EP-4737951-A1 - PRECIPITATION DOWNSCALING WITH LIMITED GROUND-OBSERVATION DATA

EP4737951A1EP 4737951 A1EP4737951 A1EP 4737951A1EP-4737951-A1

Abstract

A method to train a diffusion model for satellite observation precipitation data downscaling may include obtaining high-resolution (HR) ground observation precipitation data that has a first resolution. The method may include obtaining corresponding low-resolution (LR) satellite observation precipitation data that has a second resolution lower than the first resolution. The method may include upsampling the LR satellite observation precipitation data that has the second resolution to generate upsampled satellite observation precipitation data that has the first resolution. The method may include generating training residuals from the HR ground observation precipitation data and the corresponding upsampled satellite observation precipitation data. The method may include adding noise to the training residuals. The method may include de-noising the noisy training residuals using a diffusion model to generate predicted residuals. The method may include updating the diffusion model using a loss function that depends on the training residuals and the predicted residuals.

Inventors

USHIJIMA-MWESIGWA, Hayato
WONG, Hon Yung
DAI, Ting-yu

Assignees

FUJITSU LIMITED

Dates

Publication Date: 20260506
Application Date: 20251027

Claims (20)

A method to train a diffusion model for satellite observation precipitation data downscaling, the method comprising: obtaining high-resolution (HR) ground observation precipitation data that has a first resolution; obtaining corresponding low-resolution (LR) satellite observation precipitation data that has a second resolution lower than the first resolution; upsampling the LR satellite observation precipitation data that has the second resolution to generate upsampled satellite observation precipitation data that has the first resolution; generating training residuals from the HR ground observation precipitation data and the corresponding upsampled satellite observation precipitation data; adding noise to the training residuals; de-noising the noisy training residuals using a diffusion model to generate predicted residuals; and updating the diffusion model using a loss function that depends on the training residuals and the predicted residuals.
The method of claim 1, wherein the diffusion model comprises an equivariant diffusion model (EDM).
The method of claim 1, wherein updating the diffusion model using the loss function comprises backpropagating loss using a weighted mean square error (MSE) loss function.
The method of claim 3, wherein backpropagating loss comprises, for each training residual and corresponding predicted residual: based on the training residual and the corresponding predicted residual, calculating an error value for the corresponding predicted residual using the weighted MSE loss function; and updating learnable parameters of the diffusion model based on the weighted MSE loss function.
The method of claim 3, wherein: the weighted MSE loss function, weightedMSE, comprises: weightedMSE = 1 n ∑ i = 1 n α i y i − y ι ^ ∧ 2 n is a number of predicted values in a given predicted residual; α is a weighting factor; i is an index from 1 to n; y is a given observed value in a given training residual that corresponds to the given predicted residual; and ŷ is a given predicted value in the given predicted residual that corresponds to the given observed value y .
The method of claim 1, wherein generating the training residuals based on the HR ground observation precipitation data and the corresponding upsampled satellite observation precipitation data comprises, for each training residual, subtracting the upsampled satellite observation precipitation data for a given location and time from the HR ground observation precipitation data for the given location and time.
The method of claim 1, wherein the HR ground observation precipitation data and the LR satellite observation precipitation data used to train the diffusion model correspond to a training geographic region, the method further comprising downscaling satellite observation precipitation data corresponding to a target geographic region that is different from the training geographic region using the trained diffusion model.
The method of claim 7, wherein downscaling the satellite observation precipitation data comprises: inputting noisy combined upsampled precipitation data of the target geographic region into the diffusion model to generate a predicted downscaled residual, the noisy combined upsampled precipitation data generated from LR satellite observation precipitation data for the target geographic region and limited ground observation precipitation data for the target geographic region; and combining the predicted downscaled residual with the upsampled combined precipitation data to produce a HR precipitation estimate for the target geographic region.
The method of claim 8, wherein the LR satellite observation precipitation data for the target geographic region comprises bias-corrected LR satellite observation precipitation data for the target geographic region, the method further comprising, prior to inputting the noisy combined upsampled precipitation data of the target geographic region into the diffusion model, generating the bias-corrected LR satellite observation precipitation data for the target geographic region, including: inputting non-bias-corrected LR satellite observation precipitation data for the target geographic region into a bias correction diffusion model to generate a predicted bias correction residual; and combining the predicted bias correction residual with the non-bias-corrected LR satellite observation precipitation data for the target geographic region to produce the bias-corrected LR satellite observation precipitation data for the target geographic region.
A non-transitory computer-readable storage medium comprising computer-readable instructions that are executable by a processor to perform or control performance of the method of claim 1.
A method of downscaling satellite observation precipitation data, comprising: obtaining low-resolution (LR) satellite observation precipitation data for a target geographic region; obtaining limited ground observation precipitation data for the target geographic region; generating combined upsampled precipitation data from the LR satellite observation precipitation data and the limited ground observation precipitation data; adding noise to the combined upsampled precipitation data to generate noisy combined upsampled precipitation data; inputting the noisy combined upsampled precipitation data into a diffusion model to generate a predicted residual; and combining the predicted residual with the combined upsampled precipitation data to generate a high-resolution (HR) precipitation estimate for the target geographic region.
The method of claim 11, wherein generating the combined upsampled precipitation data from the LR satellite observation precipitation data and the limited ground observation precipitation data comprises upsampling the LR satellite observation precipitation data and applying an inverse distance weighting algorithm to combine the limited ground observation precipitation data with the upsampled LR satellite observation precipitation data.
The method of claim 11, wherein the diffusion model comprises an equivariant diffusion model (EDM).
The method of claim 11, further comprising training the diffusion model, including: obtaining HR ground observation precipitation data and corresponding LR satellite observation precipitation data for a training geographic region that is different than the target geographic region; upsampling the LR satellite observation precipitation data to generate upsampled satellite observation precipitation data, the upsampled satellite observation precipitation data having a same resolution as the corresponding HR ground observation precipitation data; generating training residuals from the HR ground observation precipitation data and the LR satellite observation precipitation data; adding noise to the training residuals; and training the diffusion model to denoise the noisy training residuals.
The method of claim 11, wherein the LR satellite observation precipitation data comprises bias-corrected LR satellite observation precipitation data, the method further comprising, prior to generating the combined upsampled precipitation data, generating the bias-corrected LR satellite observation precipitation data, including: inputting non-bias-corrected LR satellite observation precipitation data into a bias correction diffusion model to generate a bias correction residual; and combining the bias correction residual with the non-bias-corrected LR satellite observation precipitation data to produce the bias-corrected LR satellite observation precipitation data.
The method of claim 15, further comprising, prior to generating the bias-corrected LR satellite observation precipitation data, training the bias correction diffusion model, including: training the bias correction diffusion model using past satellite observation precipitation data and past ground observation precipitation data for a training geographic region that is different than the target geographic region; and fine-tuning the bias correction diffusion model using past satellite observation precipitation data and past limited ground observation precipitation data for the target geographic region.
The method of claim 11, wherein the diffusion model comprises a U-Net architecture.
The method of claim 11, wherein generating the combined upsampled precipitation data from the LR satellite observation precipitation data and the limited ground observation precipitation data comprises: upsampling the LR satellite observation precipitation data from 10 kilometer (km) resolution to 1 km resolution; and combining the upsampled LR satellite observation precipitation data and the limited ground observation precipitation data.
The method of claim 11, wherein the limited ground observation precipitation data for the target geographic region comprises rain gauge precipitation data from one or more rain gauges within the target geographic region.
A non-transitory computer-readable storage medium comprising computer-readable instructions that are executable by a processor to perform or control performance of the method of claim 11.

