A Method for Snow-cover Mapping of Forests by Optical Remote Sensing

A snow-covered forest mapping method for optical remote sensing is proposed. The method is based on linear sub-pixel reflectance modelling of the surface components snow, individual tree species and bare ground. Experiments are performed using a 100% snow-covered Landsat TM scene and aerial photos covering spruce, pine and birch forest in the Jotunheimen mountain area of South Norway. The reflectance modelling shows best results for pine forest and mixed pine and birch forest, while the modelled reflectance for birch forest and spruce forest is underestimated. The results are improved when tree shadows on the snow surface and reduced diffuse illumination due to tree crowns masking parts of the sky hemisphere are included in the reflectance model. INTRODUCTION Seasonal snow may cover up to 50 million km (34%) of the Earth’s land surface (1). Most of the seasonal snow is located in the Northern Hemisphere. Boreal forest covers 12 million km (8%) of the land surface, and most of the boreal forest is seasonally snow covered. Seasonally snow-covered forest is also present in high mountain regions of temperate latitudes. Monitoring the snow-cover extent is important both for climatological studies and for hydrological applications. Due to the high snow albedo compared to other natural surfaces, variations in the global snow-cover distribution affect the global energy balance. Hydrological applications include support to hydropower production planning and river flood predictions. Several classification methods have been developed for or adapted to optical snow-cover mapping, e.g. the SNOMAP-algorithm (2), linear spectral unmixing (3), and an empirical linear sub-pixel classification method (4). The classification methods give reasonable results for unforested areas, see e.g. (2). However, forested areas constitute a problem due to the contribution of radiance from the trees, in addition to reducing the radiance from the snow below the trees. Classification methods generally underestimate the snow cover in forest, see e.g. (5). The SNOMAP-algorithm has been extended by including the Normalized Difference Vegetation Index to map snow-covered forest (6), and verification of the algorithm is investigated by (7). The objective of this work is to study the possibilities for determining the snow coverage at subpixel level in forest by optical remote sensing. A snow-cover mapping method, based on linear subpixel reflectance modelling, is proposed. Experiments focus on physical reflectance modelling in order to understand the various factors that influence on the satellite-measured reflectance from a snow-covered forest. By examining the temporal influence of these factors on the pixel reflectance, a simplified operational model may later be derived. The experiments deal with 100% snowcovered spruce, pine and birch forest, and mixed pine and birch forest, while situations of less than 100% snow coverage will be investigated in a later work. Three successive experiments are presented and discussed: 1) The snow surface is modelled entirely illuminated; 2) The tree shadows on the snow surface are included in the reflectance model; and 3) The reduced diffuse illumination Proceedings of EARSeL-SIG-Workshop Land Ice and Snow, Dresden/FRG, June 16 – 17, 2000 EARSeL eProceedings No. 1 329 reaching the snow surface due to surrounding trees that reduce the visible sky hemisphere is modelled and included in the reflectance model. THE SNOW-COVERED FOREST MAPPING METHOD During the snowmelt season, the satellite measured radiance from a snow-covered forest is influenced by a number of surface components, illuminationand atmospheric effects. The main surface components of a snow-covered forest are trees, snow and bare ground. The spectral radiance of each surface cover is affected by the temporal natural variability of the surface cover. Based on the influence of the temporal natural variability on the radiance, a variability ranking of the surface components is made. Snow is temporally the most unstable surface component (diurnally and seasonally). The snow metamorphosis, which is a continuous process, changes the physical properties of the snow. The snow albedo may vary between 35% and 90% (1). Bare ground consists of vegetation, vegetation litter, rocks and soils. Bare ground is ranked as the second most influencing surface component, provided that the vegetation is non-green and the temporal moisture conditions are stable. Dry, snow free conifer and non-green deciduous trees are the least changing surface components, giving the most stable influence on the radiance. Added to these surface components are the direct and diffuse illumination effects, which are due to solar elevation angle, trees and topography. Shadows on the snow/ground surface and shadows in the tree crowns are direct illumination effects, and occur due to single trees. Surrounding trees may increase these shadowed areas (mutual shadowing). Trees also reduce the sky hemisphere visible from the snow/ground surface, and consequently reduce the diffuse illumination reaching the snow/ground surface. Additionally, these direct and diffuse illumination effects are affected by the topography. Among these illumination effects direct illumination effects due to single trees are considered as first order effects, while second order effects are mutual shadowing and diffuse illumination. Based on the ranking of the surface components and first order illumination effects, a linear subpixel reflectance model, which is an area-weighted sum of these components, is proposed for the modelling: , BG BG SWB SWB SWS SWS SWP SWP SW SW B B S S P P R A R A R A R A R A R A R A R A R + + + + + + + = where R is the pixel reflectance, AP, AS, AB are the area proportions of a pixel covered by pine, spruce and birch tree crowns, respectively. ASW and ABG represent the area proportions covered by illuminated snow and bare ground, respectively. RP, RS, RB are the tree-crown reflectances of pine, spruce and birch, respectively. RSW is the illuminated snow reflectance, while RBG is the bare ground reflectance. Shadows within the tree crowns are included in the model using tree-crown reflectances representing the combined illuminated and shadowed tree crowns. A shadowed snow component for pine (ASWP, RSWP), spruce (ASWS, RSWS) and birch (ASWB, RSWB) models the tree shadows on the snow. Spruce and pine tree crowns are assumed opaque, while leafless birch tree crowns are assumed transparent. Therefore, birch tree crown reflectance (RB) is modelled separately by linear mixing of illuminated snow ( SW SW R A , ′ ) and estimated effective branch area proportion ( B B R A ′ ′ , ): B B SW SW B R A R A R ′ ′ + ′ = (Figure 1). The effective branch area proportion is the proportion of birch branches within a tree crown, projected vertically on the snow surface. The effective branch area proportion is estimated using the tree height (h) as input parameter to the regression function AB’ = ah + b. Similarly, the reflectance of the shadowed snow component of birch (RSWB) is modelled separately using a linear mixing reflectance model of illuminated snow ( SW SW R A , ′ ′ ) and estimated effective shadow area proportion ( SWB SWB R A ′ ′ , ): SWB SWB SW SW SWB R A R A R ′ ′ + ′ ′ = . The effective shadow area proportion is also estimated from the tree height (h): ASWB’ = ch + d. Some assumptions are made for the snow-covered forest mapping method. When the forest is observed from nadir, parts of the snow/ground surface must be visible. The ground area covered by a pixel is assumed to be greater than the size of individual trees. The tree-crown reflectance is Proceedings of EARSeL-SIG-Workshop Land Ice and Snow, Dresden/FRG, June 16 – 17, 2000 EARSeL eProceedings No. 1 330 assumed constant for a given solar elevation angle and the area covered by tree crowns is constant during a snowmelt season. The tree crowns are assumed dry and snow free. Figure 1: Birch tree crown reflectance (RB) is modelled by linear mixing of illuminated snow and effective branch area proportions of vertically projected tree crowns. The effective branch area is estimated based on the tree height. An empirical regression function has been developed from measurements of trees of various tree heights.