AI Framework Sharpens Satellite Rainfall Maps Over the Remote Tibetan Plateau

High on the Qinghai–Tibet Plateau, where the average elevation climbs above four thousand meters and ground-based weather stations are few and far between, scientists have long struggled to answer a deceptively simple question: how much rain actually falls there? Satellite precipitation products promise an answer, but their estimates are notoriously noisy in this rugged, sparsely instrumented region. Now, a team of researchers at Nanjing University of Information Science and Technology has developed a hybrid artificial intelligence framework that dramatically improves the accuracy of one of the most widely used satellite rainfall datasets, with the largest gains arriving precisely where they matter most: during heavy and torrential downpours.

The study, published in Theoretical and Applied Climatology, targets the IMERG product, the Integrated Multi-satellitE Retrievals for GPM, which is generated by NASA’s Global Precipitation Measurement mission. IMERG stitches together observations from a constellation of satellites to produce near-global rainfall maps at high spatial and temporal resolution. Yet over the interior of the Tibetan Plateau, the product suffers from well-documented biases. It tends to overestimate both the intensity and the spatial extent of precipitation in some areas while underestimating rainfall locally in others, and its errors grow worse as rainfall intensifies. Because the plateau feeds the headwaters of major Asian rivers and drives regional hydrological cycles, these errors ripple directly into flood simulation, drought monitoring, and water resource management.

Lead author Jiaxin Zhao and colleagues, including Aili Liu, Chong Wang, Lin Zhao, and Tingting Dong, tackled the problem with a two-stage machine learning pipeline that pairs a convolutional neural network, or CNN, with Extreme Gradient Boosting, better known as XGBoost. The key innovation is that the framework is explicitly aware of precipitation intensity. Rather than treating all rainy days as a single undifferentiated class, the system first classifies each satellite estimate into rainfall categories and then uses that classification to guide the correction of the rainfall amounts themselves.

The technical architecture works as follows. The researchers trained a CNN-based classification model using IMERG estimates together with nine auxiliary topographic and meteorological variables, such as terrain characteristics and atmospheric conditions, with ground-based gauge observations serving as the reference truth. The CNN learns to recognize the spatial patterns and contextual signatures that distinguish light rain from moderate, heavy, and torrential rainfall in satellite data. Crucially, the model outputs not just a single category label but a set of probabilities across all rainfall classes. These probability vectors are then fed into the XGBoost regressor as additional features, alongside the original IMERG values and environmental variables, allowing the gradient-boosted trees to correct the magnitude of rainfall differently depending on how likely an event is to be a deluge rather than a drizzle.

This division of labor exploits the complementary strengths of the two algorithms. CNNs excel at extracting spatial features from gridded data, making them well suited to interpreting the two-dimensional structure of satellite precipitation fields and their relationship to the plateau’s complex topography. XGBoost, a scalable tree-boosting system renowned for its performance on tabular regression problems, handles the numerical correction task efficiently and robustly. By passing the CNN’s rainfall-category probabilities into the regression stage, the framework ensures that the intensity information directly shapes the bias correction, rather than being lost in a one-size-fits-all adjustment.

To evaluate the approach, the team benchmarked the corrected IMERG estimates against two independent, long-term, high-resolution meteorological datasets: the Third Pole long-term high-resolution meteorological forcing dataset, known as TPMFD, and the China Meteorological Forcing Dataset, or CMFD. Both are widely trusted reference products for the region, and comparing against two separate benchmarks helps guard against the possibility that improvements simply reflect fitting to one particular dataset’s quirks.

The results are striking. On a daily timescale, the corrected IMERG substantially outperformed the original product, with the correlation coefficient increasing by 0.15 and the root mean square error and mean absolute error decreasing by 0.46 and 0.04 millimeters per day, respectively. The framework also proved seasonally robust, maintaining satisfactory performance in both the dry and wet seasons, with correlation coefficients of 0.35 and 0.65. In a region where monsoon dynamics produce sharp seasonal contrasts in precipitation regime, that consistency matters: many correction methods tuned for the wet season degrade badly during dry months, or vice versa.

The spatial improvements are equally significant. The original IMERG product over the plateau’s interior contains spurious high-value regions where it paints unrealistic pockets of intense rainfall, alongside localized zones of underestimation. The corrected product suppressed these artifacts, producing rainfall maps whose spatial patterns align much more closely with gauge observations. When compared with the CMFD reference, the corrected IMERG reduced the overestimation of both precipitation intensity and spatial coverage, two of the most persistent failure modes of satellite retrieval over high terrain, where orographic effects and complex cloud microphysics confound passive microwave and infrared retrieval algorithms.

Perhaps the most consequential gains came in the extreme rainfall categories. For heavy and torrential rainfall events, the corrected IMERG achieved correlation coefficient increases of 0.14 to 0.31 and reductions in root mean square error of 1.47 to 2.01 millimeters per day relative to the original product, all while maintaining stable performance for light and moderate rainfall. That balance is rare. Conventional bias correction methods often improve average conditions at the cost of smearing out extremes, or they sharpen extremes at the cost of distorting the drizzle that dominates the record. An intensity-aware framework that improves both ends of the spectrum simultaneously represents a meaningful advance for hydrological extreme-event simulation, where the difference between a corrected and uncorrected heavy rainfall estimate can determine whether a flood model captures a real hazard or misses it entirely.

The implications extend well beyond the Tibetan Plateau. Roughly a third of humanity depends on water that originates in High Mountain Asia, and as climate change intensifies both the hydrological cycle and the frequency of extreme precipitation, the demand for accurate, gauge-independent rainfall estimates in ungauged mountains will only grow. The framework’s ingredients are all publicly available: the IMERG product is distributed freely by NASA, and the TPMFD and CMFD datasets are hosted by the National Tibetan Plateau Data Center. That openness, combined with the modular design of the CNN-XGBoost pipeline, suggests the approach could be adapted to other satellite precipitation products and other data-sparse mountain regions worldwide, from the Andes to the Hindu Kush. For a region where every rain gauge covers hundreds of square kilometers of some of the most inaccessible terrain on Earth, teaching satellites to see rainfall more clearly is not just a technical refinement. It is a step toward understanding, and preparing for, the water future of an entire continent.

Subject of Research: Machine learning-based bias correction of GPM IMERG satellite precipitation estimates over the interior of the Tibetan Plateau

Article Title: A precipitation-intensity-aware CNN-XGBoost framework for IMERG bias correction over the interior of the Tibetan Plateau

Article References: A precipitation-intensity-aware CNN-XGBoost framework for IMERG bias correction over the interior of the Tibetan Plateau. (n.d.). https://doi.org/10.1007/s00704-026-06606-z

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06606-z

Keywords: IMERG, Tibetan Plateau, satellite precipitation, bias correction, convolutional neural network, XGBoost, machine learning, extreme rainfall, hydrology, GPM mission, climate data, remote sensing