Modelling of wind damage to forests using artificial Intelligence techniques
Résumé
Modelling of Wind Damage to Forests using Artificial Intelligence Techniques
Wind causes more than 50% by volume of all damage to European forests. In south-west France there have been two major storms in the recent past that have threatened the viability of forestry in the Aquitaine region. On 27 December 1999 Storm Martin caused a loss of 26 million m3 of timber in the north of the region and on 24 January 2009 Storm Klaus caused 37 million m3 of timber loss further south. The damage was predominately to maritime pine (Pinus pinaster) and represented damage to 15% and 32% respectively of the maritime pine standing volume.
Recent attempts to model the observed damage patterns in Aquitaine using a mechanistic model (ForestGALES) and logistic regression have had mixed success. The models were first tested on the Nezer Forest, a small area that had a detailed survey of tree characteristics and damage following the Martin storm. Both models did a good job of predicting which trees would be expected to be damaged. However, when the models were applied across Aquitaine using data from the National Forest Inventory (NFI), the logistic regression model performed very poorly and ForestGALES only worked well in areas with soils similar to the soils from tree pulling tests used in model parameterisation.
In this paper we present a new analysis of the damage data from Aquitaine using two artificial intelligence techniques; Neural Networks and Random Forests. The results show that when trained on a sub-sample of the damage data from either the Nezer Forest or part of the Aquitaine NFI data, the models perform extremely well on predicting the trees that were damaged across the whole of Aquitaine. In all cases the artificial intelligence techniques outperformed the logistic model and ForestGALES when evaluated using Receiver Operator Curves (ROC). In addition we found that Neural Networks could accurately simulate the outputs of the ForestGALES model. These results suggest that firstly it is possible to make accurate damage prediction models for the Aquitaine Region, and secondly that in applications such as landscape design and regional forecasting Neural Networks can replace computationally intensive programs such as ForestGALES. Such approaches could be adopted in other forest regions and could be powerful techniques for guiding forest management and forest design to reduce wind damage risk.