Description

FIELD The present disclosure generally relates to precipitation downscaling with limited ground-observation data. BACKGROUND Accurate and timely precipitation data is essential for many applications, including early warning systems for natural disasters, local water management, and agricultural planning. However, ground-based weather observation equipment like weather surveillance radar (WSR) is often lacking in remote or developing areas due to its high cost. For example, while the United States and Europe combined have over 600 WSRs, Africa has fewer than 40 despite having almost one and a half times their landmass. The subject matter claimed in the present disclosure may not be limited to embodiments that solve any disadvantages or that operate only in environments such as those described above. Rather, this background may be only provided to illustrate one example technology area where some embodiments described in the present disclosure may be practiced. SUMMARY In an example embodiment, a method to train a diffusion model for satellite observation precipitation data downscaling may include obtaining high-resolution (HR) ground observation precipitation data that has a first resolution. The method may include obtaining corresponding low-resolution (LR) satellite observation precipitation data that has a second resolution lower than the first resolution. The method may include upsampling the LR satellite observation precipitation data that has the second resolution to generate upsampled satellite observation precipitation data that has the first resolution. The method may include generating training residuals from the HR ground observation precipitation data and the corresponding upsampled satellite observation precipitation data. The method may include adding noise to the training residuals. The method may include de-noising the noisy training residuals using a diffusion model to generate predicted residuals. The method may include updating the diffusion model using a loss function that depends on the training residuals and the predicted residuals. The object and advantages of the embodiments will be realized and achieved at least by the elements, features, and combinations particularly pointed out in the claims. It is to be understood that both the foregoing general description and the following detailed description are explanatory and are not restrictive of the invention, as claimed. BRIEF DESCRIPTION OF THE DRAWINGS Example embodiments will be described and explained with additional specificity and detail through the accompanying drawings in which: FIG. 1 illustrates an example system 100 for weather data processing;FIG. 2A illustrates an example process flow to train a diffusion model for satellite data downscaling;FIG. 2B illustrates a flowchart of an example method to train the diffusion model for satellite data downscaling;FIG. 3A illustrates an example process flow to downscale satellite data;FIG. 3B illustrates a flowchart of an example method to downscale satellite data;FIG. 4A illustrates an example process flow to train a diffusion regression model for satellite data bias correction;FIG. 4B illustrates an example process flow to fine-tune the pretrained diffusion regression model of FIG. 4A for satellite data bias correction;FIG. 4C illustrates a flowchart of an example method to train the diffusion regression model for satellite data bias correction;FIG. 5A illustrates an example process flow to bias correct satellite data;FIG. 5B illustrates a flowchart of an example method to bias correct satellite data;FIG. 6 compares error distributions of original and corrected IMERG data relative to LR MRMS data;FIG. 7 compares error distributions of LR MRMS and downscaled MRMS relative to HR MRMS;FIG. 8 illustrates results of applying bias correction and downscale models sequentially to downscale IMERG data to the resolution of MRMS observations, showcasing detailed cloud-like formations and precipitation patterns;FIG. 9 compares error distributions of IMERG, bias corrected IMERG, and downscaled IMERG relative to MRMS data;FIG. 10 shows results of application of a downscale model trained using data from Seattle to data from, respectively, New York, San Jose, and Portland;FIG. 11 shows results of application of a bias correction model trained using data from Seattle to data from, respectively, New York, San Jose, and Portland; andFIG. 12 illustrates a block diagram of an example computing system, in accordance with one or more embodiments of the present disclosure. DETAILED DESCRIPTION Water-related disasters, such as landslides, floods, and droughts, may constitute a significant majority of natural disasters. The past five decades have witnessed over 11,000 reported weather-related disasters globally, tragically claiming over two million lives and inflicting $3.64 trillion in economic losses. Developing countries disproportionately bear this burden, accounting for over 91% of these fatalities. Howev