INTEGRATING INSAR INFORMATION AND SPATIAL-TEMPORAL FACTORS IN MACHINE LEARNING ANALYSIS FOR LANDSLIDE PREDICTION – A CASE STUDY FOR PROVINCIAL HIGHWAY 18 AREA IN TAIWAN

Rain in one place often comes from water that evaporated far away, so losing forests can dry out regions thousands of kilometres downwind. To map these hidden links across South America, we used a weather-driven computer model to follow, day by day for thirty years, where the rain falling on each small area first evaporated. Our detailed continent-wide record can be explored on an ordinary laptop, helping researchers and decision-makers understand and manage water, droughts, and deforestation.

This study presents a new way to create detailed rainfall maps using only a small number of rain gauges. By learning rainfall patterns from radar data, the method can reconstruct realistic rainfall across an area even when measurements are sparse. This approach can improve flood forecasting and water management, especially in regions where weather radar is not available or data coverage is limited.

Heavy rainfall from typhoons can trigger sudden floods, yet the most intense bursts are brief and easily missed. We measured rain every ten seconds during three 2022 typhoons in Taiwan and examined how the patterns behaved across time scales. One storm produced exceptionally concentrated, rare rainfall, while the others were limited by record length. The results show that fast measurements matter because slower instruments smooth out these peaks and can lead us to underestimate flood risk.

We developed a framework to improve short-term rainfall forecasts by combining radar data with rain gauge observations. This approach reduces errors and uncertainty, giving more reliable predictions of when and where rain will fall. Such improvements are valuable for flood warnings, stormwater management, and other decisions that depend on timely and accurate rainfall information.

pyBL is an open-source package for generating realistic rainfall time series based on the Bartlett–Lewis (BL) model. It can preserve not only standard but also extreme rainfall statistics across various timescales. Notably, compared to traditional frequency analysis methods, the BL model requires only half the record length (or even shorter) to achieve similar consistency in estimating sub-hourly rainfall extremes. This makes it a valuable tool for modelling rainfall extremes with short records.