Joint processing of Landsat and ALOS-PALSAR data for forest mapping and monitoring

Transcription

1 IEEE Transactions on Geoscience an Remote Sensing 1 Joint processing of Lansat an ALOS-PALSAR ata for forest mapping an monitoring Eric A. Lehmann, Member, IEEE, Peter A. Caccetta, Zheng-Shu Zhou, Member, IEEE, Stephen J. McNeill, Xiaoliang Wu, an Anthea L. Mitchell Abstract Recent technological avances in the fiel of raar remote sensing have allowe the eployment of an increasing number of new satellite sensors. These provie an important source of Earth observation ata which as to the currently existing optical atasets. In parallel, the evelopment of robust methos for global forest monitoring an mapping is becoming increasingly important. As a consequence, there is significant interest in the evelopment of global monitoring systems that are able to tae avantage of the potential synergies an complementary nature of optical an raar ata. This paper proposes an approach for the combine processing of Lansat an ALOS- PALSAR ata for the purpose of forest mapping an monitoring. This is achieve by incorporating the PALSAR ata into an existing, operational Lansat-base processing system. Using a irecte iscriminant technique, a probability map of forest presence/absence is first generate from the PALSAR imagery. This SAR classification ata is then combine with a time-series of similar Lansat-base maps within a Bayesian multi-temporal processing framewor, leaing to the prouction of a time series of joint raaroptical maps of forest extents. This approach is applie an evaluate over a pilot stuy area in north-eastern Tasmania, Australia. Experimental outcomes of the propose joint processing framewor are provie, emonstrating its potential for the integration of ifferent types of remote sensing ata for forest monitoring purposes. Inex Terms Forest mapping an monitoring, remote sensing, ALOS-PALSAR, Lansat TM, multi-temporal processing, ata fusion, iscriminant analysis, conitional probability networ. T I. INTRODUCTION HE recent launch of several international initiatives emonstrates the increasing nee for global forest monitoring systems that provie estimates of carbon fluxes for the purpose of carbon accounting. Examples inclue the Unite Nations REDD program (Reucing Emissions from Deforestation an Forest Degraation, Manuscript receive September 24, 2010; revise June 17, 2011 an August 08, 2011; accepte September 11, E. A. Lehmann is with the Commonwealth Scientific an Inustrial Research Organisation (CSIRO), Division of Mathematics, Informatics an Statistics, 65 Brocway Roa, Floreat WA 6014, Australia (phone: +61 (0) , fax: +61 (0) ). P. A. Caccetta, Z.-S. Zhou, an X. Wu are also with the CSIRO Division of Mathematics, Informatics an Statistics in Floreat, Western Australia. S. J. McNeill, is with Lancare Research, Lincoln, New Zealan. A. L. Mitchell is with the Cooperative Research Centre for Spatial Information (CRC-SI), School of Biological, Earth an Environment Sciences, the University of New South Wales, Syney, Australia. initiate in 2008, an the Global Earth Observation System of Systems (GEOSS) evelope since 2005 by the Group on Earth Observations (GEO, The GEO tas on Forest Carbon Tracing (GEO-FCT, set up in 2008, aims to emonstrate that coorinate Earth observations can provie the basis for reliable forest information services of suitable consistency an accuracy to support the operation of global forest carbon estimation an reporting systems uner the Unite Nations Framewor Convention on Climate Change (UNFCCC, In response to the GEO-FCT tas, the Australian Department of Climate Change an Energy Efficiency (DCCEE) has recently launche the International Forest Carbon Initiative (IFCI, which aims to increase international forest carbon monitoring an accounting capacity in accorance with emerging international reporting an verification requirements. Optical-base lan cover monitoring proucts have been available for many years in several parts of the worl. Examples inclue the CORINE lan atabase supplie by the European Environmental Agency ( the National Lan Cover Database in the USA ( an the Global Observation of Forest an Lan Cover Dynamics program (GOFC-GOLD, In Australia, the Commonwealth Scientific an Inustrial Research Organisation (CSIRO) has contribute to the evelopment of the National Carbon Accounting System Lan Cover Change Program (NCAS-LCCP) [1], [2], in collaboration with the DCCEE an other partners. One component of this system offers the capability for fine-scale continental mapping an monitoring of the extent an change in perennial vegetation using Lansat satellite imagery, allowing for an effective estimation of greenhouse gas emissions from lan use an lan use changes [3]. This program currently uses over 7000 Lansat MSS/TM/ETM+ scenes resample to a spatial resolution of 25m for nineteen time epochs since 1972 over continental Australia; it is now upate annually, maing it one of the largest operational lan cover monitoring programs of its in in the worl. An important aspect in future evelopments of the NCAS framewor uner the International Forest Carbon Initiative involves taing avantage of new technologies in the area of space borne Synthetic Aperture Raar (SAR) sensors. While

2 IEEE Transactions on Geoscience an Remote Sensing 2 the availability of clou-free Lansat images oes not normally represent an issue for most of Australia, access to suitable optical atasets cannot always be guarantee in all geographical areas [4], an especially those locate in tropical regions or more southern latitues such as Tasmania. For instance, the extent of clou-affecte areas over Tasmania in the nineteen Lansat mosaics of the NCAS archive is about 26% on average, ranging from 4% for the best year to 100% for the worst. 1 Missing optical ata ue to clou cover is therefore a factor of particular importance in the evelopment of global lan cover monitoring systems, which nee to operate uner a variety of environmental an atmospheric conitions. In clou-affecte regions, the use of SAR ata thus represents a valuable alternative to exten the current monitoring capabilities of existing optical systems [5]. A further benefit is the increase temporal coverage resulting from aitional an/or complementary observations provie by SAR sensors. An improve accuracy in lan cover mapping can also be expecte when the complementary information from SAR ata is consiere jointly with existing optical atasets [6], [7]. Key aspects such as satellite ata interoperability (obtaining the same thematic results with ifferent sensors) an complementarity (aing thematic value by using two or more sensors) are therefore of specific interest in the evelopment of joint raaroptical processing systems. In the current literature, many stuies can be foun on the use of SAR ata for forest mapping an monitoring. In particular, several papers investigate ifferent approaches for the extraction of forest mapping information from JERS-1 an ERS-1/ERS-2 ata in tropical