📄 earth_science

Machine Learning downscaling of ERA5 temperature and precipitation in the Central Highlands of Vietnam using topographic, hydrometeorological, and climate indices

This study demonstrates that a machine learning framework incorporating engineered hydrometeorological and climate indices, such as the Orographic Index and Vapor Pressure Deficit, significantly improves the downscaling of coarse ERA5 temperature and precipitation data to station-scale resolution in Vietnam's Central Highlands, thereby providing high-fidelity climate information for agricultural and water resource management.

Original authors: Anh Quan Duong, Ngoc Quy Bui

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anh Quan Duong, Ngoc Quy Bui

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake a cake for a specific village in the mountains of Vietnam, but your recipe book only gives you instructions for the entire country. The book says, "It's usually warm and rainy here," but it doesn't know that your village is tucked behind a steep mountain that blocks the rain, or that the valley below it gets scorching hot while the peak stays cool.

This is exactly the problem scientists faced with ERA5, a powerful global weather database. It's like a high-resolution satellite photo that is actually quite blurry when you zoom in on the complex, mountainous Central Highlands of Vietnam (known as Tay Nguyen). Because the mountains are so jagged and the weather so tricky, the "blurry" data often gets the temperature and rainfall wrong, sometimes underestimating rain by a lot or missing local heat spikes.

The Solution: A "Smart Translator"

The researchers in this paper built a Machine Learning "translator" to fix these blurry pictures. Think of it like taking a rough sketch of a landscape and using a super-smart AI to fill in all the missing details based on real-life observations.

They didn't just let the AI guess based on the map's shape (like elevation or slope). Instead, they gave the AI a "cheat sheet" of dynamic weather clues to help it understand why the weather behaves the way it does.

Here are the "clues" they added to the AI's brain:

The Wind-and-Slope Interaction (Orographic Index): Imagine wind hitting a mountain. If it hits the side facing the wind, it gets forced up, cools down, and dumps rain. If it goes over the top and down the other side, it gets hot and dry. The AI learned to calculate exactly how the wind was hitting the specific slope of each village.
The "Thirst" Meter (Vapor Pressure Deficit): This measures how much "dryness" the air has. It's like knowing how thirsty the atmosphere is. If the air is very thirsty, it sucks moisture out of the soil and plants, making the ground hotter. The AI used this to predict temperature spikes during dry seasons.
The "Memory" of Rain (SPI3): The AI was taught to look at the last three months of rain. If the ground has been dry for a while, it behaves differently than if it's been soaking wet. This helped the AI predict future rain better.
The Global Mood Ring (ONI): This tracks the El Niño/La Niña cycles. It's like knowing if the whole Pacific Ocean is in a "grumpy" or "happy" mood, which changes the weather patterns in Vietnam.

The Results: From Blurry to Crystal Clear

The team tested this new "Enhanced AI" against the old, standard way of doing things. They used 36 years of data (1984–2019) from 13 weather stations to teach the AI, and then tested it on years it had never seen before.

For Rain: The old, blurry data was often off by nearly 95 mm of rain per month. The new AI, using the "cheat sheet" of clues, reduced that error significantly. It got much better at predicting how much rain would actually fall, especially in the tricky mountain areas where the old data failed.
For Temperature: The old data was sometimes off by almost 1 degree Celsius. The new AI got the temperature almost perfect, with errors dropping to just half a degree. It successfully figured out why some valleys were hotter and some peaks were cooler, even when the global data missed it.

Why This Matters (According to the Paper)

The paper focuses on the Central Highlands, which is the heart of Vietnam's coffee industry. Coffee plants are very sensitive; they need just the right amount of water and temperature to grow and flower.

Because the old global data was "blurry," farmers and water managers were working with inaccurate information. This new, sharpened data acts like a high-definition map of the local climate. It allows them to:

Plan for Droughts: By understanding exactly how dry the air and soil are, they can better predict when a drought might hit.
Manage Water: They can make smarter decisions about how much water to release from reservoirs for irrigation.
Protect Crops: They can get more accurate warnings about extreme heat or lack of rain, helping to save the coffee harvest.

In short, the researchers took a "one-size-fits-all" global weather map and used smart physics-based clues to turn it into a custom, high-precision weather forecast for the specific villages and farms of Vietnam's Central Highlands.

Technical Summary: Machine Learning Downscaling of ERA5 in Vietnam's Central Highlands

Problem Statement

Coarse-resolution atmospheric reanalysis datasets, such as the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5), are essential for climate research but suffer from significant systematic biases in regions with complex topography. The Central Highlands of Vietnam (Tay Nguyen), a critical agricultural zone producing the majority of the country's coffee exports, is highly vulnerable to dry-season droughts and localized extreme rainfall. However, the region lacks dense, long-term ground-based meteorological station networks. Consequently, researchers often rely on global gridded datasets like ERA5 (~31 km resolution), which fail to capture subgrid-scale physical processes, steep orographic gradients, and localized tropical convective precipitation. These datasets frequently underestimate extreme precipitation events and misrepresent temperature gradients, rendering them insufficient for regional agricultural and hydrological planning.

Methodology

This study proposes a physically informed machine learning (ML) framework to downscale monthly temperature and precipitation from the 0.25° ERA5 grid to station-scale resolution using data from 13 hydrometeorological stations (1984–2019).

