Using the new French Land Cover Map (OSO) as spatial inputs in forest ecological modeling
Résumé
Since 2016, a new French Land Cover Map (the OSO product) is available at a country scale with 17 land cover classes representing main land cover types (urban, agricultural and semi-natural). It is based on a fully automatic processing chain using Sentinel-2 image time series and Landsat images. The classification accuracy is around 90\% and reasonably enables to use this map as spatial inputs in species-habitat model. However, information on potential spatial uncertainty effects in these ecological models remain limited yet.In this study, we explore the impact of spatial uncertainty on a Generalized Linear Model (GLM) that investigates area and connectivity effects of 48 woodlands on the species richness of forest specialist hoverflies. We compared three different models calibrated from the three OSO products available (10m raster, 20m raster and the 20m vectorized product).Firstly, AREA and CONNECTIVITY were derived for all detected sampled forest fragments in each forest product. Then, a statistical protocol was defined in order to evaluate performances and sensitivity of each model : (1) a Leave One Out Cross Validation (LOOCV) was applied. Then, we computed mean and standard deviation for indicators usually analyzed in ecological statistical models outputs : the deviance explained (D2), the significance of regression coefficients (p-value) and the predictive capability (Root Mean Squared Error - RMSE) ; (2) We used a boostrapping technique (1000 runs) to approximate the distribution and test the stability of regression coefficients estimates for each model. Therefore, for each k-1 cross validation run, we also analyzed the standard deviation of the RMSE.Preliminar results show that spatial inputs derived from OSO products had no impacts on the predictive capability of the models (Rho = 0.85, p < 0.001). Nevertheless, we observed strong impacts on the quantity of Deviance explained (Mean Absolute Error (D2) = -0.12 ± 0.08) and the significance of the coefficients. AREA effect on the response variable remains significant with the « raster 10 m » (p-value <0.05) product but disappears in the other model’s configurations. The significance level of the connectivity variable also decreases with the increase of spatial resolution and/or simplification. It became strictly non-existent in the model based 192 on variables from the « vectorized product".