Spatial Scaling of a Remotely Sensed Surface Parameter by Contexture

Abstract

Various remote sensing sensors observe the Earth’s surface at different spatial resolutions. In deriving surface parameters using remotely sensed data, the transportability of algorithms from one resolution to another is often of great concern because of the surface heterogeneity. This article addresses this scaling issue through image degradation experiments using a Landsat TM image. It is shown theoretically that scaling problems in deriving surface parameters exist not only because of the nonlinearity in the relationships between remote sensing measurements such as NDVI (normalized difference vegetation index) and SR (simple ratio) and the parameters of interest, but also because of the discontinuity between contrasting cover types within a mixed pixel. To quantify the effects of the nonlinearity and discontinuity on scaling, it is found that contextural parameters are more effective than textural parameters. Contexture-based functions are derived for the estimation of the scaling effects on leaf area index (LAI) calculations using algorithms based on NDVI and SR separately. Based on NDVI–LAI and SR–LAI relationships that were derived at the Landsat TM scale (30 m) as part of the Boreal Ecosystem–Atmosphere Study (BOREAS), the effects of scaling on the retrieval of LAI were investigated using nine selected areas of the same size (990 m×990 m) but different water area fractions. The following conclusions are drawn from the investigation: 1) Negative biases in the estimation of LAI occur when either the NDVI or SR algorithm derived at a fine resolution (Landsat TM) is used for calculations at a coarse resolution (for example, AVHRR). 2) The amount of the biases depends on the surface heterogeneity. For a pure forest pixel, the bias caused by the nonlinearity of the NDVI algorithm was smaller than 2% and the linear SR algorithm induces no error in scaling. Therefore, the scaling problem for pure pixels may be ignored for many applications using either linear or nonlinear algorithms. 3) Large negative biases occur when a pixel contains interfaces between two or more contrasting surfaces. In the case of two contrasting surfaces between vegetation and open water, the biases can be up to 45% of the correct value depending of the water area fraction in the pixel. The biases in this case depend on contexture and little on texture. Simulations show that the most useful contextural parameter for quantifying the scaling effects in vegetation–water mixed pixels is the water area fraction within each degraded pixel. Algorithms for remote sensing applications can be transported from one scale to another, if the information on the water body size is available. This study shows the need for global water masks at high resolutions for the purpose of accurate derivation of surface parameters maps at various resolutions. In boreal regions, this is particularly important because of the large number of small water bodies. Crown