Aim: Climate modelling in viticulture introduced new challenges such as high spatio-temporal monitoring and the use of dependable time series and robustness modelling methods. Land surface temperature (LST) is widely used and particularly MODIS thermal satellite images due to their high temporal resolution (four images per day). However, this data is not completely adapted to regional scale with its medium spatial resolution (1-km). Downscaling methods can improve spatial resolution using machine learning algorithms implementing multiple predictors as topographical variables and vegetation indices. In the last decades, classical bioclimatic temperature-based indices showed a specific spatial distribution depending on topographical variables and at once a significantly non-correlation with vegetation growing trend.
Methods and Results: In the current study, an assessment of SVM Machine learning method was used to downscaling daily LST using topographical variables and vegetation indices as predictors at multiple spatial resolution. The aims of this study were to (1) evaluate daily LST time series through 2012-2018 period, (2) assess the impact of topographical variables and evolution of vegetation indices during vegetative season and (3) calculation of bioclimatic indices on the wine-growing area of the Gironde The dataset included: 1) daily time series of MODIS LST at 1-km (MOD11A1 and MYD11A1) and 2) topographical variables derived from Digital Elevation Model at 500 m (GMTED10). The first step was the pre-processing and reconstruction of time series. The second step was the downscaling of LST using SVM with topographical variables as predictors. For each day, a model was calibrated and validated to predict daily LST at finer spatial scale. The third step was the calculation of bioclimatic indices (Winkler and Huglin). The methodology was applied for the fourth LST MODIS products acquired at different times. For example, for the 2012 wine growing season Huglin index and Winkler index were calculated with the daily predicted LST (without vegetation indices as predictors but only topographical variables) on the Gironde area and have a globally similar spatial structure. The lowest values (≈ 1900°C for Huglin and 1340°C for Winkler) are concentrated on the coastline to the west and south of the Gironde. The highest index values (> 2000°C for Huglin and > 1700°C for Winkler) are located from the centre of the Gironde to the north-east. These warmer sectors are concentrated in the valley bottoms of the Dordogne and Gironde with higher values in the south of Libourne. LST predictions should be downscaled for the whole period (2012-2019) and the second experiment of the downscaling method includes vegetation indices as predictors.
Conclusion: The advantage of LST is their temporal and spatial covers in all the areas. However, data availability and bias must be taken into account and minimized.
Significance and Impact of the Study: At the scale of Gironde region, this downscaling method has been tested for the first time with MODIS Land Surface Temperature derived from thermal satellite images in a wine-growing context.
Authors: Gwenaël Morin1*, Renan Le Roux2, Pierre-Gilles Lemasle1 and Hervé Quénol1
1LETG-Rennes, UMR 6554 CNRS – Université Rennes 2, Place du Recteur Henri Le Moal, Rennes – France
2CIRAD, Forêts et Sociétés, F-34398 Montpellier, France
Keywords: Climate modelling, topographical downscaling, thermal satellite imagery, bioclimatic indices, Gironde