and 30 person hours to cover the same area using GPR. Resultantly, the scale of blanket peat coverage across Dartmoor, UK (406 km 2 ) would preclude the use of both of these methods.
As an alternative to measuring peatdepth in situ, some studies have modelled peatdepthusing remotely sensed data. Holden and Connolly (2011) modelled peatdepth for the Wicklow mountains, Ireland using an exponential relationship with slope constrained by elevation (na- tional DTM) and disturbance mapped using satellite imagery. Parry et al. (2012) also used exponential relationships with airborne Inter- ferometric Synthetic Aperture Radar derived slope and/or elevation, this time constrained by previously mapped soil/vegetation units to model peatdepth for Dartmoor. Rudiyanto et al. (2016) used Shuttle Radar Topography Mission derived digital elevation model to derive topography, slope, aspect, wetness index and distance to river metrics. They then applied a quantile regression function and cubist regression tree models to model tropical peatdepth in Indonesia. In a more recent study of Indonesian tropical peat depths, Rudiyanto et al. (2018) ap- plied machine learning to 37 potential covariates derived from satellite- based remote sensing data. They found elevation, radar images (a proxy for wetness), valley bottom ﬂatness (indicative of areas of deposition) and distance to the nearest river to be the main controls on peat thickness. These models varied in resolution (30 m to 1 km) and coef- ﬁcient of determination from 0.52 ( Parry et al., 2012 ) to 0.97 ( Rudiyanto et al., 2018 ) all showing the potential of modelling peat depths across larger extents. However, these models do not account for the underlying, and often complex topography commonly smothered by blanket bogs. In addition, they assume a direct relationship between peat depths and present accumulation rates controlled by topography, elevation, slope, aspect and wetness.
infrastructure asset management and the role of noncontact remote sensing technologies for effective asset management have been discussed in depth by Hudson, Haas, and Uddin (2). Multi-purpose, self-contained vehicles have become popular among highway agencies in the last two decades, particularly as a part of a pavement management system (PMS). Self-contained condition survey vehicles are useful for collecting detailed highway inventory and condition monitoring data at high speeds using noncontact photography, video, laser, acoustic, radar, and infrared sensors (3). These data are primarily related to pavement attributes. However, even these ground based, noncontact technologies may suffer limitations resulting from time of day and traffic congestion, due to the traffic levels and proximity to urban locations, respectively. Additionally, both in-vehicle based and traditional ground survey can be quite hazardous. The new developments in airborne and spaceborne remote sensing technologies can now provide high-resolution georeferenced digital images of transportation infrastructure corridors. These options can provide complementary inventory data on the network level. The digital elevation model (DEM) generated by these remote sensing databases can be used for detailed rehabilitation strategies, design of drainage structures, and watershed modeling.
The study area has seven land cover classes, including forest, grassland, shrubland, fallow, road, wetland-clear and wetland-algae. We also included another class for land that was in shadow, regardless of its use. Although this application is focused on wetland mapping, these additional land cover categories are required for “upland” clas- sification. Note that wetlands were formally separated in two classes (clear and algae dominated) to minimize the model confusion when classify the wetland pixel with distinct spectral response. These classes are useful in this study, but they do not represent the standard classi- fication delivered by some NWIs, which also includes the wetland type. Regarding that, a field survey was conducted in August 2018 and some wetland locations were verified (Fig. 3). The marshes were identified as the common wetland systems, with open freshwater and emergent soft- stemmed plants (e.g., bulrush). Since the study area is relatively small with the dominance of marsh systems, our classification is focused on wetland location and distribution. Stratified training samples were then generated for each class (Table 1), and a total of 2260 points were la- beled based on visual interpretation on WorldView-3 image and aux- iliary data (ESRI service layers and 1-m NAIP 2017 aerial photo). The sample size was greater than 100 per class. These points were cate- gorized as fallow (∼11.8 %), forest (∼13.4 %), grassland (∼14.4 %), road (∼7.7 %), shadow (∼12.1 %), shrubland (∼11.2 %), wetland- clear (∼16.6 %), and wetland-algae (∼12.8 %) classes. These samples were randomly split into training (1470) and testing (790) samples, where the training set is used to train the models (Section 3.2.2 and 3.3) and testing samples were used for further evaluation of DNN results (Section 4.2.). In addition, a reference wetland map (polygons) was
• Wetland versus upslope location, with 1 denoting wetland, 0 otherwise. These plot-specific assignments were further checked by cross-referencing with (i) GIS data layers pertaining to forest inventory and wetland cover and type, and (ii) aerial photographs . LiDAR-derived 1-m resolution digital ele- vation models (DEMs) were used to generate the cartographic depth-to-water index (DTW) for each site (Figure 2). This index determines the difference in elevation between the ground surface and the nearest open-water features, such as flow channels and water pools. As such, it also emulates the gradation in soil drainage from very poor, poor, imperfect, moderate, well and excessively well  . Table 2 provides a summary of all transect and plot determined nu- merical and categorical variables, by map-versus plot-based categories. The map-based variables were used for province-wide moss distribution mapping, whereas the plot-determined variables were intended to reveal finer resolution and therefore not-yet projectable moss-distribution differences.
