Snow Cover Mapping using Satellite Remote Sensing Data

This paper discusses neural network based approach to generate the spatial distribution of snow accumulation using multi-channel Special Sensor Microwave/Imager (SSM/I) data. Five SSM/I channels (19H, 19V, 22V, 37V, and 85V) were used to remotely sense snow accumulation during 2001/2002 winter season. Ground snow depth measurements were acquired from the National Climatic Data Center (NCDC) through the Cooperative Observer Network for snow monitoring in the United States. The snow depths were compiled and gridded into 25 km x 25 km grid to match the final SSM/I spatial resolution. Neural network based approach was tested and compared with the filtering algorithm developed by Grody and Basist(1) in the Northern Midwest region of the United States. The results indicate that the neural-network-based approach has a great potential in identifying snow pixels from SSM/I data by providing a significant improvement in snow mapping accuracy over the filtering algorithm. Having accurate estimations of snow cover characteristics during the snowmelt season is indispensable for efficient hydrological modeling and snowmelt runoff forecasting (2). Direct measurements of snow depth at a single station are generally not very useful in making estimates of accumulation over large areas. Additionally, the traditional field sampling methods and the ground-based data collection are often very sparse, time consuming, and expensive compared to the coverage provided by remote sensing techniques. Moreover, direct measurements of snow depth at a single station are generally not very useful in making estimates of distribution over large areas since the measured depth may be highly unrepresentative of the study areas even under the same snowfall conditions. At present, most hydrological models that require snowpack information are using maps obtained by gridding standard point gauge measurements or data derived from physically based models (2-4). The estimation of snow depth and snow water equivalent from passive microwave measurements requires a deep understanding of surface and volume emissivity of snowpack and its underlying ground. The measured brightness temperature of the snow-covered surface is a function of both ground and snow cover properties, includes: surface roughness, surface temperature, vegetation cover, snow cover density, snow water equivalent, and snow grain size distribution (3). Many empirical models have been developed to estimate snow depth from spaceborne passive microwave sensors; most of these models make the simple assumption that the snow depth and brightness temperature differences, generally between channels 19 and 37 GHz, are linearly related. The Meteorological Service of Canada (MSC) model, for example, currently uses them to produce real-time SWE maps for the Canadian Prairies (5, 6). In forest environments, SWE retrieval becomes more complicated due to the attenuation of the ground microwave signal propagating through the canopy as well as the vegetation contribution to the brightness temperature (7, 8). Neural network has been successfully applied to a wide range of non-linear problems in several disciplines. Multi-layer perceptron trained by the backpropagation algorithm has also been successfully applied to image classification, and it has shown great potential in the classification of different types of remotely sensed data. A useful review of the application of neural networks in remote sensing can be found in (9, 10). The rapid increase in neural network applications in remote sensing is mainly due to their ability to perform more accurately than other classification techniques especially when the intent is to classify features with overlapped spectral signatures that cannot be easily associated with defined statistical functions. Generally, a neural network is capable of storing a complex functional relationship between its inputs (pixel values) to the outputs and it is proficient in approximating any function with a finite number of discontinuities.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.