Résumé
Among the 88 Malagasy ebony species belonging to the Diospyros genus and considered as large trees, 52% are threatened by illegal timber logging and trade. The availability of a rapid and easy-to-use method such as Near-Infrared Spectroscopy (NIRS) for wood identification is of crucial importance to ensure the sustainable management of these resources. Several NIRS spectrometers, di↵ering in their physical characteristics (size,weight, etc.) and technical specifications (spectral range, spectral resolution, etc.), have been calibrated in Madagascar to identify precious woods. This study aims to compare the performance of two of these NIR spectrometers (handheld Viavi MicroNIR 1700 and desktop Bruker MPA II) to discriminate 10 malagasy Diospyros based on their wood spectra. For each spectrometer, 714 NIR spectra were measured on 119 wood cores belonging to the 10 analyzed species. A 4-bloc cross-validation, repeated 20 times was performed on the ¾ of both training data set to chose the best preprocessing method and to tune the hyperparameters of PLS-LDA discrimination models. The tuned models were then applied to class the ¼ remaining spectra. The performance of the models were evaluated using the accuracy metric. There is a significant di↵erence in the performance of the two discrimination models, with accuracies of 92.8% and 57.8% respectively for the Bruker MPA II and the MicroNIR. The low accuracy of the MicroNIR is especially due to its restricted spectral range. For operational purpose, these findings show that for on-site applications, NIR spectroscopy can serve as a helpful identification tool. However, using other methods such as wood anatomy with NIR spectroscopy, or conducting multi-blok analysis by combining both types of group variables, is interesting to improve good identification rates. In cases where a higher level of identification is required, the use of the Bruker MPA II in the laboratory for more in-depth identifications is recommended.