Mapping forest cover and forest cover change with airborne S-band radar

: Assessments of forest cover, forest carbon stocks and carbon emissions from deforestation and degradation are increasingly important components of sustainable resource management, for combating biodiversity loss and in climate mitigation policies. Satellite remote sensing provides the only means for mapping global forest cover regularly. However, forest classiﬁcation with optical data is limited by its insensitivity to three-dimensional canopy structure and cloud cover obscuring many forest regions. Synthetic Aperture Radar (SAR) sensors are increasingly being used to mitigate these problems, mainly in the L-, C- and X-band domains of the electromagnetic spectrum. S-band has not been systematically studied for this purpose. In anticipation of the British built NovaSAR-S satellite mission, this study evaluates the beneﬁts of polarimetric S-band SAR for forest characterisation. The Michigan Microwave Canopy Scattering (MIMICS-I) radiative transfer model is utilised to understand the scattering mechanisms in forest canopies at S-band. The MIMICS-I model reveals strong S-band backscatter sensitivity to the forest canopy in comparison to soil characteristics across all polarisations and incidence angles. Airborne S-band SAR imagery over the temperate mixed forest of Savernake Forest in southern England is analysed for its information content. Based on the modelling results, S-band HH- and VV-polarisation radar backscatter and the Radar Forest Degradation Index (RFDI) are used in a forest/non-forest Maximum Likelihood classiﬁcation at a spatial resolution of 6 m (70% overall accuracy, κ = 0.41) and 20 m (63% overall accuracy, κ = 0.27). The conclusion is that S-band SAR such as from NovaSAR-S is likely to be suitable for monitoring forest cover and its changes.


Remote Sensing and Forest Mapping
Forests play a pivotal role in biogeophysical feedbacks of the terrestrial biosphere with the climate system [1] and constitute one of the main components of the global carbon cycle [2] in the form of woody aboveground biomass (AGB) [3]. Availability of information on forest cover and accurate mapping algorithm are essential for monitoring and managing forests [4]. Accuracies of forest cover information vary depending on the type of tools and methods used. For example, the United Nation's Food and Agriculture Organization (FAO), has provided estimates of global forest cover at 5-10 year intervals since 1946 based on national statistics [5], largely relying on forest inventory data, models and expert opinion. However, there is a lack of consistency in terms of defining forest and assessment methods of different countries [6]. Moreover, there are data gaps in national forest inventories in many parts of the world, particularly areas with inaccessible problems. To overcome these uncertainties on forest cover extent, surveys based on remote sensing are now being used operationally [7].
For a consistent and systematic study on global monitoring of forested areas, use of Earth observation data from satellite offers robust and continuous data with large area coverage and repetitive observations [8] and can reduce the cost and time of field inventories. At the small scale, remote sensing also provides forest maps on inventory estimates with precision for limited field data and can be used for purposes related to timber production, procurement and ecological studies [9]. Therefore, remotely sensed data have been incorporated into operational forest inventories. Over the past decades, different forest cover maps at continental to global scales have been generated utilising optical sensors with fine, medium and coarse spatial resolutions [4,[10][11][12]. The majority of the optical-derived forest/non-forest (F/NF) maps have been generated from optical data acquired in a single year or period due to persistent cloud cover using time-series composite data. Additionally, they have been used for mapping forest cover and AGB of forest where relatively simple stand structure and gradients of cover exists such as wooded savannas [13].
For the first time, forest cover maps with the annual forest gain and loss from 2000 to 2012 have been produced using multi-temporal fine-resolution Landsat data at global scale [4]. The existing optical-derived global maps showing areas of F/NF reveal differences in the forested area estimations ranging from 3.5 billion ha to 4 billion ha primarily due to the varying definition of forest and methodologies. In general, forest is defined as all contiguous area >0.5 ha with at least 10% canopy cover and 2-5 m height [5]. Moreover, the spectral reflectances (both absorbed and scattered) of forest canopy components (leaves, needles and branches) at optical wavelengths are sensitive to weather conditions and closely related to two-dimensional structure (i.e., amount of foliage cover, leaf area index) and their biochemical properties such as chlorophyll, nitrogen, lignin and cellulose content and the difficulty in separating mixed vegetation types has resulted in varying accuracies between 66.9% (IGBP DISCover 1.1 km) and 78.3% (MODIS 1 km) at global scale [14]. Information related to radiation scattering from leaves, branches, trunks (i.e., forest structure particularly height) and the ground is highly desirable for differentiating woody from herbaceous or other mixed vegetation where the use of optical data is often problematic. Specific information related to stands of different species, age or density [15] including stand information with tree trunk diameter and stem volume [16] will aid in classification of forest cover types. Forest structural properties related to crown architecture at tree level and spatial heterogeneity distribution at stand level will also be useful in forest cover mapping [15].

