Prédiction de l’usage des sols sur un zonage régulier à différentes résolutions et à partir de covariables facilement accessibles

Abstract

RésuméNous proposons dans ce travail d'évaluer dans quelle mesure un modèle de prédiction à partir de covariables facilement accessibles permet de donner une image de l'usage des sols comparable à celle obtenue à partir de l'information complète au niveau individuel souvent coûteuse. L'utilisation des sols est mesurée par l'occupation physique telle que définie dans l'enquête Teruti-Lucas, et nous considérons cinq catégories : urbain, agricole, forêts, prairies et sols naturels.Notre approche comporte deux étapes : la première étape permet de modéliser l'utilisation du sol au niveau des points Teruti-Lucas tandis que la deuxième étape propose une méthode pour en déduire l'utilisation des sols sur un maillage défini par des carreaux. Le modèle de la première étape fournit des prédictions à un niveau fin. La deuxième étape agrège, en comparant plusieurs méthodes, ces prédictions sur les carreaux du maillage. Nous envisageons différents maillages réguliers du territoire en carreaux pour étudier la qualité de restitution en fonction de la résolution. Nous montrons qu'avec des variables facilement accessibles on obtient une qualité de prédiction acceptable au niveau point et que l'amélioration de la qualité est importante dès la première étape d'agrégation.

Spatial scale in land use models: Application to the Teruti-Lucas survey

“…Another interesting conclusion from our paper in the context of land use is that the Teruti-Lucas survey contains enough observation points in order to t a good classication model for estimating the vector of probabilities. However, averaging the estimated probabilities obtained at the T-L locations only, as we did in Chakir et al (2016c), is not very ecient. Since the explanatory variables we use in our model are available at any location in the Midi-Pyrénées region and not only at the T-L locations, one is better o using this information.…”

“…Table 1 describes the source of each explanatory variable. More detail about the data sets and variables can be found in Chakir et al (2016c), in particular the construction of the ve categories. The nest information comes from the Corine Land Cover data, and the variables population density and altitude.…”

“…Indeed, for the assessment of ecological eects of land uses, since the ecological process of interest, such as habitat quality or dispersal of species are relevant at ner scales, a model of land use at the department level within which ecological conditions vary considerably would be less relevant. Chakir et al (2016c) propose a simple model explaining the T-L data with easily accessible covariates that allow to predict land use at the point level in ve categories. Despite the diculty of this classication problem, they show that 65% of the locations can be correctly classied using land cover and altitude and that additional variables (meteorological, socio-economic and biophysical, spatially lagged explanatory variables) improve only marginally the prediction.…”

“…Despite the diculty of this classication problem, they show that 65% of the locations can be correctly classied using land cover and altitude and that additional variables (meteorological, socio-economic and biophysical, spatially lagged explanatory variables) improve only marginally the prediction. Starting from a good point level model, such as the one obtained in Chakir et al (2016c), the objective of the present paper is to propose an ecient way for predicting land use at dierent aggregated levels.…”

“…The results of the analysis of the prediction error are given in Section 4 at the point level and in Section 5 at aggregated levels. Finally Section 6 concludes by the following recommendation for practitioners: when the objective is to predict land use at a given aggregated level from the Teruti-Lucas survey (or alternatively from a model derived from this survey, as the model presented in Chakir et al (2016c)), one should rst look at the spatial scale of explanatory variables to determine a ne grid of locations, then use the model to compute estimated probabilities on this ne grid of locations before averaging them to obtain the predictions at any larger scale.…”