnet Demil https://getnetdemil.github.io
 PhDing in Use of Image Processing for Snow Water Characteristics Estimation
 Under DIWA project


PptxGenJS Presentation

Scene 1 (0s)

Scene 2 (12s)

Scene 3 (29s)

Scene 4 (51s)

Scene 5 (1m 17s)

Scene 6 (1m 41s)

Scene 7 (2m 12s)

Scene 8 (2m 39s)

Scene 9 (3m 6s)

Scene 10 (3m 32s)

Scene 11 (4m 7s)