SatVITS-Flood: Satellite Vegetation Index Time Series Flood detection model for hyperarid regions

Abstract

<p>We present a model for detecting historical floods in ungauged remote hyperarid regions using satellite vegetation indices (SatVITS-Flood). Our model is based on observations that floods in hyperarid regions are the primary cause of change in the growth and expansion of local vegetation. We used 40 years of hydrological data from reliable gauge stations at four sites from three regions over three continents to develop and evaluate the model. MODIS, Landsat, and AVHRR time series of the normalized difference vegetation index (NDVI), the modified soil-adjusted vegetation index (MSAVI), and the normalized difference water index (NDWI) were used to detect vegetation changes. The model uses two different time series analysis metrics, (1) a trend change detection from the Breaks For Additive Season and Trend (BFAST-trend) and (2) a newly developed seasonal change metric based on Temporal Fourier Analysis (TFA) and the growing-season integral anomaly (TFA-GSI<sub>anom</sub>). The idea was to capture major changes in perennial species following extreme, rare floods with BFAST-trend and small changes in the ephemeral vegetation following more frequent, less severe floods with TFA-GSI<sub>anom</sub>. We compared these metrics with flood events and the magnitude of change with the flood volume (Mm<sup>3</sup>) and duration (days). Our model was able to predict flood occurrence with an accuracy of 78% and precision of 67% (Recall = 0.69 and F1 = 0.68; <em>p</em><0.01), and the flood volume with NSE of 0.79 (RMSE = 15.4 Mm<sup>3</sup> event<sup>–1</sup>) and flood duration with R<sup>2</sup> of 0.69 (RMSE = 5.7 days). SatVITS-Flood proved useful for detecting historical floods and may provide valuable hydrological information in poorly-documented areas, which can help understand the impacts of climate change on the hydrology of hyperarid regions.</p>