Drought characterization in hyper-arid endorheic basins requires multi-index approaches that capture distinct hydrometeorological processes. This study investigates spatio-temporal drought dynamics in the Tarim Basin (TB)—China’s largest inland arid region—using two complementary remote sensing indices: the Temperature Vegetation Dryness Index (TVDI) for landscape-scale moisture and the Crop Water Stress Index (CWSI) for agricultural drought. Based on 2000–2024 remote sensing and meteorological data, we employed a pixel-wise Random Forest framework with spatial cross-validation and permutation importance analysis to quantify climatic drivers across the TB. Results reveal a fundamental “core-periphery” dichotomy: TVDI identifies persistent extreme drought in the Taklamakan Desert core, while CWSI reveals alleviating water stress in peripheral oasis farmlands (73.21% showing significant decrease, p < 0.05). Despite regional warming-wetting trends, TVDI exhibited an insignificant decrease (54.72% of the basin), contrasting with CWSI's significant agricultural drought alleviation. Vapor Pressure Deficit (VPD)—a key atmospheric dryness indicator—exhibited high relative permutation importance for both drought indices (72–75%), considerably exceeding the values obtained for precipitation (6–8%) within the Tarim Basin. Secondary drivers diverge by land surface type: TVDI responds to Relative Humidity (8.2%) and Precipitation (6.1%), while CWSI is modulated by Land Surface Temperature (9.4%) and Sunshine Hours (7.8%). Partial correlation analyses controlling for topography and temperature confirm VPD’s independent effect on drought severity. Large-scale climate oscillations, particularly the Arctic Oscillation (AO) and ENSO-PDO interactions, significantly modulate interannual drought variability (r = 0.74–0.75, p < 0.01). This study provides the first pixel-scale quantification of the relative dominance of atmospheric water demand over precipitation in driving drought evolution in the Tarim Basin, with VPD contributing 72–75% of the total permutation importance compared to 6–8% for precipitation. This quantitative benchmark offers actionable parameters for drought monitoring systems in arid regions and underscores the need to integrate VPD and large-scale climate signals into early warning frameworks.

Flash drought (FD) is an extreme climate event that intensifies within days and exerts severe socio-environmental impacts. Its onset and evolution remain difficult to predict. Here, we quantify the spatio-temporal dynamics of FD across northern Xinjiang from 1961 to 2023 and identify the dominant driving factors. We apply linear trend detection, wavelet analysis, change-point detection, random forest (RF) modeling, and Pearson correlation. Results show that winter is becoming significantly wetter, whereas the annual signal and the other three seasons exhibit drying trends. After 1980, both FD frequency and FD duration increased; the longest single event lasted 40 days. Spatially, FD is concentrated in the Ili River Valley and the Altay region; the Akdala station recorded the highest count (nine events). Duration, rather than frequency, peaks on the northern slope of the Tianshan Mountains, where the maximum length reaches 40 days. RF importance ranks the Pacific Decadal Oscillation (PDO) as the leading driver (20.9%), followed by air temperature (17.8%); the sunspot index contributes only 6.1%.