Forest/Non-Forest Mapping Using SAR
Approximately 47% of the Earth's land surface was assumed to be covered by closed forest about 8000 years ago based on numerous global and regional biogeographic information [17]. Since the last decades, nearly 30% of the world's land area is forested due to wildfires, drought, logging and large-scale deforestation particularly in the tropics [18]. Within this area, different forest types occur based on climatic (global), soil, topography and disturbance factors at landscape level [19]. They also constitute about 80% of terrestrial AGB as a function of forest growth (e.g., regeneration) and mortality (e.g., logging, deforestation) [18]. The extent of progressive losses in forest area and forest biomass (productivity) caused by human activities has been substantial, resulting in significant greenhouse gas emissions and regional to global climate changes and carbon budgets [20]. Therefore, accurate measurement of the forested areas and spatio-temporal variations in the biomass of these forests at local, national to global scales is needed.
For an operational forest cover monitoring system at high spatial and temporal resolutions, Synthetic Aperture Radar (SAR) systems have emerged as robust tools which can penetrate clouds and record signals which originate from scattering within the vertical canopy layer [21,22]. For example, Shimada, Itoh, Motohka, Watanabe, Shiraishi, Thapa and Lucas [22] have generated global F/NF cover maps (2007-2010) at 25 m resolution from the Advanced Land Observing Satellite (ALOS) Phased Arrayed L-band Synthetic Aperture Radar (PALSAR) HV gamma-nought (γ 0 ) reporting around 85% overall accuracy. For the past three decades, significant efforts in microwave remote sensing have been devoted to interpreting the radar backscatter [23] for improved forest classification accuracy [24] and biophysical retrievals [25,26]. For information related to forest attributes at the stand level such as cleared-felled or canopy degradation, S-band radar data at 6 m resolution from NovaSAR-S could potentially provide a more useful data product than is currently available from Sentinel-1, ALOS-2 and other SAR sensors that provide global coverage.
Understanding the relationships between SAR backscatter at different wavelengths and polarisation combinations and forest canopy parameters is critically important for developing forest monitoring methods. However, the sensitivity of the radar backscatter to different canopy elements (leaves, branches and trunks) cannot easily be quantified based on image data alone (without complementary field data) due to the inter-relationship between forest parameters and radar parameters. Simulations from microwave canopy radiative transfer models provide insight into the complex interactions of the radar wave with the different components of the canopy as a function of wavelength, incidence angle and polarisations [26][27][28][29]; for instance, increasing level of penetration with increasing wavelength. Such models have shown that the low frequencies (P-and L-band), in particular co-polarized (horizontal transmit and horizontal receive-HH, vertical transmit and vertical receive-VV) channels produce stronger scattering from larger scattering elements (branches, trunks) with high frequencies (X-and C-band) from leaves and needles. This can be understood by radar being most sensitive to features in a scene with dimensions characteristic of the scale of the probing wavelength, combined with increasing canopy penetration wave with increasing wavelength. In contrast to X-, C-, L-and P-band, the nature of the S-band microwave scattering mechanism in forest canopies is poorly studied to date due to the historic lack of availability of S-band SAR data. The S-band (3.1-3.3 GHz) lies between the longer L-band (1-2 GHz) and the shorter C-band (5-6 GHz).