regions, mainly for the purpose of etecting eforestation events [8], [9]. Examples of SARbase methos for lan cover iscrimination inclue the use of ecision tree classifiers [10], spectral mixing moeling techniques [11], support vector machine [12], [13], an hybri learning classifiers [14]. Other existing wors also consier the use of ifferent SAR atasets such as RADARSAT [15], ALOS-PALSAR [16], an SIR-C/X-SAR [17]. In [18], an empirical investigation using multi-frequency SAR imagery (C-, L- an P-ban) acquire from airborne platforms highlights the usefulness of longer-wavelength SAR ata for forest monitoring an biomass estimation, thus paving the way towars monitoring systems using satellite ata. Driven by the emergence of these new SAR technologies, further stuies have also been recently carrie out to investigate the synergetic potential of SAR an optical ata; such comparative stuies can be foun, e.g., in [5] an [19]. Several articles also propose various methos for a combine processing (ata fusion) of SAR an optical atasets for forest mapping an lan cover classification, as escribe for instance in [20], [13], [6] an [7] (an references therein). The use of ata-level fusion techniques is base on the assumption that the ifferent atasets uner consieration are coincient in time. In the frame of the present research, a 1 NCAS aims to select the best available (most clou-free) Lansat imagery in each epoch for forest mapping/monitoring across the Australian continent. series of comparative stuies have also been carrie out on the forest iscrimination capabilities of combine SAR an optical ata uner the assumption of temporal coincience. These stuies are base on linear iscriminant analysis an maximum lielihoo classification for the same type of ata as use in the present wor (ALOS-PALSAR an Lansat TM), an the results are reporte in [21]. The main outcomes inicate that: 1) consiering the SAR an optical sensors jointly provies a better forest classification than either use inepenently, an 2) the respective contribution of each of the optical an SAR bans to the separation of ifferent types of forest an nonforest lan covers varies significantly. Reaers are referre to [21] for more information on the specific forest iscrimination properties of the ata use in the present wor (uner the assumption of temporally coincient atasets). The concepts escribe in this paper focus on a case where the assumption of coincient SAR an optical ata cannot be mae, ue to constraints relate to clou cover an ata source availability, for instance. This wor thus investigates the interoperability of the SAR an optical sensors, where one sensor might be use to generate the information normally provie by the other. Current an future environmental monitoring systems will have the ability to raw on the (spatial an temporal) information provie by many satellite sensors operating at ifferent wavelengths an with ifferent acquisition times, an this paper presents a methoology for the integration of such multi-temporal ata into a consistent probabilistic framewor. One avantage of the propose approach is that it inepenently combines the thematic information erive from the SAR an optical ata at ifferent times. This allows the interoperable use of one ataset when no information is available from the other, as in the case of clou-affecte Lansat imagery for instance. Two or more ata streams (time-series) from ifferent sensors can thus be efficiently integrate over time for improve forest monitoring capabilities (compare to the use of a single stream). In support of the GEO-FCT tas of forest mapping an monitoring, the propose methoology is emonstrate here by consiering a time series of ALOS-PALSAR an Lansat images over a test site in Tasmania. These ata, together with an inepenent reference ataset, are escribe in Section II. Joint SARoptical processing is achieve by first converting the SAR an Lansat ata into maps of forest probabilities. A irecte iscriminant technique (canonical variate analysis, CVA) is use for this purpose, as explaine in Section III. The resulting maps of forest probabilities are then processe jointly within the Bayesian framewor of a conitional probability networ (CPN), which is escribe in Section IV. Finally, Section V conclues this paper with a iscussion of the main results presente in this wor. A. Pilot stuy area II. DATA AND STUDY AREA This wor represents a pilot stuy carrie out for a 66m 50m emonstration area in the Ben Lomon region in north-

3 IEEE Transactions on Geoscience an Remote Sensing 3 HH: B -7.6 B HV: B B HH-HV: 2.6 B 11.6B Fig. 1. ALOS-PALSAR ataset for 2008, HH/HV/HH-HV in R/G/B (σ in B). Left: mosaic over Tasmania, with a box showing the 66m 50m stuy area. Right: PALSAR ata for the area of interest (centere on the latitue/longitue coorinates S/ E); white pixels correspon to the shaow mas. The color bars show the scales for the right-han image. eastern Tasmania, Australia (see Fig. 1, left). The centre of this stuy site is locate at the latitue coorinate S an longitue coorinate E. This area inclues one of Australia s national calibration sites (Mathinna region) efine in the framewor of the GEO-FCT initiative. It contains a variety of lan covers incluing rainforests, wet an ry eucalypt forests, exotic plantations for silviculture, agricultural lan, as well as other cleare lan (eforestation) an urban areas. Significant topographic variations can be foun across the stuy site, with the terrain elevation varying between 85m an 1510m above sea level. B. ALOS-PALSAR ata The ALOS-PALSAR ata use in this wor was acquire at L-ban (~23.6cm wavelength) in fine-beam ual-polarization moe (HH an HV), in an ascening orbit an with the efault off-nair angle of The SAR ata in the stuy area consists of a (croppe) mosaic of two PALSAR scenes, namely the scenes with path/row 381/6340 in the west (acquire on October 04, 2008) an path/row 380/6340 in the east (acquire on September 19, 2008). The single-loo complex ata (SLC level 1.1) was pre-processe accoring to the following steps: multilooing (loos in range an azimuth, respectively) resulting in a pixel size of 29.8m 25.1m 2. specle filtering by means of aaptive 5 5 Lee filter [22] 3. raiometric calibration [23], [24], using a igital elevation moel (DEM) with a 25m cell resolution, an raiometric normalization (moifie cosine moel) [25] 4. geocoing to 25m pixel size (to match the spatial resolution of Lansat) using the 25m DEM, an calculation of the shaow areas 3 5. terrain illumination correction [27], using the 25m DEM 6. creation of a two-scene mosaic (graient mosaicing) 7. masing of the ata using the shaow mas. This processing sequence generates a mosaic of orthorectifie, terrain-correcte an raiometrically calibrate PALSAR scenes at a resolution of 25m (Fig. 1, right), with values ranging from -22.5B to -7.6B in the HH channel, an from -30.7B to -12.8B in the HV channel (99% clip, i.e., excluing the first an last 0.5% of the ata). The terrain illumination correction in Step 5 is necessary to compensate