12/11/2026 –, Sala Geomática 7
Accurate drought prediction in data-scarce urban settings is a challenge with
direct implications for water management. This study evaluates whether combining
ground-based drought indices with MODIS and ERA5-Land satellite data improves
RandomForestdrought index reconstruction across six meteorological stations in Mex
ico City. Two indices — SPI and SPEI — were computed at 3, 6, 12, and 24-month
scales. Four feature configurations and two training strategies were compared: a global
modelusingallstationssimultaneously,andaspatialpoolingapproachthatconcatenates
all station records to increase training sample size from 192 to 1,152 monthly obser
vations. Feature selection (top-30 predictors by importance score) and spatial pooling
substantially improved predictive performance. Average 𝑅2 for SPEI-12 increased from
0.51 (global model) to 0.82 (pooling), and for SPEI-3 from 0.29 to 0.43. The main
finding is that in small-data contexts, expanding the training set through spatial pooling
is more effective than adding satellite predictors alone.