Annual mapping of large forest disturbances across Canada's forests using 250 m MODIS imagery from 2000 to 2011

Disturbances such as fire and harvesting shape forest dynamics and must be accounted for when modelling forest properties. However, acquiring timely disturbance information for all of Canada's large forest area has always been challenging. Therefore, we developed an approach to detect annual forest change resulting from fire, harvesting, or flooding using Moderate Resolution Imaging Spectroradiometer (MODIS) imagery at 250 m spatial resolution across Canada and to estimate the within-pixel fractional change (FC). When this approach was applied to the period from 2000 to 2011, the accuracy of detection of burnt, harvested, or flooded areas against our validation dataset was 82%, 80%, and 85%, respectively. With FC, 77% of the area burnt and 82% of the area harvested within the validation dataset were correctly identified. The methodology was optimized to reduce the commission error but tended to omit smaller disturbances as a result. For example, the omitted area for harvest blocks greater than 80 ha was less than 14% but increased to between 38% and 50% for harvest blocks of 20 to 30 ha. Detection of burnt and harvested areas in some regions was hindered by persistent haze or cloud cover or by insect outbreaks. All resulting data layers are available as supplementary material. Key words: boreal forest, National Burned Area Composite, remote sensing, regression tree, decision tree, change detection. Les perturbations telles que le feu et la coupe faconnent la dynamique forestiere et doivent etre prises en compte lorsqu'on modelise les proprietes de la foret. Cependant, l'acquisition rapide de l'information au sujet des perturbations pour l'ensemble des vastes superficies de foret du Canada a toujours presente un defi. Nous avons par consequent developpe une approche pour detecter les changements annuels de la foret dus au feu, a la coupe ou aux inondations en utilisant l'imagerie MODIS avec une resolution spatiale de 250 m partout au Canada et pour estimer le changement fractionnel (FC) a l'echelle du pixel. Lorsque appliquee a la periode allant de 2000 a 2011, la precision de la detection des superficies brulees, coupees ou inondees sur la base de notre jeu de donnees de validation atteignait respectivement 82,80 et 85%. Avec le FC, 77% des superficies brulees et 82% des superficies coupees ont ete correctement identifiees dans le jeu de donnees de validation. La methodologie a ete optimisee pour reduire les erreurs dues a la detection de faux changements avec le resultat cependant qu'elle a tendance a ne pas detecter les plus petites perturbations. Par exemple, les superficies non detectees pour les aires de coupe plus grandes que 80 ha etaient inferieures a 14%, mais atteignaient entre 38 et 50% pour les aires de coupe de 20 a 30 ha. La detection des superficies brulees et coupees etait difficile dans certaines regions du a la brume ou au couvert nuageux persistant, ou par les epidemies d'insecte. Toutes les couches de donnees resultantes sont disponibles dans la section Materiel supplementaire. [Traduit par la Redaction] Mots-cles: foret boreale, Composite nationale des superficies brulees, teledetection, arbre de regression, arbre de decision, detection de changement.
Introduction The managed forests of Canada cover approximately 229 Mha (Environment Canada 2012). Much of this area is in the boreal forest, which is characterized by slow growth and significant [...]