for illumination ifferences ue to the local variations in topography an the viewing geometry of the SAR sensor. In areas of steep terrain, these factors typically lea to forwarfacing slopes (facing towar the SAR sensor) appearing brighter an bacwar-facing slopes (facing away from the sensor) appearing arer. These effects limit the ability to extract relevant thematic information from the ata an must therefore be correcte. Application of the terrain correction algorithm [27] to the SAR ata was necessary as the SARscape prouct generate following Steps 3 an 4 still containe significant terrain-relate artifacts (espite the correction for scattering area carrie out as part of the raiometric calibration in Step 3, see [24]). C. Lansat TM ata an erive proucts 1) Data The optical ata use in this wor was obtaine from the existing archive of calibrate Lansat MSS/TM/ETM+ images prouce as part of the NCAS-LCCP system. Within this 2 Steps 1, 2, 3, 4 an 6 in this processing sequence were carrie out by means of the SAR processing software SARscape version ( 3 In the stuy area, regions of SAR layover account for less than 1% of the pixels, an are therefore not treate separately in this wor; in an operational setting, further processing of layover pixels can be implemente as escribe, for instance, in [26].

4 IEEE Transactions on Geoscience an Remote Sensing 4 Fig. 2. Overview of Lansat ata an forest/non-forest prouct. Left: 2008 Lansat TM ataset for the consiere stuy area (bans 5/4/2 in R/G/B). Ben Lomon is the prominent rocy feature locate in the centre-south of the image. Right: corresponing single-ate forest probability map from the NCAS program for 2008, with forest probabilities of 100% shown in green, 0% in blac, an values in between in yellow. In both images, white areas inicate regions mase out ue to missing ata (clous an sensor eficiencies). processing framewor, all images nee to be properly coregistere an calibrate in orer to achieve meaningful lan cover change information an to minimize errors relate to misregistration. All Lansat scenes are processe accoring to strict quality stanars [28] using the following steps: 1. orthorectification to a common spatial reference, using a rigorous earth-orbital moel an a cross-correlation feature matching technique [2] 2. top-of-atmosphere reflectance calibration (sun angle an istance correction) [29] 3. correction of scene-to-scene ifferences using biirectional reflectance istribution functions [30] 4. calibration to a common spectral reference using invariant targets [31] 5. correction for ifferential terrain illumination [32] 6. removal of corrupte ata such as regions affecte by smoe, clous an sensor eficiencies 7. mosaicing of the iniviual Lansat scenes into 1:1,000,000 map sheets. Key aspects of these processing steps are iscusse in [28], [2], while full operational etails are given in [33]. This process was use to generate a calibrate an orthorectifie time series of Lansat mosaics over the Australian continent for a total of nineteen time slices between 1972 an For 2008, the optical ata over the consiere stuy area is a mosaic of two Lansat scenes acquire on January 14 an February 24, respectively (Fig. 2, left). 2) Forest probability maps The main tas initially unertaen as part of the NCAS program was the mapping of forest presence/absence (forest/non-forest, F/NF) from the Lansat time series. Continental maps of the forest extents across Australia were generate using a process involving the following steps: 1. ientification of stratification zones, within which lan cover types have similar spectral properties an are thus consiere homogenous 2. creation of a base forest cover probability image by: a. specification of a single-ate classifier for each stratification zone, using groun an satellite ata b. ientification of (soft) threshols to etermine the ecision bounaries between F/NF classes, leaing to a probability of forest cover for each pixel 3. creation of forest cover probability images for other epochs by means of a matching process [2], using the base probability image as reference 4. refinement of the single-ate probability images using a spatial-temporal moel for classification (Bayesian conitional probability networ [34], [35]). In Step 2, the single-ate classifier is base on a irecte iscriminant technique calle canonical variate analysis (CVA) [36]. This metho etermines the linear combination of image bans proviing the best separation between the F/NF classes by maximizing the ratio of between-class to within-class variance (more etails are provie in Section III.A). The image on the right in Fig. 2 shows the NCAS-LCCP forest probability image erive from the Lansat ata for the area of interest in The subsequent use of a multi-temporal conitional probability networ (CPN) in Step 4 improves the forest mapping accuracy an temporal consistency [1]. A comparison of alternative single-ate classifiers was consiere in [37] an emonstrates that the resulting ifferences in the forest presence/absence classifications are negligible after processing the single-ate results with a CPN. As will be shown in Section IV, the CPN framewor can also be use for the purpose of combining SAR an optical atasets into a single an consistent framewor for forest mapping. The present wor consiers the integration of the PALSAR ataset within this operational Lansat-base methoology (NCAS-LCCP). Some issues can be expecte to arise when substituting ata from a ifferent sensor into this existing (legacy) framewor. In particular, the results will be influence by the specific biases introuce by each sensor, as

5 IEEE Transactions on Geoscience an Remote Sensing 5 accuracy (best 80% of chec GCPs) in the orer of 34m (stanar eviation of 13.7m) with respect to 1:100,000 topographic maps. As the DEM use in this wor is also erive from a state topographic map, an given the lac of systematic spatial isplacement between the PALSAR an Lansat ata mentione above, this geolocation accuracy was eeme to provie a sufficient corresponence between the DEM (use in the ata pre-processing steps) an each of the SAR an optical ata for the purpose of this pilot stuy. Fig. 3. Overview of broa vegetation communities over the stuy area (TASVEG map). Circle marers inicate the locations of the training sites selecte for the PALSAR forest/non-forest classification (see text for etail). iscusse in [38]. SAR respons to forest structure an ielectric content, while optical sensors rely on a more biochemical response. Discrepancies in raar an optical classifications are therefore incurre ue to the way in which raar an optical sensors see the lan surface an cover. 