Potential of S-Band Data for Forest Mapping
The first S-band SAR observations from space were provided by the Russian Almaz-1 satellite, which carried a single-polarized HH S-band SAR sensor operating between 1991 and 1992. Since the Almaz-1 mission, interest in S-band SAR has led to the launch of the Environment and Disaster Monitoring Satellite Huan Jing -1 Constellation (HJ-1C) from China [30] and new S-band SAR missions are planned, e.g., the UK NovaSAR S-band [31] and the joint NASA-ISRO Synthetic Aperture Radar (NISAR) missions [32]. NovaSAR is planned to provide S-band data with StripMap and ScanSAR resolution modes at 6 m and 20 m respectively.
Few published studies have investigated S-band radar responses to agricultural crops and forest canopies to date. The sensitivity of S-band backscatter to maize crops was dominantly volume scattering due to the random distribution over the whole plants [33]. The Michigan Microwave Canopy Scattering (MIMICS) model suggests that S-band backscatter is sensitive to the temporal dynamics of the structure of wheat crops [34]. In the young fir, S-band backscatter arises from the needles and branches resulting in lower volume scattering due to the conical structure of conifer species [35]. The capability of S-band backscatter to map agricultural crop canopies [36] and forested areas [37] using an integration of polarimetric scattering entropy H/average alpha angle α and Pauli's decomposition techniques and salt marsh habitat mapping using a Random Forest algorithm [38] have also been explored using airborne very-high resolution S-band data. Studies by Yatabe and Leckie [39] and Fransson, et al. [40] have demonstrated the mapping of clear-felled and forested areas using S-band Almaz-1 in boreal forest of Canada and Sweden.
The fine 6 m resolution data from NovaSAR could provide a useful and better than other available datasets from Sentinel-1, ALOS-2 and other SAR sensors that provide global coverage. This means that the radar pulse at this frequency can provide the information in relation to the small-scale changes in the forest ecosystem particularly degraded and disturbed areas due to its sensitivities to the forest canopy and its structural components. Modelling studies have highlighted the sensitivity of longer wavelength at L-band to the varying sizes of branch canopy and canopy density [29] and similar sensitivity to varying canopy density, branching level and branch size have also been found to be sensitive to S-band radar pulse.
The main objective of this study is to classify S-band radar backscatter data to produce F/NF cover and change maps using a Maximum Likelihood Classification (MLC) algorithm at different spatial resolutions that match the anticipated acquisition modes of NovaSAR-S. Scattering mechanisms in the canopy and from the ground have been investigated from radiative transfer modelling experiment with MIMICS-I [29]. This paper is organised as follows: Section 2 gives the details of test site, the dataset involved in the study and the used methodology. Section 3 mainly deals with the results and discussion and the conclusions are drawn in Section 4.

