Centre International de Recherche

sur l’Environnement et le Développement

Nos tutelles

CNRS Ecole des Ponts CIRAD EHESS AgroParisTech

Nos partenaires



Accueil > Actualités

Séminaire CIRED : Améline Vallet

par Arancha Sánchez - publié le , mis à jour le

Important : Le séminaire du CIRED reprend au créneau habituel mais en visioconférence. Pour vous connecter, merci de contacter Antoine Missemer.

Résumé / Abstract

Efforts to mitigate climate change through UNFCCC mechanisms (such as Clean Development Mechanism or Reduced Emissions from Deforestation and Degradation), payments for ecosystem services, or offsetting projects require to map and monitor soil carbons stocks and emissions related to land-use change. The quality of these assessments depend on the soil data available, and their spatial distribution. However, most of existing global soil databases are rather incomplete, and poorly cover some regions of the world. For instance, the World Soil Information Service database contains only 159 soil profiles for Peru (less than 0.15 per 1000 km2), and these are mainly located in tropical areas. In this research project, we build a new soil database for the Southern Peruvian Andes by compiling soil analyses from scientific publications, university student thesis, soil laboratories archives, institutional projects. This unique database contains 3526 soil observations, coming from 1876 sampling sites and 28 sources. We also develop predictive models of bulk density, a key physical property of soils which influences carbon stocks, but which is often missing from soil analyses. To do so, we use a subset of 1653 soil samples from our database, located in the Apurímac region only, an important agricultural and mining region in the Peruvian Andes. We compare different machine learning algorithms (such as random forests and gradient bosting), trained with sets of explanatory variables of increasing size, from one to 17 inputs each (soil physical and chemical properties). For each set of explanatory variables tested, we identified the best performing algorithm, as well as the 2-3 other well performing algorithms that did not decrease model accuracy by more than 2%. Over all considered algorithms and sets of inputs, random forests trained with texture, organic matter and other soil fertility-related inputs showed the very best accuracy (RMSE = 0.095, NSE = 0.777 and r2=0.790 based on an independent test dataset). Depth had a limited effect on model accuracy, and could thus be omitted when predicting BD. Next steps of the project will be to compare different soil carbon stocks mapping approaches using our soil database and the BD estimations where necessary.

Télécharger l’affiche / Download the poster

+ information sur le séminaire du CIRED / about the CIRED Seminar


Ajouter un événement iCal