3) Coregistration with PALSAR ata For a proper integration of the thematic information extracte from both atasets, it is crucial for the Lansat an PALSAR ata to be accurately coregistere with each other. The registration accuracy between these ata was establishe with the use of a graient cross-correlation technique [39], which provies a sub-pixel coregistration assessment by etecting an matching features in both images over a winow containing many pixels. The isplacement between the features in the two images is etermine by iteratively searching for the shift (istance an irection) that maximizes the cross-correlation between the ata winow in each image. To evaluate the coregistration accuracy between the PALSAR an Lansat ata, a total of 299 image features were selecte over the stuy area so as to be uniformly istribute spatially, in both flat an mountainous regions. The average isplacement between these features was 0.69 pixel (stanar eviation of 0.34 pixel) with 98% of the shifts smaller than 1.5 pixels (at 25m pixel size). Also, the irections of the features isplacements i not show any apparent signs of systematic coregistration error between the two atasets. Base on these results, the coregistration accuracy between the PALSAR an Lansat images in the region of interest was thus consiere satisfactory for the purpose of joint SARoptical processing. An assessment of the absolute registration accuracy of the rectifie Lansat imagery against inepenent groun control points (GCPs) was carrie out as part of the NCAS program [33]. This analysis emonstrate a Lansat registration D. TASVEG reference ata TASVEG is a Tasmania-wie vegetation map prouce by the Tasmanian Vegetation Mapping an Monitoring Program within the Department of Primary Inustries an Water ( [40]. This map comprises 154 istinct vegetation communities mappe at a scale of 1:25,000, an provies a single reference ataset that can be use for a broa range of management an reporting applications relating to vegetation in Tasmania. Aerial photographic interpretation is the primary ata collection metho for TASVEG, generally at a scale of 1:25,000 an 1:20,000, or 1:42,000 where a better resolution is unavailable. The photo-interpretation is assiste by fiel verification of representative polygons, an by the use of various vegetation, ecology an geology texts as well as maps containing species information. The resulting line wor is then vectorize to form a igital vegetation layer with a positional accuracy of 20m or less, an with a smallest polygon representation of about 4ha. TASVEG is continually revise an upate to reflect changes in the natural environment. The present wor raws on the information containe in Version 2.0 of this prouct, which was release on February 19, In the region of interest (Ben Lomon bioregion), fiel wor for revision mapping commence in March 2005 an was complete in October Over this area, TASVEG shows that most of the lan cover belongs to one the following main classes: ry eucalypt forest an woolan wet eucalypt forest an woolan rainforest an relate scrub non-eucalypt forest an woolan plantation (silviculture) regenerating cleare lan urban an water areas other (non-forest), such as agriculture an alpine shrubs. Fig. 3 shows the istribution of these main vegetation communities within the stuy area. In terms of recentness, spatial resolution an coverage, TASVEG represents the best available groun-truth prouct for Tasmania at present time. In this wor, the TASVEG map (together with other groun an satellite ata) is use in the process of selecting training an valiation sites for the forest/non-forest classifications. III. SAR FOREST CLASSIFICATION Several methos for F/NF classification using SAR ata can be foun in the literature (see references in Section I) an coul be potentially applie to the ata consiere here. The

6 IEEE Transactions on Geoscience an Remote Sensing 6 present wor aopts the NCAS-LCCP approach to forest classification, which involves the following steps (escribe in more etail in the following subsections): 1. selection of training sites for the classes of interest 2. use of CVA to erive linear iscriminant functions (canonical vectors) 3. selection of separation inex an soft classification threshols using irecte contrasts 4. pixel-base classification of the SAR ata using the erive inex an threshols. To the best of the authors nowlege, this paper represents the first attempt to apply CVA to the problem of F/NF classification of SAR ata (see iscussion an references in Section I for relate classification methos) an emonstrates here that routine processes of the operational NCAS forest mapping system can be reaily applie to SAR ata. A. Discriminant analysis CVA is a metho to fin uncorrelate linear combinations of variables (canonical vectors, CVs) with maximum betweenclass variability relative to the within-class variability of a set of training ata. Consier the case where a total of n i (i = 1, 2,..., C) training sites have been selecte for each of C ifferent classes. In class i, the j-th training site is represente by the observation vector y ( j = 1, 2,..., n i ) which contains the average values of the B image bans compute over the site s p pixels: y = ( 1/ p ) = x 1, where x is a column vector of the B image bans for pixel, an p represents the number of pixels in the j-th training site of class i. With n i j = 1 m = ( 1/ ) y representing the mean of the i-th class, i n i the between-class variation matrix B is then efine as C 1 T B = n i ( mi m)( mi m) (1) N 1 i= 1 m = m (2) C n i N i= 1 i where N = n 1 + n n C. With E{ } representing the statistical expectation operator, the within-class variation matrix W (which is the same for all i an j) is efine as T W = E{( y m )( y m ) }. (3) i i Canonical variate analysis involves fining the vectors f ( = T T 1, 2,..., B) maximizing the ratio µ = f Bf / f Wf of between-class to within-class variation; the variable µ is referre to as the canonical root for vector f. This approach leas to the following generalize eigenvalue problem: ( B µ W) f = 0, for 1, 2,..., B, (4) = which nees to be solve for µ an f subject to T 1 if i = j f i Wf j = (5) 0 otherwise. The conition in Eq. (5) means that the canonical root represents the magnitue of the between-class variation in the irection f of maximum between-class variation. This process can be seen as projecting each class ata onto a line such that the variance of the means of each class on the line is as large as possible relative to the average variance of the observations within each class. Successive projections are efine similarly, subject to being orthogonal (uncorrelate) with previous ones. The canonical roots µ then provie a measure of separation between classes. The first few canonical vectors f (with the largest separation) can be use to generate the linear combinations s T µ = f y (CV scores), which provie a low-imensional ata reuction for escribing class ifferences. For instance, a plot of the vector [ s s, 1 2 ]T enables a visual representation of the ata nown as CV plot. Reaers are referre to [36] for further information about the CVA technique. B. SAR classification parameters In this wor, the concepts escribe in Section III.A are applie with C = 2 classes (forest/non-forest) an B = 2 PALSAR bans (HH an HV). A total of N = 160 training sites were selecte across the whole area of interest in