Analisis Spatiotemporal Varying Coefficients Model Berbasis Generalized LASSO pada Data Semi-continuous: Studi Kasus Banjir Sumatra

Abstract

Banjir merupakan permasalahan hidrometeorologi yang dipengaruhi oleh berbagai faktor hidroklimatologis dan lingkungan yang hubungannya dapat berbeda antarwilayah dan antarwaktu. Penelitian ini menerapkan generalized LASSO untuk memodelkan asosiasi spatiotemporal antara kerentanan banjir dan empat peubah penjelas, yaitu runoff, Normalized Difference Vegetation Index (NDVI), potential evapotranspiration (PET), dan Palmer Drought Severity Index (PDSI), di 10 provinsi Pulau Sumatra periode Januari 2024 hingga Desember 2025. Peubah respons semi-continuous ditangani melalui transformasi logaritma dengan penambahan jitter normal, yang menurunkan skewness dari 3,103 menjadi 0,092. Tiga struktur ketetanggaan spasial (queen's contiguity termodifikasi, kNN k=2, dan k=3) serta dua metode pemilihan parameter penalti (ALOCV dan GCV) dibandingkan. Model terbaik diperoleh pada struktur queen's contiguity termodifikasi dengan GCV, menghasilkan reduksi kompleksitas 89,8% dari 960 koefisien awal menjadi 98 derajat bebas efektif. Hasil menunjukkan asosiasi heterogen secara spasial dan temporal. Runoff berasosiasi positif konsisten dan menguat menjelang akhir periode pengamatan, NDVI berasosiasi negatif stabil, PET menunjukkan perubahan arah asosiasi, sedangkan PDSI menunjukkan variasi paling beragam. Generalized LASSO mampu mengidentifikasi perbedaan pola asosiasi faktor hidroklimatologis dan lingkungan menurut wilayah dan waktu, sekaligus menyederhanakan model melalui fusi koefisien.

Flooding is a hydrometeorological hazard influenced by various hydroclimatological and environmental factors whose relationships may vary across regions and over time. This study applied generalized LASSO to model spatiotemporal associations between flood vulnerability and four explanatory variables, namely runoff, the Normalized Difference Vegetation Index (NDVI), potential evapotranspiration (PET), and the Palmer Drought Severity Index (PDSI), across 10 provinces in Sumatra from January 2024 to December 2025. The semi-continuous response was addressed through a logarithmic transformation with normally distributed jitter, reducing skewness from 3.103 to 0.092. Three spatial structures (modified queen's contiguity, kNN k=2, and k=3) and two penalty selection methods (ALOCV and GCV) were compared. The best model used modified queen's contiguity with GCV, achieving an 89.8% complexity reduction from 960 initial coefficients to 98 effective degrees of freedom. The results revealed spatiotemporal heterogeneity. Runoff showed a consistently positive association that strengthened toward the end of the observation period, NDVI showed a stable negative association, PET displayed a shift in association direction, and PDSI showed the greatest variability. Generalized LASSO successfully identified differing hydroclimatological and environmental association patterns across regions and time while yielding a more parsimonious model through coefficient fusion.