significant impacts including a large soil erosion event, significant reduction in koala (Phascolarctos cinereus) and brush-tailed rock-wallaby (Petrogale penicillata) populations and the destruction of 50 houses near the park boundary (Burned Area Assessment Team 2013; unpublished government report). Much of the area was severely burned with total crown consumption, but there were areas of much milder severity that acted as refuges for the biota. Park managers observed widespread vege- tation regrowth following the fire and wanted to know the pattern of post-fire recovery in terms of biodiversity value and fuel hazard for future fire regime management. Here we present a quantification of the spatial pattern of mid-story vegetation regrowth (defined as all vegeta- tion 0.5 – 3.0 m from the ground) following the fire and statistical modelling of potential reasons why the response varied spatially. This study was used to test two opposing hypotheses about fire severity and vegetation response: (1) high-severity fire causes high biological impact and hence slow recovery (i.e., the public ’ s percep- tion of vegetation response to severe wildfire) or (2) high-severity fire creates the conditions for strong recov- ery (i.e., fire stimulates germination and recruitment; Ooi et al. 2006, Liyanage and Ooi 2015). Gordon et al. (2017) found widespread vegetation regrowth following the fire (particularly of Acacia shrubs, which have long- lived, fire-responsive, soil seed banks), including in areas burned at high severity; Acacia cover was 4.3 times greater 2 yr after than before the fire, cover was posi- tively correlated with fire severity and the mean maxi- mum height of recruiting seedlings was 2.4 m. Therefore our main hypothesis was that recovery would be more vigorous in areas of high fire severity. Moreover, we expected patches of vigorous regrowth to be large (several hectares), reflecting the size of the high-severity patches (85% of the high-severity fire occurred within discrete patches > 10 ha in area). Such large areas of dense regrowth could pose a significant fire hazard in the park, which was something that park managers were particularly concerned about both in the short and long term. As well as the ecological aspects, this study tests the use of airborneLiDAR technology to map post-fire vegetation recovery, a technique that could be used rou- tinely if it is found to be accurate and precise. While air- borne LiDARdata has been applied to measure many aspects of vegetation structure (especially in forests; Gonzalez-Olabarria et al. 2012, Levick et al. 2012, Brosofske et al. 2014, Harpold et al. 2015), its use to map post-fire vegetation recovery across large areas, especially in Australia, is rare (Kane et al. 2013, 2014).
Abstract—This paper evaluates rasterization strategies and the benefit of hierarchical representations, in particular attribute profiles, to classify urban scenes issued from multispectral LiDAR acquisitions. In recent years it has been found that rasterized LiDAR provides a reliable source of information on its own or for fusion with multispectral/hyperspectral imagery. However previous works using attribute profiles on LiDAR rely on elevation data only. Our approach focuses on several LiDAR features rasterized with multilevel description to produce precise land cover maps over urban areas. Our experimental results obtained with LiDARdata from university of Houston indicate good classification results for alternative rasters and even more when multilevel image descriptions are used.