such a way as to provie representative samples for each of the broa vegetation communities in the stuy area; the locations of these training sites are shown as circle marers in Fig. 3. Each site contains roughly 150 to 200 pixels to provie consistency when computing the sites means. In orer to facilitate the attribution to one of the available classes, the sites were selecte on the basis of information provie by a combination of aerial photographs, high-resolution imagery (Spot an Ionos ata), the TASVEG ata an Lansat imagery. Base on the selecte training sites an the PALSAR ata, CVA provie the following two canonical scores: s = 76 HH HV 1 (6) s2 = 1.38 HH HV (7) where HH an HV represent the PALSAR ata (expresse in B) average over the j-th training site in the i-th class. It can be seen that the CV analysis here results in s 1 being largely efine by the cross-polarization component HV, while s 2 represents a contrast between the co- an the crosspolarization components. Fig. 4 shows a plot of the vectors [ s1, s2 ] T for all training sites in the resulting CV space. Note that this plot uses colors (an shapes) to represent various subclasses of the two main forest an non-forest groups; this is one for illustration purposes only, an this information (broa

7 IEEE Transactions on Geoscience an Remote Sensing 7 to be classifie as either forest, non-forest or uncertain cover on the basis of their inex value. The uncertain cover category contains regions that are not fully separable in the given image, an prouce a forest probability value between 0% an 100%. The forest probability P (in %) for a given pixel in the SAR image is therefore compute accoring to the following formula: 100 I I P = 100 I thr,f I 0 thr,nf thr,nf if I if I if I I thr,nf I thr,f < I thr,nf < I thr,f (9) Fig. 4. Plot of training sites means in CV space (CV plot). The ashe lines inicate the separation between the forest an non-forest classes achieve with the selecte classification inex an soft threshols (see text for etail). vegetation communities) is not actually use in the classification algorithm. This plot shows that most of the forest sites can be clearly separate from the non-forest samples, except for the plantation sites (in green). These sites were selecte in areas where forestry activities are taing place, an inclue samples from both harveste as well as growing/mature plantations, thus explaining the presence of these sites in both the forest an non-forest clusters. Using the plot in Fig. 4, a spectral inex for the forest/nonforest separation is achieve by contrasting two ifferent subsets of training sites (contrast-irecte CVA, [41]), where the irection of maximum separation between the two subsets is etermine in a manner similar to the erivations in Section III.A. This was achieve here by contrasting all the forest sites against a sub-group of non-forest sites obtaine after removing all the clearly separable non-forest sites, namely all sites with s in Fig. 4. This proceure le to the following separation inex, which essentially correspons to the linear combination of spectral bans allowing the best iscrimination between the selecte sub-groups of F/NF sites using the available PALSAR ata: I = 5.36 HH HV. (8) Here, the separation inex I can be interprete physically as showing that the maximum class separation is achieve by a strong epenence on the HV polarization (two orers of magnitue stronger than the HH coefficient), which is liely to result from volumetric scattering in forest targets. This strong epenence of the SAR F/NF classification on HV (an the limite influence of HH) corroborates the results obtaine in previous literature stuies (see, e.g., [42]). Further information about the forest iscrimination capabilities of the HH an HV channels for ifferent cover types can also be foun in [21]. When classifying the PALSAR image, two (soft) threshols were selecte to ientify certain forest (I thr,f ) an certain non-forest (I thr,nf ) regions in the CV space, allowing the pixels where I is the value of the separation inex for pixel, as compute accoring to Eq. (8). Initial values for the soft threshols were ientifie by inspection of the CV plot for the training sites (Fig. 4). These threshols were subsequently fine-tune manually by visual inspection of the classification results so as to minimize the level of commission error (nonforest pixels assigne to the forest category) an omission error (forest pixels assigne to the non-forest category) over the whole SAR image. This process is similar to that use operationally in the frame of the Lansat-base NCAS monitoring program (see [33] for further etail). In this wor, the resulting soft threshols were set as follows: I thr,nf = 2470 an I thr,f = The ashe lines in Fig. 4 show the corresponing iscrimination lines (orthogonal to the irection of maximum separation) in CV space. C. SAR forest classification Base on the selecte separation inex an threshols, a pixel-base classification is applie to the PALSAR ata, with results presente in Fig. 5. The plot on the left shows the resulting SAR forest probability map, which is comparable to the Lansat-base NCAS map (Fig. 2, right). The secon image in Fig. 5 shows a comparison of the SAR classification against the TASVEG ata. For this purpose, both atasets were collapse into binary maps of F/NF pixels. For the SAR ata, this was achieve by thresholing the forest probability map at the 50% probability level. For TASVEG, pixels from the non-eucalypt forest, rainforest, wet/ry eucalypt forest an plantation classes were groupe into a single forest class, with the remaining pixels labele as non-forest. The two resulting F/NF atasets are then isplaye in a re/green composite image, proviing a visual valiation of the results that shows the pixels in agreement in blac an yellow, an isagreement pixels in re (labele as forest in TASVEG) an green (labele as forest in the SAR classification). This composite image inicates two main sources of isagreement. First, large areas of inconsistency can be seen as re patches across the stuy area. As seen in Fig. 3, these regions all correspon to plantations, where the TASVEG ata is interprete as being forest regarless of the state of the plantation stans when the SAR imagery was acquire. The SAR ata here correctly ientifies the re areas in Fig. 5 as