Site Description
Savernake forest (51˝23'13"N, 1˝43'19"W) is located near Marlborough in England, UK ( Figure 1). The study site of Savernake forest was chosen based on its diverse land cover types. A field campaign was carried out to collect ground data. The forest is composed of mixed stands of temperate deciduous and coniferous species with varying tree densities, canopy height, growing stock, and age class. Main deciduous species include ancient beech (Fagus sylvatica), birch (Betula pendula) and oak (Quercus spp.) with mixed species of yew, ash, common lime, crab apple, elm, field maple, hazel, horse chestnut etc. Dominant coniferous species consist of Scots pine (Pinus sylvestris), Corsican pine (Pinus nigra), Norway spruce (Picea abies) and Western hemlock (Tsuga heterophylla). The non-forest comprises three main land cover classes, as grassland, clear-felled areas and bare-ground. The existence of mix forest/non-forest classes in the study site could potentially lead to lower accuracy level and confusion in any classification algorithm. Savernake forest is maintained by the Forestry Commission (FC) and the forest is divided for management purposes into sub-compartments of 4 ha average size. A series of management operations have been undertaken in Savernake forest between December 2012 and July 2013. For example, two sub-compartments of sizes 3.45 ha and 5.2 ha have been completely clear-felled followed by plantation with Douglas fir and oak seedlings. Otherwise, most of the sub-compartments were subject to thinning operations to enhance timber production. Topography is relatively flat with an average elevation of 107 m and 1% slope with a south-eastern aspect [41]. The forest receives an average of 750 mm of annual precipitation and has an average annual temperature of 11.3˝C. Geologically, the parent soil material is characterized by Jurassic Clay (Oxford) with Eutric vertisol, and has a soil pH of 3.9-6.2.

SAR Data and Processing
A number of SAR airborne campaigns have been undertaken in Europe to investigate the capabilities of different SAR sensors to forest characteristics. For example, the MAESTRO-I campaign during August 1989 utilising the NASA-JPL AIRSAR sensor in four European test sites [28,42] focusing on SAR bands at P-, L-, C-and X-frequencies. Now, for the first time, extensive S-band SAR campaign have been conducted over the temperate region of UK known as the 'AirSAR Campaign' in 2010 and 2014. The Airborne SAR Demonstrator Facility 'AirSAR' is a collaborative project operated by Airbus Defence and Space (UK) with the Natural Environment Research Council (NERC) and the Satellite Applications Catapult [43].
The frequency band used in the study is S-band (3.1-3.3 GHz) with a swath width of 1.92 km and incidence angles from 16° to 42.5°. The Single-Look Complex (SLC) imagery is provided with 0.75 m pixel in both ground and cross-range spacing with 4.48 looks in azimuth and 1 look in range direction [43]. Their characteristics are summarized in Table 1.

