FOREST MAPPING AND MONITORING IN TASMANIA USING MULTI-TEMPORAL LANDSAT AND ALOS-PALSAR DATA E.A. Lehmann 1 , Z.-S. Zhou 1 , P. Caccetta 1 , A. Milne 2 , A. Mitchell 2 , K. Lowell 3 , and A. Held 4 1 Commonwealth Scientific and Industrial Research Organisation (CSIRO), Division of Mathematics, Informatics and Statistics, Perth WA, Australia 2 Cooperative Research Centre for Spatial Information (CRC-SI), School of Biological, Earth, and Environmental Sciences, The University of New South Wales, Kensington NSW, Australia 3 CRC-SI, Dept. of Infrastructure Engineering, University of Melbourne, Carlton VIC, Australia 4 CSIRO AusCover Facility, Canberra ACT, Australia ABSTRACT Developing a large-scale forest monitoring system able to take advantage of the complementary nature of optical and radar remote sensing data presents a number of technical and conceptual challenges. This paper investigates the issue of sensor interoper- ability in a time series of Landsat and ALOS-PALSAR data for purposes related to forest mapping and monitoring. The proposed approach relies on the processing methods developed in the frame of an existing and operational Landsat-based forest monitoring system. These methods are here applied to a PALSAR dataset within a bioregion of north-eastern Tasmania, Australia. Particular attention is given to the selection of training data in an attempt to generate results comparable to those obtained with the original Landsat-only time series, thereby allowing for a relevant assessment of interoperability. Results are presented in the form of forest maps and areal forest estimates. Despite similar gross amounts of forest extents, these results highlight differences in the forest (and change) classifications produced using different sensors. Combinations of sensors should therefore be carefully considered in light of what is required of the monitoring system. Index Terms—Forest mapping, remote sensing, multi- temporal, multi-sensor, interoperability. 1. INTRODUCTION In conjunction with historical time series of mostly optical data extending back to the 1970s [1], more recent remote sensing (RS) technologies, such as synthetic aperture radar (SAR), offer new and promising opportunities for enhanced monitoring of the changing extents of the world’s forest resources [2]. The successful integration of RS data originating from sensors operating at very different wavelengths, however, presents a number of challenges for the development of large-scale multi-sensor forest monitoring systems, mainly due to the different type of thematic information provided by each sensor. Thus, there is significant scientific interest in the development of forest information systems that are able to take advantage of the synergies and complementary nature of optical and SAR data. Initiatives such as the Forest Carbon Tracking task of the Group on Earth Observations (GEO-FCT), for instance, aim to demonstrate that coordinated Earth observations can provide the basis for reliable forest information services of suitable consistency and accuracy to support the operation of global forest carbon estimation and reporting systems as part of the UNFCCC (United Nations’ Framework Convention on Climate Change) and REDD (Reducing Emissions from Deforestation and Forest Degradation) agreements [3]. Such initiatives highlight the need for robust and standardized methods to generate forest information products at local, national and global scales, on the basis of RS data acquired by various optical and radar sensors. Key aspects of data inter- operability (obtaining the same thematic results with different sensors) and complementarity (adding thematic value by using two or more sensors) are therefore of specific interest in the develop- ment of such multi-sensor monitoring systems. As part of GEO-FCT, the island state of Tasmania, Australia, was selected as one of seven “national demonstrator” regions to showcase the use of RS data for consistent, wall-to-wall time-series estimates of forest and land cover change. Significant progress has been achieved so far in the frame of the Tasmania demonstrator program, particularly in the area of joint SAR-optical processing using time series of Landsat MSS/TM/ETM+ and ALOS-PALSAR data [4, 5]. While this earlier research has provided some valuable insights into the complementarity and interoperability of the Landsat and PALSAR sensors for forest mapping, the forest/non- forest (F/NF) classifications were typically obtained independently for the optical and radar datasets. The Landsat time series, extending back to 1972 in these earlier studies, was processed according to the methods developed within Australia’s National Carbon Accounting System (NCAS) [6, 7], which was originally developed in the early 2000s. On the other hand, the newer ALOS- PALSAR dataset (2007 onwards) was processed using more recent processing tools and training data [4, 5]. In particular, the use of different training sets for the SAR and optical F/NF classifiers typically leads to discrepancies in the classifications that are not necessarily sensor-related (chemical sensing for optical data vs. structural/dielectric sensing for SAR), which can potentially affect the resulting interpretation of sensor interoperability when comparing the time series of optical and radar F/NF data [5]. Ideally, an operational forest monitoring system should provide a consistent and repeatable framework within which all datasets are subjected to the same pre-processing and analysis methods, so as to ensure consistency of the forest information outputs. In support of GEO-FCT, the focus of this work is on addressing the issue of interoperability of PALSAR and Landsat data in mapping forest extents. Within the framework of a multi-temporal multi-sensor forest information system, standardized analysis techniques from the NCAS program are applied to a mixed time series of optical and SAR data. The resulting forest information maps are then