TACKLING WITH THE LIMITATIONS OF THE RASTER DATA MODEL: VECTOR-BASED PREPROCESSING OF MODIS REFLECTIVE BANDS

摘要

The Moderate Resolution Imaging Spectroradiometer (MODIS) instruments onboard NASA's Terra and Aqua satellites provide continuous observations and have been increasingly used in different disciplines since 1999. A whole suite of derived products is also produced in a systematic manner. However, these data have some intrinsic characteristics that make their processing a delicate issue when used on pixel- or subpixel-scale studies. Due to the wide field of view, the actual ground pixel size changes continuously off-nadir both in along-track and along-scan directions. All kinds of resampling carried out due to the fixed cell size implemented in the raster data model inevitably result in artifacts in the data. The MODIS gridding process also has a large influence. The average overlap between grid cells and "real" observations is less than 30%. Our results based on synthetic MODIS images show that in our case, an artificial variance of 28% is introduced by the only procedure of raster-based nearest neighbor resampling. Therefore, the accuracy of current MODIS data at its original resolution is not sufficient for carrying out pixel- or subpixel-level studies such as spectral unmixing or pixel-level change detection. Spatial or temporal aggregation is widely used to cope with these issues, but may lead to a considerable loss of information. In this work, we propose and demonstrate an alternative solution based on the vector data model. By using MODIS swath products and geolocation datasets, a polygon layer is created, in which each polygon represents the real pixel footprint sensed during image acquisition. To test the increase of accuracy yielded by the new method, correlation with same-day high resolution (HR) SPOT images is used as an indicator. The results are encouraging: the original 250-m MODIS product in the near infrared domain shows an r2 value of 0.61 with the reference SPOT data, whereas the new method yielded almost 0.8 with the same conditions.