SAR Data and Processing
A number of SAR airborne campaigns have been undertaken in Europe to investigate the capabilities of different SAR sensors to forest characteristics. For example, the MAESTRO-I campaign during August 1989 utilising the NASA-JPL AIRSAR sensor in four European test sites [28,42] focusing on SAR bands at P-, L-, C-and X-frequencies. Now, for the first time, extensive S-band SAR campaign have been conducted over the temperate region of UK known as the 'AirSAR Campaign' in 2010 and 2014. The Airborne SAR Demonstrator Facility 'AirSAR' is a collaborative project operated by Airbus Defence and Space (UK) with the Natural Environment Research Council (NERC) and the Satellite Applications Catapult [43].
The frequency band used in the study is S-band (3.1-3.3 GHz) with a swath width of 1.92 km and incidence angles from 16˝to 42.5˝. The Single-Look Complex (SLC) imagery is provided with 0.75 m pixel in both ground and cross-range spacing with 4.48 looks in azimuth and 1 look in range direction [43]. Their characteristics are summarized in Table 1. An antenna pattern correction with polynomials of different order was applied, and the 5th order polynomial provided the best fit. To reduce the speckle, conversion of the SLC imagery into multi-look complex images using a 5ˆ5 kernel corresponding to 3.75 m pixel size resolution was performed. Further image processing was carried out with seven different adaptive filters using 3ˆ3, 5ˆ5 and 7ˆ7 kernels where Enhanced Frost filter performed best for reducing speckle with a 5ˆ5 kernel. These include Kuan, Lee, Enhanced Lee, Frost, Enhanced Frost, Gamma and Local Sigma filters. The analysis was based on the mean and standard deviation (SD) of the DN values on the assumption that the filter that retains the mean of the DN values of the raw data while reducing the SD to the minimum, is the ideal filter for reducing the speckle effect. For all the polarisations and in different kernels (3,5,7), the Enhanced Frost, Enhanced Lee, Local Sigma and Kuan filters produce similar mean values to the raw data. However, the SD increases from Enhanced Frost up to Kuan filters. Frost, Gamma and Lee filters produces maximum mean values and highest SD relative to the unfiltered raw data. Figure 2 represents the performance of different filters for reducing the SAR speckle with HV polarisation (2014 data).  An antenna pattern correction with polynomials of different order was applied, and the 5th order polynomial provided the best fit. To reduce the speckle, conversion of the SLC imagery into multi-look complex images using a 5 × 5 kernel corresponding to 3.75 m pixel size resolution was performed. Further image processing was carried out with seven different adaptive filters using 3 × 3, 5 × 5 and 7 × 7 kernels where Enhanced Frost filter performed best for reducing speckle with a 5 × 5 kernel. These include Kuan, Lee, Enhanced Lee, Frost, Enhanced Frost, Gamma and Local Sigma filters. The analysis was based on the mean and standard deviation (SD) of the DN values on the assumption that the filter that retains the mean of the DN values of the raw data while reducing the SD to the minimum, is the ideal filter for reducing the speckle effect. For all the polarisations and in different kernels (3,5,7), the Enhanced Frost, Enhanced Lee, Local Sigma and Kuan filters produce similar mean values to the raw data. However, the SD increases from Enhanced Frost up to Kuan filters. Frost, Gamma and Lee filters produces maximum mean values and highest SD relative to the unfiltered raw data. Figure 2 represents the performance of different filters for reducing the SAR speckle with HV polarisation (2014 data). As the study area has only a minimal topographic slope, geo-coding of the SAR imagery was done using an Ordnance Survey map (Projection: United Kingdom, Datum: Ordnance Survey of Great Britain-OSGB 1936) as reference with 30 widely distributed ground control points (GCPs), a second-order polynomial and nearest neighbour re-sampling to 6 m and 20 m pixel resolution achieving a root mean square error (RMSE) of half a pixel. For the absolute radiometric calibration of the airborne demonstrator data, corner reflectors were placed in the near, centre and far range of the swath with corresponding incidence angles of 22.11°, 30.38° and 39.96°. The radar backscatter coefficient was calculated using Equation (1) according to an Airbus Defence and Space technical report [44]:  As the study area has only a minimal topographic slope, geo-coding of the SAR imagery was done using an Ordnance Survey map (Projection: United Kingdom, Datum: Ordnance Survey of Great Britain-OSGB 1936) as reference with 30 widely distributed ground control points (GCPs), a second-order polynomial and nearest neighbour re-sampling to 6 m and 20 m pixel resolution achieving a root mean square error (RMSE) of half a pixel. For the absolute radiometric calibration of the airborne demonstrator data, corner reflectors were placed in the near, centre and far range of the swath with corresponding incidence angles of 22.11˝, 30.38˝and 39.96˝. The radar backscatter coefficient was calculated using Equation (1) according to an Airbus Defence and Space technical report [44]: rσ 0 " 10 log 10 pDN 2 q´Kcals To implement forest cover change detection, a multi-temporal image calibration (or normalization) of the 2010 and 2014 data was performed to ensure absolute radiometric calibration [45]. This was done by re-calibrating 2010 data based on 2014 data using the basic statistics (minimum, maximum, mean To implement forest cover change detection, a multi-temporal image calibration (or normalization) of the 2010 and 2014 data was performed to ensure absolute radiometric calibration [45]. This was done by re-calibrating 2010 data based on 2014 data using the basic statistics

Ancillary Data
The Forestry Commission Geographical Information System (FC GIS) database is a spatially comprehensive vector database of forest type, species, planting year, planting density and yield class for each forest stand. This database contains 215 sub-compartments for Savernake forest. The digital sub-compartment data were projected to the United Kingdom projection of the SAR image data with OSGB 1936. Information on forest type and tree species is used to train the MLC algorithm. Additionally, the ancient tree database contains approximately 2640 point locations within the image area. Each point contains detailed attribute information including diameter at breast height (DBH), canopy height (H) and tree species having a maximum tree DBH of 350 cm and canopy height of 35 m. In August 2012 and March 2015, a total of 70 sample plots were surveyed in the field, with two circular sample areas of 20 m diameter each of 35 sub-compartments with different tree species of both deciduous and coniferous forest types.