Detailed tree-level forest management requires individual trees to be segmented from the LiDAR point clouds. Although numerous tree segmentation methods have been developed, they have major focus on conifer forests or forests with relatively open canopy where assumptions about size and shape of tree crowns are made . Deciduous forests present considerably more complex vegetation conditions due to large variation in tree shapes and sizes, larger number of species, and denser canopy where individual trees are considerably more challenging to segment . In addition, retrieval of understory trees usingairborneLiDAR is much harder because of the reduced amount of LiDAR points penetrating below the main cohort formed by overstory trees . Typically, the detection rate of overstory (dominant and co-dominant) trees is around 90% while the detection rate of understory (intermediate and overtopped) trees is below 50%. Trivi- ally, a minimum point density is the basic requirement for a reasonable segmentation of trees, which is typically not satisfied for understory trees in a dense forest [11, 12]. Furthermore, LiDARdata covering an entire forest is much more voluminous than the memory of a worksta- tion and may also take an unreasonable time to be sequentially processed using an external memory algorithm. Because large-scale LiDARdata is typically arranged in several tiles for efficient management and delivery, distributed processing of different tiles is straightforward. However, the data representing tree crowns located across tile boundaries are split into two or more pieces that need to be processed by different computing units. Only few studies have considered distributed processing of large geospatial data addressing the boundary problem —specifically there are no studies considering forest data. This is increasingly important when obtaining tree-level information for areas other than small plots, which is often the ideal objective. Moreover, continuous advancements of sensor technology and platforms  is resulting in point clouds to be acquired with greater resolutions, increasing the need for more efficient and scalable processing schemes.
LiDAR is an active remote sensing technology . It actively transmits pulses of laser light toward an object of interest, and receives the light that is scattered and reflected by the objects . An airborneLiDAR system is typically composed of three components: a laser scanner unit, a GPS receiver and an IMU , , , . The laser scanner unit consists of a pulse generator (Nd:YAG laser) with a wavelength in the range of 0.8 µm to 1.6 µm (typically, 1.064 µm or 1.500 µm) and a receiver to get the pulses scattered and reflected by the targets , . The laser pulses are emitted at a rate of up to 250 kHz to the Earth surface . The distance (range) between the LiDAR sensor and the target is calculated by multiplying the speed of light by half of the time it takes for the light to travel from the sensor to the target and back . The GPS receiver is used to record the aircraft trajectory and the IMU measures the attitude of the aircraft (roll, pitch, and yaw or heading) . The calculated range between the scanner and the target and the position and orientation information obtained from the GPS and IMU units are used to determine the target location in three-dimensional space . The three-dimensional LiDAR points are initially represented by latitude, longitude, and ellipsoidal height based on the WGS84 reference ellipsoid. They can then be transformed to a national or regional coordinate system. During this process, elevations are converted from ellipsoidal heights to ortho- metric heights based on a national or regional height datum , .
adjustable through a definition in C. The two dimensional array that represents the entire map is declared as map[MAP_SIZE][MAP_SIZE], where MAP_SIZE is defined as ten times double the max range the LiDAR can scan, divided by the scale of each index. For example 10 * (24000 / 20), where 24000 is the double the max range, 20 is our scale. A multiple of ten is used to give us a large enough map of 57,600 meters squared, while only using about 10% of memory; theoretically we can use a multiple of 25 to use up almost all of the RAM memory, and generate a map that is 360,000 meters squared in area. Double the range of the LiDAR is used to account for the fact that we initialize the location of the robot to be at the center of the map. Indexes into the two dimensional array are done in an X,Y fashion(map[y][x]), where index (0,0) is the farthest north-west index, (max,max) is the farthest south-east index, and (max/2,max/2) is the center of the map. Refer to figure 4 in the appendix E for a visual depiction of the map.
The LIDAR technology has few constraints typical to conventional topographic survey methods. It can survey day and night, at altitudes between 300 and 900 m (1,000 to 3,000 ft) above ground, over any terrain, and through most vegetation and canopy. Most of the highway application surveys are conducted at a height of 500 m (1,500 ft) above ground level, as illustrated in Figure 1. The flexibility of day and night missions is subjected to usual constraints of flying aircraft at relatively low altitudes due to applicable aviation rules. An airborne platform provides non-intrusive operation and no interference with highway traffic. A twin-turbo-prop Cessna aircraft is used for ALTM missions. Before a flight, a ground-based GPS is set up on a known point in the survey area. Flight planning determines optimal LIDAR settings and aircraft parameters. A typical survey can collect data at a rate of up to 81 sq. km (20,000 acres) per day.
