Palsar tropical forest and cover mapping, mosaicing and validation, case study Borneo

The production of spatially detailed maps of (very) large areas, and time series of these maps, requires dedicated processing approaches. PALSAR data collection is done in partly overlapping strips or swaths. Since one complete observation cycle of the ALOS satellite is 46 days, adjacent strips, typically, feature differences in acquisition time of several days or weeks. Because of rainfall events, flooding dynamics and incidence angle effects, mosaicing of the backscatter data is tedious. The best approach (which is discussed in this paper) is to make pre-classifications strip by strip and mosaic the strips during the final stage of the classification process. Another challenge is related to the very large volume of the data and, consequently, the need to use computer time in a very efficient way. This problem will be approached by statistical analysis of the data (before classification) to derive appropriate (radar) legends (for each strip) in a highly automated way. In the final mapping stage the legend is made compliant with user needs (based on radar legends of individual strips). This paper shows how mixture modelling and Markov Random Field classification can be utilised for (seamless) mosaicing. New tools are discussed to automate statistical cluster matching between series of adjacent strips. One of the new tools utilises polarimetric data sampling and a new polarimetric decomposition approach. Validation is done with large ground data sets and other reference sets spread over Borneo. The final maps (based on the 2007 FBS and FBD strips) feature high accuracy, large thematic detail with a fair number of forest classes and other land cover classes, and compliance with LCCS and IPCC guidelines. These maps may be of key interest to develop REDD for the humid tropics