Data Sources

Observations: Monthly precipitation and daily mean temperature from 13 stations managed by the Vietnam Meteorological and Hydrological Administration (VNMHA).
Reanalysis: ERA5 monthly data (2m temperature, total precipitation, relative humidity, wind components, solar radiation, evaporation).
Static Features: 30m SRTM Digital Elevation Model (DEM) for elevation, slope, aspect, and distance to the sea; MODIS land cover data.

Feature Engineering

To bridge the gap between static geographic coordinates and dynamic atmospheric processes, the authors engineered several time-varying predictors:

Orographic Index (OI): Quantifies wind-slope interaction ( $Wind Speed \times \sin(Slope) \times \cos(Wind Direction - Aspect)$ ) to represent mechanical uplift (windward) or subsidence (leeward).
Moisture Flux (MF): A proxy for moisture availability ( $Relative Humidity \times Wind Speed$ ).
Vapor Pressure Deficit (VPD): Represents atmospheric drying power, critical for temperature-moisture coupling.
Standardized Precipitation Index (SPI3): A 3-month window metric to capture antecedent soil moisture memory and drought conditions.
Dry Season Length: Cumulative count of months where precipitation is less than potential evapotranspiration.
Oceanic Niño Index (ONI): Categorical representation of ENSO phases (El Niño, La Niña, Neutral) to capture interannual climate oscillations.

Model Architecture and Validation

Five regression algorithms were evaluated: Random Forest (RF), XGBoost, LightGBM, CatBoost, and a Stacked Ensemble (combining RF, CatBoost, and XGBoost via Ridge regression).

Configurations: Two setups were compared:
- Standard: Static topographic features + raw ERA5 variables.
- Enhanced: Standard features + engineered meteorological indices (OI, MF, VPD, SPI3, DryLength, ONI).
Validation: A TimeSeriesSplit approach was used, training on 1984–2005 and testing on 2006–2019 to ensure generalization to unseen temporal windows.
Metrics: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Coefficient of Determination ( $R^2$ ), and Kling-Gupta Efficiency (KGE).

Key Results

Baseline Bias

Raw ERA5 data exhibited systematic errors:

Precipitation: Underestimated monthly rainfall by an overall mean of 13.58 mm (RMSE: 94.76 mm). Errors were particularly severe in high-altitude areas (e.g., underestimation of 44.56 mm in Bao Loc) and overestimation in others (e.g., Da Lat).
Temperature: Overall MAE of 0.78°C, with biases varying by elevation (overestimation in Da Lat, underestimation in low-lying valleys).

Model Performance Comparison

Temperature: Tree-based ensemble models significantly outperformed linear baselines. The enhanced CatBoost model achieved the highest performance with an $R^2$ of 0.955 and RMSE of 0.451°C under standard settings.
Precipitation: Under standard feature settings, simple linear regression outperformed complex ML models (RF $R^2$ : 0.557 vs. Linear $R^2$ : 0.614). The authors attribute this to the lack of physical descriptors in static features, causing tree-based models to overfit geographic coordinates.

Impact of Enhanced Features

The inclusion of engineered dynamic features dramatically improved performance, particularly for precipitation:

Precipitation: The enhanced Random Forest model improved $R^2$ from 0.669 (standard) to 0.706 and reduced RMSE from 98.98 mm to 93.34 mm.
Temperature: The enhanced model improved $R^2$ from 0.944 to 0.949 and reduced RMSE from 0.582°C to 0.555°C.
Extreme Events: During the 2015–2016 El Niño drought, the enhanced model maintained a stable RMSE of 87.42 mm for precipitation and 0.542°C for temperature, successfully capturing rainfall deficits and SPI3 anomalies.

Feature Importance

Analysis revealed that Orographic Index (OI) and Vapor Pressure Deficit (VPD) were the most critical predictors. OI captured local topographic influences on precipitation (windward uplift vs. rain-shadow), while VPD captured the thermodynamic coupling between atmospheric demand and surface temperature, particularly during the dry season.

Sensitivity Analysis

When observed station lags were replaced with ERA5-derived lags to generate spatially continuous grids, the model's $R^2$ decreased slightly (0.706 to 0.672) but still outperformed the standard topographic-only model, confirming the viability of this strategy for gridded product generation.

Significance and Claims

The paper claims that integrating hydrometeorological coupling (via VPD, Moisture Flux) and climate anomaly indices (SPI3, ONI) into machine learning frameworks significantly improves statistical downscaling in complex terrains.

Physical Representation: The study demonstrates that moving beyond static geographic coordinates to include dynamic, engineered features allows ML models to better represent the physical processes driving local climate extremes, such as orographic lifting and land-atmosphere energy partitioning.
Application: The resulting high-fidelity, station-scale historical climate data addresses the scarcity of ground observations in the Central Highlands. This provides a reliable foundation for:
- Water resource planning and reservoir management.
- Agricultural management, specifically for coffee production which is sensitive to dry-season droughts and temperature anomalies.
- Development of localized early-warning drought systems.

The authors conclude that while the framework is robust, future work should address non-stationarity under climate change, integrate time-varying land cover data (to account for deforestation and coffee expansion), and extend the approach to daily temporal resolution using deep learning architectures.