8 IEEE Transactions on Geoscience an Remote Sensing 8 Fig. 5. Single-ate SAR forest classification results for Left: classification image of forest probabilities (blac: 0%, green: 100%, yellow: other values). Right: comparison with TASVEG (re/green composite image), with SAR F/NF results in the green layer an TASVEG F/NF ata in the re layer. White areas correspon to the SAR shaow mas. Ben Lomon is the prominent (quasi-rectangular) feature locate in the bottom-centre of the images. TABLE I CONFUSION MATRIX BETWEEN TASVEG AND THE SAR CLASSIFICATION (SINGLE-DATE F/NF RESULTS). TASVEG (%) forest non-forest forest PALSAR (%) non-forest Values obtaine using a total of 5,006,048 pixels (273,952 pixels in shaow areas mase out). TABLE II CONFUSION MATRIX BETWEEN TASVEG AND THE LANDSAT CLASSIFICATION (SINGLE-DATE F/NF RESULTS). TASVEG (%) forest non-forest forest Lansat (%) non-forest Values obtaine using a total of 5,089,413 pixels (190,587 pixels mase out ue to clous an sensor eficiencies). being harveste (non-forest), an the corresponing pixels thus cannot be consiere as being erroneous. A secon type of iscrepancy can be foun in higheraltitue rocy plateaus such as the Ben Lomon area (large blac an green area in the bottom-centre of the composite image) as well as the Mt. Barrow region (smaller green area in the centre-left of the image). Both areas culminate at an altitue of about 1300m, an the commission errors in the SAR classification result from the presence of alpine/subalpine heath an sege in these regions (an possibly also ue to the influence of surface moisture in these areas of low stature vegetation an bare groun at the time of the SAR acquisition). A CV analysis was performe with aitional test sites selecte in these areas an showe that these typically cluster together with sites belonging to ensely foreste regions, thus inicating that this particular type of vegetation (alpine shrubs an bushes) is not fully separable from the forest class using ual-polarization PALSAR ata only. Minor ifferences between TASVEG an the SAR classification map also appear ue to the ifferent mapping resolutions of these two proucts. For instance, thin features such as roas an water streams are not always prominent in the SAR imagery an are consequently classifie as forest. Other minor ifferences may also result from the ifferent acquisition times between the atasets (February 2009 revision for TASVEG, September/October 2008 for PALSAR). Outsie areas of forestry activities (plantations), the lan cover is mostly relate to either stable native forests or agricultural lan, an such ifferences can therefore be consiere minimal for the stuy region over the consiere time frame. A further valiation of the SAR classification results was establishe through a pixel-base comparison with TASVEG an inicates an overall agreement of 85.91%. Table I presents the confusion matrix between TASVEG an the SAR F/NF classification, showing the percentages of agreement an isagreement pixels between classes. The ifferences existing between the SAR-base classification results an other atasets offer an opportunity to further investigate the complementarity of SAR sensors for forest mapping. For completeness, the confusion matrix for the single-ate Lansat classification (Fig. 2, right) is also given here in Table II. Both results can be seen to be similar. The overall agreement between the Lansat classification an TASVEG (84.0%) is here slightly lower than that obtaine for the SAR ata (85.91%), mainly ue to a localize but relatively large area of forest commission error obtaine from the optical ata in the Ben Lomon region. This ifference between the optical an SAR mapping accuracy can here be explaine by the fact that the Tasmania-wie NCAS forest map is compare to the locally-traine (over the stuy area) SAR classification, which is use in this pilot stuy for the main purpose of emonstrating the conceptual integration of the PALSAR ata within the existing legacy system. Because of the implicit use of ifferent training ata, it is important to note that the SAR an optical results presente in Table I an Table II o not necessarily provie appropriate information regaring which sensor is best at extracting forest information over the stuy area. This in of etermination (i.e., which sensor is best)

9 IEEE Transactions on Geoscience an Remote Sensing 9 oes not represent the main focus of the present wor, an further research on the specific forest iscrimination properties of the PALSAR an Lansat ata is reporte in [21]. IV. JOINT PROCESSING OF SAR AND OPTICAL DATA In essence, two istinct approaches can be consiere for a combination of the available Lansat an PALSAR atasets. One approach is to consier fusion at the ata level, where the SAR an optical images are merge into a single ataset which is subsequently use as input to the single-ate classification algorithm. This approach, which represents the focus of most of the current literature on SARoptical ata fusion (see, e.g., [57], [19], [20]), is not irectly applicable in the present stuy since the PALSAR ata is only available for a single epoch (vs. nineteen for Lansat) an is not exactly coincient with the corresponing 2008 Lansat ata. This approach can also be problematic in areas where ata is missing from one of the consiere atasets (e.g., clou-affecte Lansat images). The secon approach consists in consiering the alternative ata source as an inepenent aition to the existing time series. The ata available at ifferent times (i.e., PALSARan Lansat-base single-ate forest classifications) can then be assimilate using a multi-temporal methoology, where each forest map is consiere as a iscrete observation of a continuous process (forest presence/absence) evolving over time. As escribe in the following, this approach is ieally suite to aress issues relate to the use of ifferent sensors an non-coincient acquisition times. The multi-temporal approach use in this wor was originally presente in [35] an is currently implemente operationally within the Lansatbase NCAS system (see, e.g., [1], [2]). Its use in conjunction with SAR ata can be seen as an extension of the NCAS framewor to aress the issue of multi-sensor integration. A. Multi-temporal processing Conitional probability networs provie a framewor allowing for the assessment an propagation of uncertainty in the classification of multiple ata sources of varying quality or accuracy [35]. To improve the forest mapping accuracy, a CPN uses a moel which incorporates temporal an spatial rules as well as error rates of the initial classifications. Let us assume that each image in the time series contains the same number K of pixels. For the -th image pixel ( = 1, 2,..., K) an m-th epoch (m = 1, 2,..., M) in the time series, let x enote the vector of B image bans, an { 1, 2, K, C} m enote the corresponing class label estimate by the singleate classifier base on the image ata x (an selecte training sites). The temporal information for the -th image pixel can be gathere by grouping the above variables into sets: X = x, x, K, x } an L = l, l, K, l }. Also, { 1 2 M l m m { 1 2 M let R represent the group of eight pixels spatially ajacent to pixel (8-pixel neighborhoo). Given the multi-temporal ata X, the quantities of interest are the true class labels l m of each pixel in the time series: L = l, l, K, l }. { 1 2 M As applie here, the conitional probability networ is a first-orer hien Marov moel (HMM) representing the relationships between the image ata an the unerlying process of interest. For the -th image pixel, this joint spatialtemporal moel is efine as follows (see [2], [35]): M p( X, L, L, R ) = Q Q Q Q (10) 1 m m m= Q = p( x l ) (11) Q = p ( l m l ) (12) 2 m 3 p( l m l ( 1) m Q = ) (13) Q = p ( l m R ) (14) 4 where for Q 3, p ( l 1m l 0m ) is simply efine as p ( l 1m ). In this moel, Q 2 represents error rates (sensor bias) that weigh the estimate lan cover class given the true labels, while Q 3 correspons to temporal rules inicating the lielihoo of transition between classes from one epoch to the next. In practice, Q2 an Q 3 are typically represente as contingency tables that are specifie or estimate from the ata available. The term Q 4 weighs a pixel s label towars that of the ominant neighborhoo labeling. Following [43], Q 4 is efine as p l R ) exp( α + βc ), where c is the number 3 ( m of pixels having the same label as pixel in its neighborhoo R from the previous iteration of the moel; α an β are user-specifie parameters set in this wor to 0 an 1, respectively. Assuming a uniform istribution of the class priors p ( l m ), Q 1 correspons to the lielihoo of the estimate class labels given the ata at each pixel. This quantity effectively correspons to the forest probabilities P m (compute for the -th pixel an the m-th epoch) erive in Eq. (9), i.e., p( x l = "F") = 1 p( x l = " NF" ) P. m m Applying Bayes rule an marginalization to Eq. (10), the probability p L X ) is compute, proviing the maximum ( lielihoo solution for the unobserve true class label given the (multi-temporal) ata. This is here achieve by using a cyclic ascent algorithm, iteratively cycling over all pixels until convergence on L is achieve. The etails of this algorithm can be foun in [34], [35]. The outputs from this spatialtemporal moel also represent probability images, one for each time slice, where the forest probabilities have been refine by the temporal rules an error rates efine in the moel, so as to provie temporally consistent estimates. Other important properties of the CPN approach can be summarize as follows: propagation of uncertainties in the inputs an calculation of uncertainties in the outputs prouction of har- an soft-ecision maps hanling of missing ata by using all available (spatial an temporal) information to mae preictions existence of well-evelope statistical tools for parameter estimation. This approach provies an efficient probabilistic framewor for combining isparate ata since variations in ata quality 4 m m m

10 IEEE Transactions on Geoscience an Remote Sensing 10 TABLE III CONFUSION MATRIX BETWEEN TASVEG AND MULTI-TEMPORAL CLASSIFICATIONS FOR TOP: OPTICAL-ONLY TIME SERIES. BOTTOM: COMBINED SAR-OPTICAL TIME SERIES. all-optical (%) TASVEG (%) forest non-forest forest non-forest SARoptical (%) TASVEG (%) forest non-forest forest non-forest Fig. 6. Comparison of the 2008 classification outputs from the spatialtemporal moel (CPN), with combine SARoptical forest probabilities in the green layer, an all-optical forest probabilities in the re layer (re/green image composite). an/or missing ata can be easily accommoate. As such, this approach thus also provies a useful methoology for the purpose of ata fusion of the Lansat an PALSAR atasets consiere in this wor. B. Joint SARoptical processing The concept of combine LansatPALSAR processing is emonstrate here by means of the following scenario. The NCAS-LCCP time series of Lansat-erive forest probability maps contains a total of nineteen images between 1972 an 2010 (non-uniformly space), incluing one in These images are use as an input to the CPN, resulting in a first (optical-only) time series of refine forest cover maps. In a secon experiment, the Lansat-base forest map for 2008 is remove from the time series an replace with the forest probability image obtaine on the basis of the PALSAR ataset (as escribe in Section III.C). This leas to a secon time series of moel outputs, corresponing to refine forest maps obtaine from a combination of the SAR an optical atasets into a single prouct. It must be emphasize here that the moel effectively blens the ata provie in each forest map with the information available from all the preceing an subsequent time slices. In the secon scenario, the resulting maps of forest extents for all time steps thus raw on the information containe in both the PALSAR an Lansat ata. The example presente here can be seen as representative of several situations where the complementary nature of SAR can prove avantageous. For instance, the SAR ata coul be use to fill in potential gaps in the existing optical time series, as in the case of missing ata ue to clous, sensor eficiencies, etc. Alternatively, entire maps of SAR-erive forest probabilities coul also be ae as new time slices so as to improve the temporal resolution of the original time series. ALOS- PALSAR ata is available from 2007 an with the anticipate launch of ALOS-2, the continuity of L-ban SAR ata an contribution to complementary forest monitoring is ensure. 1) Joint classification results Fig. 6 presents a comparison of the results for 2008 obtaine from the two scenarios consiere above. It shows a composite image containing the refine optical-only forest probability map in the re layer, an the combine PALSAR Lansat forest probability map in the green layer (threshole at the 50% probability level). A pixel-base comparison of the corresponing F/NF results inicates that these images are 95.8% in agreement. The main ifferences can be seen to originate from iscrepancies between the Lansat an PALSAR single-ate forest probability images (see Fig. 2 an Fig. 5). Most of these inconsistencies can be explaine by one of the following reasons: ifferent thematic information provie by the raar an optical sensors with respect to some groun features ifferences between the Tasmania-wie optimization of the Lansat-base classifier an the locally-optimize SAR classification genuine ifferences on the groun between the acquisition ates of the PALSAR an Lansat atasets (as observe in the ata itself, e.g., mature plantation harveste an cleare sometime between February an October 2008). For instance, it can be seen that the Lansat-base singleate forest probability map (Fig. 2, right) erroneously assigns most of the Ben Lomon highlan (south of the stuy area) to the forest class. Because this area is labele as forest in all the Lansat epochs, the SAR-base non-forest estimates for 2008 appear as spurious in the time series an are consequently relabele as forest in the joint classification. Base on this in of assessment, further research will focus on reucing such iscrepancies by improving the Lansat an PALSAR classification, an ientifying opportunities for improvements resulting from the complementarity of these two sensors. As a further valiation of the joint classification results, Table III shows the confusion matrices for the all-optical an SARoptical time series classifications for 2008, as compare to the TASVEG ataset. Here again, both results are very similar with agreement rates of 86.17% (all-optical) an 87.44% (SARoptical), proviing a further emonstration of the interoperable nature of the optical an SAR ata for the tas of forest mapping an monitoring.

