Suivi spatio-temporel du couvert nival du Québec à l’aide des données NOAA-AVHRR

L’imagerie satellitaire dans le visible et l’infrarouge permet de cartographier le couvert nival à grande échelle, ce qui n’est pas facilement réalisable à partir des observations locales conventionnelles. Cependant, en raison de leur résolution spatiale inadéquate ou de la faible durée de leurs séries d’observations, les produits satellitaires actuellement disponibles sont inutilisables pour l’étude à long terme du couvert nival. Par conséquent, l’objectif de la présente étude a été de développer un algorithme opérationnel de cartographie de la neige à l’aide des données du capteur AVHRR (Advanced Very High Resolution Radiometer) embarqué à bord du satellite NOAA. Cette procédure doit permettre de suivre l’évolution spatio-temporelle de la neige au sol sur une longue période de temps et avec une bonne résolution spatiale. Les résultats de la cartographie ont été validés par rapport aux observations de l’occurrence et de l’épaisseur de la neige au sol. L’algorithme a été appliqué au territoire du Québec sur trois périodes spécifiques : 1998-1999, 1991-1992 et 1986-1987. L’algorithme a réussi à identifier la catégorie de surface (neige/non-neige) avec un taux de succès global moyen de 87 %. Les performances de l’algorithme ont été supérieures dans la détection de la neige (90 %) qu’elles l’ont été pour les surfaces sans neige (82 %). Également, l’algorithme a permis de situer le début des périodes de formation et de fonte de la neige, et ce tant au niveau local qu’à l’échelle du bassin versant.This work is part of a multidisciplinary study designed to validate the elements of the hydrological cycle of the Canadian regional climate model simulations (CRCM) over Quebec (Canada). These simulations, carried out over a 20-year period (1979-1999), aim at examining the annual and inter-annual hydrological budgets of a dozen catchments. Snow cover is a key factor in the modeling of the hydrological budget as well as the climatic changes. The remote sensing component of the project involves the use of satellite data in order to validate CRCM simulations of snow cover characteristics (i.e., snow cover extent), which are impossible to validate using conventional in situ snow observations.Satellite data in the visible and infrared spectra as well as passive microwaves represent an alternative source of information on snow cover. Various satellite snow products have been available since the middle of the 1960’s and a few are available in real time and online. However, their quality varies considerably with respect to sensor and platform characteristics, image processing procedures and snow classification techniques. Consequently, these operational products cannot be used for the validation of the CRCM simulations because of their limited spatial extent, or their coarse spatial resolution, or the lack of a continuous and homogeneous series of observations covering the targeted period (1979-1999). In addition, the coarse temporal resolution and the small areal coverage of high-resolution satellites limit their use for the temporal monitoring of snow cover on a regional scale. Consequently, it was decided to explore the potential of NOAA-AVHRR data for the space-time monitoring of snow on the ground and to produce snow cover maps. These maps would then be used to validate CRCM simulations. Among the 20 years concerned by the study (1979-1999), six winter seasons were targeted to be used in the validation process.The objective of this work was thus to develop a simple procedure of space-time monitoring of snow cover over the province of Quebec using AVHRR images. The algorithm was calibrated and validated over three winter seasons: 1998-1999, 1991-1992 and 1986-1987. In order to monitor snow cover, especially during snow setting and melt phases, the daily images from October 1st to December 15th and from April 1st to May 31st of each of the three periods were used. Images at the beginning of the afternoon were preferred since they are less sensitive to topographic effects and variation in illumination conditions. Only the images presenting a minimal cloud cover were retained (164 images out of the 411 initially identified). These selected images were used for the calibration and validation of the snow cover mapping algorithm. Selected AVHRR images were calibrated and corrected radiometrically and geometrically. A sub-region (82°30’ W, 58°N; 60° W, 46° N) covering the territory being studied was therefore extracted from each image.The classification algorithm used herein was developed from published classification techniques. This algorithm is based on sequential hierarchical thresholds in order to classify the AVHRR images into three surface categories: snow, no-snow and clouds. It consists of a combination of six sequential thresholds. The thresholds go from least restrictive to most severe. A pixel that successfully passes through all the thresholds is classified as snow; if the pixel does not pass through all the thresholds, it is categorized either as clouds or no-snow. The thresholds were established empirically and are consequently specific to Quebec conditions. The classification results were validated at the temporal and spatial levels using ground observations, specifically snow occurrence at Environment Canada’s meteorological stations.The algorithm was calibrated using pixel samples extracted from each selected image, above areas representing the three surface categories present within the scene. These areas were identified visually and delimited manually. Thereafter, radiometric data samples from all selected images were put together and their percentiles were calculated. The percentiles were used to build the values of the algorithm thresholds.For each of the three studied periods, two dates were chosen for the spatial validation of the snow maps produced using AVHRR images: one during the snow cover setting period (at the end of October) and the other for the snow melt period in spring (at the end of April). For these six dates, ground snow occurrence at meteorological stations was compared to the classification results. For temporal validation, snow occurrence observations at 15 meteorological stations during each of the three winter seasons were used for the classification algorithm. Corresponding ground observations were compared to the occurrence of snow class within 3 x 3-pixel windows centered on each station and the total accuracy statistics were therefore calculated. When 50% or more of the 3 x 3-pixel windows were classified as cloudy, the results for the corresponding station were excluded from the comparison.The classification results were quite accurate, with 87% of the pixels around validation meteorological stations being correctly identified. The algorithm successfully detected the presence of snow with a precision of 90% and 82% for no-snow surfaces. The algorithm performances in spring and autumn were similar. Also, the algorithm detected the presence of snow more accurately in open lands than in forested areas. We demonstrated that the algorithm allowed the location of the beginning of snow formation and melting periods at the local level as well as at the watershed scale, especially under clear sky conditions. The algorithm also captured interannual dynamics and spatial variations in the establishment and disappearance of snow cover. The use of high spatial resolution imagery (LANDSAT or SPOT) would improve the accuracy assessment of the algorithm results according to soil occupation types and pixel fractional snow coverage. The main limitation of the algorithm application is the presence of persistent clouds