Traditional methods for rainforest boundary determination in an area similar to our study area were based on the interpretation of aerial photographs or field work. These methods are labour intensive and time consuming. In addition, aerial photographs (and even the satellite images) are only able to provide information on horizontal forest extent. The LiDARdata have advantages over these traditional remote sensing data in that vertical forest structure (or the interior of the canopy and understorey vegetation) can be delineated in much detail . However, there is still considerable scope for developing advanced methodology to take maximum advantage of the information extracted from the LiDARdata for effective determination of rainforest boundaries. Therefore, the overall objective of this study is to develop a fuzzy logic model usingairborneLiDARdata for determining the boundary between cool temperate rainforest and adjacent forests in the study area. Fuzzy analysis of LiDAR-derived variables will be performed to determine fuzzy membership values describing degrees to which each area belongs to a certain forest type in the study area, and the confusion index values will also be calculated to illustrate the transition zones (or ecotone areas) and determine the boundaries between the rainforest and non-rainforest.
It can be seen that the increases in the accuracies achieved in some of the datasets by the addition of cer- tain variables are not large. RF produces a measure of the variable importance by analyzing the deterioration of the predictive ability of the model when each predictor vari- able is replaced in turn by random noise (Vincenzi et al. 2011). In general, the texture measures and radar data have low importance scores. The class-specific contribu- tions of different variables to the models are shown in Figure 3. Due to their negligible influence, the texture measures (optical and radar) have been omitted. In all three study areas, all models strongly relied on distinct spectral bands and band ratios. The influence of the ancil- lary data is variable between classes and study sites. The RF models were applied across the entire study areas to obtain vegetation cover for the whole regions (see Fig. 4), while the upland subsets in these study areas are shown in Figure 5. These maps were created using the (vii) data- set, without the inclusion of the soils and elevation ancil- lary data. A 3 9 3 pixel majority filter was applied to the thematic outputs to improve the homogeneity of the final product. As can be seen from Figure 4, the dominant veg- etation cover in all areas is grasslands, and this is relatable to most areas in Ireland. There is very little forest cover on the Dingle Peninsula, while both the Galtee and Comeragh study areas have considerably larger forest areas, especially along the lower slopes of the upland areas. These areas usually represent lands that are mar- ginal for agriculture and since the 1950s, large extents have been afforested, supported through various govern- ment and EU incentive programmes. Concentrating on the upland subsets in Figure 5, the true value of upland areas in terms of habitat diversity is apparent. Mount Brandon (Fig. 5A) has extensive areas of wet heath, semi- improved (dry-humid acid) grasslands, blanket bog and dry siliceous heath. Large areas of montane heath are observed, especially along the western edge of the area making it quite distinctive when compared to the Galtee and Comeragh Mountains. From Figure 5(B), the domi- nant classes for the Galtee Mountains are dry-humid acid grassland along the north-west of the area, dry siliceous heath and blanket bog. Wet heath occurs less frequently, compared to the Mount Brandon area, though there are increased areas of wet grassland. Similar to the Galtees, the dominant classes in the Comeragh Mountains area (Fig. 5C) are blanket bog, dry siliceous heath and dry- humid acid grassland. Small areas of wet heath are scat- tered throughout the area and areas of dense bracken are prevalent along the eastern edges of the upland area.
An LVIS waveform essentially consists of a signal with amplitudes proportional to energy re ﬂected from intercepted surfaces within canopy and ground. LVIS footprints are geo-located to the global reference ellipsoid WGS 84, using a combination of GPS and Inertial Navigation System (INS) information ( Blair et al., 1999; Hofton & Blair, 2002 ). Ground elevation is determined by identifying the center of the lowest mode in the waveform greater than mean signal noise ( Fig. 2 ). Canopy elevation is the height at which the signal increases beyond a certain threshold (usually 3 σ) at the top of the waveform ( Hofton & Blair, 2002 ). The difference between canopy elevation and ground elevation gives the canopy height metric or height of 100% laser energy return (RH100). The 1999 LVIS data were reprocessed using new algorithms for ground detection and an improved horizontal geolocation algorithm ( Blair et al., 2006 ). For each LVIS waveform, quartile heights of laser energy return i.e. height of 25% (RH25), 50% (RH50) and 75% (RH75) energy return were calculated in addition to RH100 ( Fig. 2 ). Canopy cover was calculated from the ground energy return of each waveform normalized by the canopy and background re ﬂectivity ratio ( Ni Meister et al., 2001 ). We used a ratio of 1.6, derived from a previous study ( Hyde et al., 2005 ). LVIS metrics were calculated for lidar shots within 1 ha plots and summarized to obtain minimum, maximum, mean and standard deviation for all metrics.