S-Band Radar Scattering Characteristics in Forest/Non-Forest Areas
This study uses the radiative transfer model MIMICS-I [29] as a tool for investigating S-band backscatter predictions and dominant scattering from forest canopies. The input parameters for MIMICS-I can be divided into two major classes related to: (1) microwave characteristics (polarisation and incidence angle); and (2) physical descriptions of the canopy (layers of leaves, needles, branches and trunk) as well as soil properties (soil moisture content and surface roughness). MIMICS-I model

Ancillary Data
The Forestry Commission Geographical Information System (FC GIS) database is a spatially comprehensive vector database of forest type, species, planting year, planting density and yield class for each forest stand. This database contains 215 sub-compartments for Savernake forest. The digital sub-compartment data were projected to the United Kingdom projection of the SAR image data with OSGB 1936. Information on forest type and tree species is used to train the MLC algorithm. Additionally, the ancient tree database contains approximately 2640 point locations within the image area. Each point contains detailed attribute information including diameter at breast height (DBH), canopy height (H) and tree species having a maximum tree DBH of 350 cm and canopy height of 35 m. In August 2012 and March 2015, a total of 70 sample plots were surveyed in the field, with two circular sample areas of 20 m diameter each of 35 sub-compartments with different tree species of both deciduous and coniferous forest types.

S-Band Radar Scattering Characteristics in Forest/Non-Forest Areas
This study uses the radiative transfer model MIMICS-I [29] as a tool for investigating S-band backscatter predictions and dominant scattering from forest canopies. The input parameters for MIMICS-I can be divided into two major classes related to: (1) microwave characteristics (polarisation and incidence angle); and (2) physical descriptions of the canopy (layers of leaves, needles, branches and trunk) as well as soil properties (soil moisture content and surface roughness). MIMICS-I model describes the vegetation layer as canopy and trunk layers where different layers of branches (viz., primary to sixth branch level) with different sizes are considered.
Forest biophysical parameters were assumed to have remained unchanged between the 2010 and 2014. Therefore, tree density, stem diameter at breast height, canopy height and branch size were assumed constant. A small number of model parameters need to be adjusted for the different imaging conditions of the two sets of radar observations in 2010 and 2014, namely soil moisture content and structure of the stands. The soil moisture information was derived from Wytham Woods test site within the AirSAR campaign, being collected by the Centre for Ecology and Hydrology in 23 June 2014. Different percentages of soil components for sand, silt and clay were derived from 1 km soil data set of Batjes [46]. Ground measurements such as density of leaves, needles and branches including their spatial orientation are extremely difficult to obtain and were not available for this site. Therefore, an informed estimate based on the literature is assumed for such structural canopy parameters. All the estimated values are derived from the MIMICS-I model technical report [47]. A list of input parameters used in the MIMICS simulations is given in Table 2. The sensitivity of S-band radar backscatter to forest canopy properties and soil characteristics was investigated as a function of polarisation and incidence angle. For this study in context to forest/non-forest assessment, only the total backscatter return from forest canopy (up to secondary branch levels) and soil properties were assessed and reported. Details related to the canopy transmissivity and contribution of the individual interactions to the total canopy backscatter at S-band will be reported in a companion paper by Ningthoujam et al., (accepted).

