Mitchell A.L.,University of New South Wales |
Tapley I.,CRC SI |
Milne A.K.,University of New South Wales |
Williams M.L.,CRC SI |
And 5 more authors.
Remote Sensing of Environment | Year: 2014
This paper addresses the shortfall in L-band SAR data availability for the purpose of extracting spatially explicit information on forest cover, and the capacity to fill the gap using shorter wavelength C-band data. Specifically, comparisons are made of forest/non-forest (F/NF) extent derived from independent classification of ALOS PALSAR and RADARSAT-2 data acquired over Tasmania. Focussing on a temperate forest landscape, it was demonstrated that C-band SAR data can be used interchangeably with L-band data to produce a comparable estimate of forest/non-forest cover. Only partial interoperability is achieved however, given the limitations of dual polarisation C-band SAR in discerning forests of different growth stage and biomass.Ambiguities in F/NF status were more prevalent in the C-band classification, despite inclusion of topographic (surface elevation and slope) and textural features in the training dataset, with the key observations as follows: (i) Reduced dynamic range and greater overlap amongst F/NF classes; (ii) Similarities in volume scattering from harvested/regrowth eucalyptus areas and background native forest; (iii) Confusion between young pine plantation and harvested/regrowth areas due to comparable roughness; (iv) Less variation in backscatter and poorer separation of intact and managed eucalyptus forest; and (v) Reduced capacity for discrimination of forest types. In almost every case, the use of L-band data was preferable. The one exception was the limited separation of young pine plantation and harvested/regrowth areas in both C- and L-band data. A similar level of performance was achieved in the discrimination of mature pine plantation, and between young eucalyptus plantation and harvested/regrowth.The findings were restricted to single-date classification of C- and L- band data. The potential to extend a time-series of L-band observations over forest using dense time-series (i.e., intra-annual) C-band observations acquired in dual or quad polarisation mode, warrants further investigation. Where a positive trade-off exists between the benefits and costs of integrating these data, multi-frequency (e.g., C- and L-band) and multi-sensor approaches (e.g., SAR and optical) are a viable way forward for operational forest monitoring and carbon accounting. © 2014 Elsevier Inc.
Lehmann E.A.,CSIRO |
Caccetta P.,CSIRO |
Lowell K.,University of Melbourne |
Mitchell A.,University of New South Wales |
And 4 more authors.
Remote Sensing of Environment | Year: 2015
In light of the growing volumes of remote sensing data generated by multiple space-borne platforms, integrated multi-sensor frameworks will continue to generate a significant interest in the frame of international forest monitoring initiatives. This work investigates the interoperability of synthetic aperture radar (SAR) and optical datasets for the purpose of large-scale and operational forest monitoring. Using a discriminant technique known as canonical variate analysis, we investigate the level of discrimination (between forest and non-forest training sites) achieved by different datasets, thereby providing an assessment of complementarity between Landsat data and SAR data acquired at C-band (RADARSAT-2) and L-band (ALOS PALSAR), as well as related texture measures. Spatio-temporal methods developed as part of Australia's Land Cover Change Program (an established forest mapping and carbon accounting scheme operating at continental scale) are subsequently used for the integration of Landsat and (segmented) PALSAR data. To highlight specific operational aspects of the multi-sensor framework, this approach is demonstrated over the Australian state of Tasmania (approximately 6.8. million. ha), one of several national demonstrator sites defined by the Forest Carbon Tracking task of the Group on Earth Observations (GEO-FCT). In terms of complementarity, the combination of Landsat and L-band SAR data is found to provide most of the forest discrimination, while texture information and single-date C-band SAR data are found to provide only limited additional discrimination improvement in the frame of the considered monitoring system. The interoperability of optical and SAR data is assessed by comparison of forest maps resulting from the spatio-temporal processing under different scenarios, including: i) Landsat-only time series, ii) PALSAR-only time series, and iii) mixed Landsat-PALSAR time series. A comparison of the single-date optical and SAR-based forest classifications indicates a good agreement over Tasmania, with some bias towards forest in the PALSAR classifications. Significant differences are evident when considering the case of forest conversion (deforestation and afforestation) over large areas, thereby compromising the full interoperability of SAR and optical data within the framework of Australia's carbon accounting system. © 2014 Elsevier Inc.