development are intimately related. Some models simply propose that pipes result in landform change (e.g., topographic depressions or gullying) [e.g., Pearsall, 1950; Jones, 1990; 1994a; 1994b; 1997b] or that existing topography promotes enhanced flow in concentrated areas of hillslopes which promotes piping [Anderson and Burt, 1982]. However, I propose that the nature of the underlying topography (and its associated drainage conditions) promotes differential build up of the peat deposits. This occurs because of the development of micro pools and larger bog pool systems on hilltops and toes which are colonized by a mosaic of plants with specialist positions within the microtopography. The remains of these plants are then incorporated into the peat as it thickens, resulting in a peat of variable properties throughout its profile. In addition, bog pool development tends to be a cyclic process; pools disappear from one spot in a peatland while new ones form elsewhere as
When taking a spectral measurement, we record five separate, sequential spectral data files in rapid succession, typically taking about five seconds duration. Post processing comparison of those five files allows a first-order check on data consistency. We compare averages of the VNIR, SWIR1 and SWIR2 spectral regions between the files. If they vary by a predetermined tolerance (typically 1% for the VNIR), then the data is disregarded as having excessive variability. Typically, this occurs for two reasons: the illumination was changing on a time scale of seconds or (2) a malfunction in the control software handshaking between the DOS-based ASD control and Windows-based video control programs that are running simultaneously on the system.
It is estimated that the total area of tropical peatland is in the range of 30-45 million hectares (ha), approximately 10-12% of the global peatland resource, and about 16.8-27.0 million ha are found in Indonesia (Immirzi et al., 1992; Rieley et al., 1996; Page & Banks, 2007; Page et al., 2010). Tropical peatlands act as sinks and store huge amounts of carbon as peat consist of dead, incompletely decomposed plant material that has accumulated over thousands of years in waterlogged environments that lack oxygen, which makes them one of the largest near-surface pools of terrestrial organic carbon (Sorensen, 1993; Page & Rieley, 1998; IPCC, 2007). Typically tropical peatlands are located at low altitudes in alluvial floodplains where peat swamp forests form smooth convex shaped peat domes up to 20m thick (Anderson, 1983; Rieley et al., 1996; Rieley & Page, 2005). Indonesian peatlands feature a characteristic micro-topography consisting of hummocks and hollows (Rieley & Page, 2005). Hummocks are up to 0.5m in height and are usually formed around tree bases and comprise of large proportion of both living and dead tree roots (Rieley & Page, 2005). These hummocks are interspersed with hollows of similar depth which form an interconnected network that carries water from the interior peat dome to its periphery (Rieley & Page, 2005). The ability of tropical peatlands to sequester and store these huge amounts of carbon is threatened by persistent anthropogenic impacts (Rieley & Page, 2005; Jaenicke et al., 2008; Hooijer et al., 2010). Especially drainage and forest clearance disturb their hydrological stability (Page & Rieley, 1998) and make these otherwise waterlogged ecosystems susceptible to fire (Langner et al., 2007). After peatland drainage there are four largely irreversible processes that lead to a drop in peat surface level, called subsidence (Rieley & Page, 2005): (a) Consolidation is the compression of saturated peat due to increased ‘overburden’ (no peat matter is lost; increase of bulk density); (b) Shrinkage is the gradual volume reduction of peat in the unsaturated zone due to a loss of water from pores (no peat matter is lost; increase of bulk density); (c) Oxidation is the gradual volume reduction of peat in the unsaturated zone due to decomposition of organic matter; and (d) Fire which results in the complete loss of peat organic matter from the burned zone (carbon dioxide (CO 2 ) to a lesser degree
