SACADEAU: A Decision-Aid System to Improve Stream-Water Quality
by Marie-Odile Cordier
Water quality is a critical environmental issue. In this project, we use a pesticide transfer model to simulate effects on the quality of stream-water. Since a large number of parameters are involved in pollution phenomena, modelling, simulation and machine learning are useful techniques for acquiring knowledge in this poorly understood domain.
The objective of the SACADEAU is to build a decision-aid tool to help specialists in charge of catchment area management to preserve stream-water quality. This is done by coupling a qualitative transfer model (simulating pesticide transfer through the catchment area) with a qualitative management model (simulating farmers’ decisions concerning weeding strategies and herbicide application). This has two main advantages: it allows the impact of recommendations to be evaluated by simulating high-level scenarios, and it allows simulation results to be analysed by using machine learning and data mining techniques to discover discriminating variables and to acquire knowledge in this poorly understood domain.
The experimentation site (the Fremeur catchment area) is located in Brittany, France and covers about seventeen square kilometres.
A Transfer Model Coupled with Three Input Models
The transfer model simulates river contamination by pesticides. It models pesticide transfer through a catchment area and simulates the daily river contamination that this causes. This phenomenon depends on numerous parameters, including human activities, climate, soil type and catchment area topology. Since these parameters are complex and difficult to formalize, we created three sub-models to describe them (see Figure).
These are as follows:
- A decision model, which models farmers' strategies. This provides herbicide application characteristics (date, substance, quantity) and agricultural interventions (soil preparation, seeding date, weeding dates) according to predefined farmers' strategies and weather conditions.
- A climate model, which provides daily weather data such as the temperature and the quantity of rainwater.
- A spatial model, which is in charge of the spatial distribution of agricultural activities, according to the catchment area topology.
Using the outputs of these three sub-models, a biophysical transfer model determines pesticide transfer from application locations, through the catchment area, to the river. The model takes into consideration all the possible ways in which rainwater can flow through the catchment area (run-off and leaching).
A High-Level Language
In order to achieve qualitative results, a high-level language for inputs and outputs of the model is required. The aim is to describe qualitatively, via a scenario, numerical inputs and outputs of the model. This can be seen as the process of discretization of quantitative data. This process is fundamental if we want to construct comprehensive results for decision-makers. The initial step was to gather a set of scenarios suggested by experts; for example, "What happens if a post-emergence weeding strategy rather than a pre-emergence strategy is applied on all plots close to the river?" or "What is the impact of pesticide application dates on stream-water quality?"
To simulate a scenario, a methodology was defined that consists in generating a large set of instances of the scenario. These instances are then simulated and the results generalized to give a qualitative description (in response to a qualitative question) using machine-learning techniques. For example, given the question above concerning the impact of application dates, a response could be: "Concentration peaks appear when pesticide application dates are close (less than two days) to significant showers (quantity > 10mm)".
Learning from Simulation Results
The global model generates pesticide quantities and concentrations according to the parameters mentioned above. The set of simulation inputs and outputs is called an example in machine-learning vocabulary, and the set of examples is the set of instances of a scenario we have simulated. We used the inductive logic programming software ICL (http://www.cs.kuleuven.ac.be/~wimv/ICL/) to learn a set of rules summarizing the examples. These rules can be formatted according to the task being addressed. For example:
- qualitatively predicting water pollution
- identifying which variables play an important role in water pollution
- characterizing important risks (eg whether pollution is due to too many concentration peaks, or to a constant and significant source of pollution).
First results have been obtained with a simplified model, and according to the experts, they show both expected and surprising relationships. For instance, concerning stream-water quality, a post-emergence weeding strategy did not show better results than a pre-emergence strategy. This was unexpected, and a discussion on the impact of weeding strategies was generated, with some possible explanations given by experts.
Future work is three-fold, and involves validating the model, refining the high-level language, and providing recommendations to experts based on relationships discovered through the model.
The SACADEAU project is in its third year of development. Contributing members are M. O. Cordier, V. Masson, A. Salleb (IRISA/Univ. Rennes 1, Rennes), C. Gascuel-Odoux, F. Tortrat, P. Aurousseau, R. Trepos (INRA-ENSAR/UMR SAS, Rennes), F. Garcia (INRA/BIA, Castanet Tolosan), B. Chanomordic (INRA/LASB, Montpellier), M. Falchier, D. Heddadj and L. Lebouille (Chambre d'agriculture de Bretagne). SACADEAU is funded by Conseil Général du Morbihan and INRA.
Link:
http://www.irisa.fr/dream
Please contact:
Marie-Odile Cordier, Université Rennes 1 / IRISA, France
Tel : +33 2 9984 7100
E-mail: Marie-Odile.Cordier
irisa.fr
