Land use and land cover inference in large areas using multi-temporal optical satellite images

This paper describes a new land use and land cover (LULC) classification method for classifying multi-temporal high-resolution satellite data in large areas. The classification method uses combined value of both reflectance of a pixel and its observation date as an input data, and calculates its probability distribution among all LULC classes via Bayesian inference based on a generative model estimated by kernel density estimation. This method can be easily applied to multi-temporal data to exploit phenological change information of vegetation, even if available multi-temporal data have a seasonal bias. In this paper, we conducted the classification over the entire land mass of Japan, using the multi-temporal data observed by the Advanced Visible and Near Infrared Radiometer type 2 (AVNIR-2) aboard the ALOS, and we evaluated its accuracy in comparison to conventional methods.