The hydraulic conductivity values from 10 cm to 80 cm depth in blanket peat are generally an order of magnitude lower than those measured in the Somerset Levels (Baird and Gaffney, 1994). Analysis of variance (ANOVA) suggests that, for depths equal to or greater than 10 cm, depth is not a significant control on k* (Table 2). Even at 10 and 20 cm depth hydraulic conductivity can be as low as 3.43 x 10 -7 cm s -1 and 1.78 x 10 -7 cm s -1 respectively. There is often an assumption that k decreases gradually with depth (or decomposition) in peat (e.g. Ingram, 1983). Rycroft et al. (1975) extensively reviewed reported hydraulic conductivity values from peats (using rigid soil theory and hence comparison must be treated with caution) and their table of values suggests k for blanket peats ranging from 1.1 x 10 -5 cm s -1 at 30 cm depth (Galvin and Hanrahan, 1967) to 6 x 10 -8 cm s -1 at 1 m (Ingram, 1967). Values in other peats tend to be slightly higher (e.g. Dai and Sparling, 1973; Neuman and Dasberg, 1977). In poorly decomposed fenland peats values as high as 5 x 10 -3 cm s -1 have been reported at 1 m depth (Rycroft et al., 1975). In the Moor House blanket peats mean k* between 10 and 80 cm depth is 2.9 x 10 -6 cm s -1 (although this would have been calculated at 1.28 x 10 -5 cm s -1 if rigid soil theory was used at t 90 ). There is no evidence therefore to suggest that
1.4 – Carbon Budgets and Modelling of Peatlands
Estimating the net carbon budget, i.e. the total net ecosystem uptake or loss of carbon for a given area has become an important area of environmental science, particularly with respect to forecasting future changes in carbon stocks. A common method to estimate carbon balance on a national/international scale is to use temporal changes in national inventories of soil carbon (e.g. Bellamy et al., 2005), coupled with estimates of the effects of land-use/climatic/vegetational changes on the system (e.g. Janssens et al., 2005; Joosten, 2009; Armentano and Menges, 1986). This approach is limited primarily by the quality of the data (e.g. resolution, comprehensiveness etc.), flux estimates (e.g. comprehensiveness, suitability) and conceptual model(s) available. For example, Armentano and Menges (1986) estimated changes in temperate zone peatland carbon stocks, using a deterministic empirical model. However, the model had a number of assumptions that were verifiably false. They assumed, for instance, that only drained peatlands would export DOC, whereas DOC is lost from pristine peatlands (e.g. Worrall and Burt., 2005) as well as drained ones (e.g. Wilson et al., 2011), and DOC typically represents 10% of carbon released from peat (e.g. Holden, 2005b). The results and estimates of the Armentano and Menges (1986) study were used, in part, for Janssens et al. (2005) the UK carbon balance study. In addition to the errors introduced by Armentano and Menges (1986), Janssens et al. (2005) is limited by the fact that it considers only peatland drainage for arable or forestry, rather than peatland drainage for heathland, which is the most common reason for peatland drainage in the UK (Holden et al., 2007). As such, Janssens et al. (2005) could be underestimating carbon losses from peatland drainage, while their use of estimates from Armentano and Menges, (1986) means they may also be underestimating losses of carbon as DOC from pristine peatlands.
LiDAR generated terrain models are a suitable tool for the determination of geomorphologic changes. Erosion and deposition zones can easily be identified, presenting an overview of the geomorphological situation of the surveyed region. This applies particularly for a balance of eroded or deposited material – independently of the analysed process. Inaccuracies have been found in identifying pseudocontrol points and ranges of steep slope situations. Here larger un- certainties may be expected. Pseudocontrol points surrogate the lack of ground control points, making a correction of dig- ital elevation models possible. Their identification can be difficult whereas the filtering of LiDARdata can have a dis- tinct impact on various morphologic features (houses, roads, bridges, etc.). For the analysis of landslides as well as for debris flows a potential elevation-error as a function of the slope inclination has to be considered. For analysing chan- nel based processes with changing erosion- and deposition behaviour (as for fluvial sediment transport) it is therefore recommended to discretise the active stream into homoge- neous slope zones. LiDAR generated surface models could also be used to investigate erosion processes of streamside trees linked to an estimation of the magnitude of potential woody debris.