Comparison of methods for the automatic classification of forest habitat types in the Southern Alps-Application to ecological data from the French national forest inventory

The monitoring of habitats at plant association level, has been developed by the French-National Forest Inventory (NFI) progressively since 2011, whereas ecological and floristic data exist since the mid-1980s. The NFI habitat monitoring is the French tool of surveillance of forest habitats decreed by Natura 2000 Directive (article 11). Determination of plant association in NFI plots concerns all the habitats, whether they are of community interest or not. The objective of this study is to compare different methods of automatic classification of floristic and ecological surveys into forest habitat groups. Indeed, enriching the old surveys, which contain only ecological, floristic and trees data, with information on habitats would increase the accuracy of the calculated statistical results on habitats. The uncertainty of the attribution of a habitat outside the field (ex-situ) by experts was quantified by comparison with the determination in the field (in situ). This result was used as a benchmark to compare to the error rates obtained by two methods of automatic classification: an algorithm inspired by the habitat identification key used in the field and Random forest, a learning classification method. The classification performance was evaluated for three levels of habitat groupings. The results showed that the lower the level of clustering, the higher the error rate. Depending on the classification method used and the level of aggregation, the error rates varied between 5 and 15%. In all cases, the error rates were below the estimated uncertainty of the expert attribution of ex-situ habitat.

Automated summary

The monitoring of habitats at plant association level, developed by the French-National Forest Inventory (NFI), is an important tool for surveillance of forest habitats in France. This study compares different methods of automatic classification for floristic and ecological surveys into forest habitat groups. The results show that the lower the level of clustering, the higher the error rate, ranging from 5 to 15%. These findings can help improve the accuracy of forest habitat monitoring and classification.