11 IEEE Transactions on Geoscience an Remote Sensing 11 2) Summary of forest extents an change The main tas consiere in this wor is to provie estimates of the forest extents in each epoch as well as changes in forest cover over time (e.g., for carbon accounting purposes). The tables provie in the Appenix present a summary of these parameters compute over the stuy area for each epoch in the time series. Several observations can be mae on the basis of these results: 1. in each epoch (incluing the 2008 SAR image), the single-ate forest maps contain a significant amount of missing ata ( null column in Table IV) 2. the joint spatial-temporal classification preicts labels for the missing ata using all other available ata, an prouces forest maps with non-null probabilities for each pixel in each epoch of the time series (Table V) 3. the joint spatial-temporal classification reuces false transitions between classes (noise). Furthermore an most importantly, the areas of forest change provie in the F NF an NF F columns (conversion between the forest an non-forest classes) in Table V can be seen to be in goo agreement between the all-optical an SARoptical time series at the CPN output. Given the focus of this wor on the ynamics of eforestation an reforestation (lan cover change), this result represents a factor of crucial importance for the interoperability of the optical an SAR sensors in forest mapping an monitoring. V. CONCLUSION The wor presente in this paper aims to establish a methoology for the classification an the multi-temporal integration of ata acquire by ifferent sensors. It provies a realistic an promising emonstration of the integration of multi-temporal ALOS-PALSAR an Lansat imagery for forest mapping an monitoring, an escribes the ey processing steps involve in a potential approach to this tas. In particular, it is shown that the Bayesian concept of a conitional probability networ is able to integrate the PALSAR-base forest probability ata within a time series of similar Lansat-erive forest maps, proucing estimates of forest extents for each time epoch in the presence of missing SAR an optical observations. Within the context of a legacy optical system, this methoology is able to incorporate observations from a newer senor an prouce estimates comparable to those which woul otherwise be obtaine. Where consistent time-series observations cannot be obtaine from a single sensor, this feature may be avantageous for monitoring purposes. This wor provies some insight into the interoperability of the SAR an optical sensors, an the framewor consiere here coul thus also incorporate other optical (e.g., SPOT, CBERS, Sentinel 2) an SAR-base (e.g., Raarsat-2, TerraSAR-X, ENVISAT/ASAR) ata, although the impacts of sensor biases woul nee to be consiere an will be the topic of future stuies. Research wor on this project is ongoing an will focus on proviing further valiation results, as well as extrapolating the finings presente in this article to larger geographical areas (e.g., to the rest of Tasmania). One aspect of interest when up-scaling the propose methoology will be to etermine whether the use of SAR-specific stratification zones (as one in the Lansat-base system) also has the potential to improve the SAR forest classifications. Future wor will also incorporate further PALSAR forest probability maps (from 2007 onwars) an evaluate their influence (bias) on the time-series outputs. The capacity for improve forest extent mapping using multifrequency SAR ata (C- an X-ban) will also be investigate. APPENDIX The tables in this appenix provie a summary of forest extents an lan cover change over the stuy region (total area of 330,000ha) for each of the nineteen years in the consiere time series, from 1972 to All values provie in the tables are in hectares (ha). In the first three columns, Table IV presents the areas of forest (F), non-forest (NF) an missing ata (null), compute from the single-ate forest classification maps. The other columns summarize the areas of change (transition from one class to another) for each epoch compare to the preceing one. All values are for the Lansat ata except for the rows flagge with an asteris in the first column (year), which are relate to the SAR ata. The row labele 2009* contains the F/NF/null values obtaine from the Lansat image (no SAR ata available in 2009), but the last nine columns inicate transitions from the SAR-base classes in 2008* to the Lansat-base classes in Note that in the consiere stuy area, no ata is available for the years 1972 an 1977 (Lansat MSS imagery) ue to clou cover. Table V contains the forest extents an change values corresponing to the multi-temporal classification results. The left half of the table shows the results for the Lansat-only time series while the right half correspons to the combine SARoptical results. These values show how the null pixels in Table IV (an transitions to an from the null class) have been assigne to the forest or non-forest classes following the multi-temporal processing, using information from all available years. Table V also shows that similar hectare counts are obtaine for eforestation an re-forestation ( F NF an NF F ) from the all-optical an the SARoptical time series, which is of importance for a consistent assessment of lan cover change using interoperable SAR an optical sensors. ACKNOWLEDGMENT This project was conucte as part of the GEO-FCT initiative. The authors woul lie to acnowlege Prof. Tony Milne (Cooperative Research Centre for Spatial Information, Syney) an Dr. Alex Hel (CSIRO AusCover Facility, Canberra) for their help with the acquisition of the PALSAR ata, an Dr. Ian Tapley an Prof. Kim Lowell (CRC-SI, Melbourne) for their involvement in this project. We woul also lie to than the anonymous reviewers for their helpful comments provie uring the preparation of this paper.