A machine learning approach for evaluating the impact of land use and management practices on streamwater pollution by pesticides

Abstract : Streamwater pollution by pesticides is a critical environmental issue in farmed catchment areas. Many important factors are involved in this pollution phenomenon, like weather, area topology, land use and crop management practices, which all influence streamwater quality. The purpose of the ongoing study presented in this paper is to evaluate the impact of land use and management practices on streamwater pollution. We use modelling, simulation and machine learning techniques for acquiring knowledge about this complex domain. Our main objective is to learn qualitative rules relating the pollution factors to the temporal distribution of the stream pesticide concentration. The study area is the farmed catchment of Fremeur (~17 km2), located in Brittany, France. Our approach relies on a simulation model, called SACADEAU, based on two main components: a transfer model and a management model. The biophysical transfer model is the core of the model. It aims at simulating on a daily basis the pesticide transfer through the catchment area from application locations on maize parcels to the river. The management model simulates farmers\\\' operations like tillage, sowing, weeding (pesticide treatments described by dates, molecules, quantities) on maize crop parcels. A climate model, which provides daily weather data such as temperature and rainfall amount and a spatial model, which distributes in space the agricultural activities according to the fields and catchment area topology, are the two other components of the SACADEAU model. Using the outputs of these three sub-models, a biophysical transfer model determines herbicide outflow, modelling transfer from application locations, through the catchment area, to the stream. This simulation model is used for generating a large number of trajectories of the catchment system, considering different weather series or spatial distributions of land use and agricultural activities. The complexity of the model inputs and outputs makes difficult interpretation of results. Qualitative trajectories, or scenarios, are thus obtained by using machine learning techniques on these simulation data. ICL, an inductive logic programming software, is used for learning a set of rules which summarizes the scenarios. The Sacadeau model is not already fully implemented and first results have been obtained with a simplified model. We were able to check the coherence and the feasibility of our approach, and to build a first view of the role of some attributes in stream-water quality. Thanks to the complete model, more specific rules declined in terms of temporal and spatial variations should be established in the future.
Type de document :
Communication dans un congrès
Modelling and Simulation Society of Australia and New Zealand. MODSIM'05 (International Congress on Modelling and Simulation), 2005, Melbourne, Australia. 2005
Liste complète des métadonnées

https://hal.inria.fr/inria-00511101
Contributeur : René Quiniou <>
Soumis le : lundi 23 août 2010 - 17:36:46
Dernière modification le : mercredi 16 mai 2018 - 11:23:02

Identifiants

  • HAL Id : inria-00511101, version 1

Citation

Marie-Odile Cordier, Frederick Garcia, Chantal Gascuel, Véronique Masson, Jordy Salmon-Monviola, et al.. A machine learning approach for evaluating the impact of land use and management practices on streamwater pollution by pesticides. Modelling and Simulation Society of Australia and New Zealand. MODSIM'05 (International Congress on Modelling and Simulation), 2005, Melbourne, Australia. 2005. 〈inria-00511101〉

Partager

Métriques

Consultations de la notice

496