Forest Cover Classification
In the first classification approach, only forest and non-forest are defined as the main land cover classes (Figure 4). The ratio between HH-and HV-backscatter was calculated and utilised for the classification. It is also known as the 'Radar Forest Degradation Index' (RFDI) because HH backscatter is sensitive to both volume and double-bounce scattering while HV backscatter is mostly sensitive to volume scattering from forest canopies [48]. In the second classification approach, non-forest class was classified further into three classes (grassland, clear-felled areas and bare-ground). The definitions and interpretation guidelines used in this methodology are given in Table 3. These classes of non-forest and forest are defined based on the field sample plots and FC GIS data points and are used for training the classification algorithm. forest class was classified further into three classes (grassland, clear-felled areas and bare-ground). The definitions and interpretation guidelines used in this methodology are given in Table 3. These classes of non-forest and forest are defined based on the field sample plots and FC GIS data points and are used for training the classification algorithm.  Table 3. Definition of forest and non-forest classes.

Forest
Land dominated by deciduous, conifer and mixed trees with canopy height ≥3 m, >20 years old and at least 50 tonnes/ha of aboveground biomass. Grassland Land dominated by non-woody annual vegetation less than 1 m in height.

Clearfelled
Open area previously occupied by forest due to stand-replacement disturbance. The area is composed of left-over dry leaves, branches and grasses with few dead stumps. This took place between December 2012 and July 2013 in few sub-compartments and planted with Norway spruce and oak seedlings.

Bareground
Land surface without any vegetation. This class includes natural and artificial bare surfaces e.g., bare soil, roads and pathways between sub-compartments in the forest.
Prior to the classification, the different scattering behaviour from forest and non-forest stands at S-band was examined for all polarisations. From the backscatter histograms of the forest and nonforest classes, a bimodal shape of the distributions was observed for all polarisations. The S-band scattering behaviour for all polarisations for the F and NF stands based on training sample points of FC sub-compartments is shown in Figure 1. Figure 5 shows box-plots illustrating the backscatter returns at all polarisations in 105 forest and 105 non-forest plots, including different non-forest classes with the training plot locations depicted in Figure 1. Stronger backscattering was recorded by the radar sensor from forested areas in comparison to non-forested ones. HH-and HV-polarisation are best at discriminating F from NF. The VV-backscatter coefficient is highest amongst all polarisations in both forest and non-forest classes. To further compare with the MIMICS-I simulated backscatter, SAR backscatter across the radar incidence range between 15° to 45° covering both forest types and bare-ground were extracted and analysed.  Table 3. Definition of forest and non-forest classes.

Forest
Land dominated by deciduous, conifer and mixed trees with canopy height ě3 m, >20 years old and at least 50 tonnes/ha of aboveground biomass.

Grassland
Land dominated by non-woody annual vegetation less than 1 m in height.

Clear-felled
Open area previously occupied by forest due to stand-replacement disturbance. The area is composed of left-over dry leaves, branches and grasses with few dead stumps. This took place between December 2012 and July 2013 in few sub-compartments and planted with Norway spruce and oak seedlings.
Bare-ground Land surface without any vegetation. This class includes natural and artificial bare surfaces e.g., bare soil, roads and pathways between sub-compartments in the forest.
Prior to the classification, the different scattering behaviour from forest and non-forest stands at S-band was examined for all polarisations. From the backscatter histograms of the forest and non-forest classes, a bimodal shape of the distributions was observed for all polarisations. The S-band scattering behaviour for all polarisations for the F and NF stands based on training sample points of FC sub-compartments is shown in Figure 1. Figure 5 shows box-plots illustrating the backscatter returns at all polarisations in 105 forest and 105 non-forest plots, including different non-forest classes with the training plot locations depicted in Figure 1. Stronger backscattering was recorded by the radar sensor from forested areas in comparison to non-forested ones. HH-and HV-polarisation are best at discriminating F from NF. The VV-backscatter coefficient is highest amongst all polarisations in both forest and non-forest classes. To further compare with the MIMICS-I simulated backscatter, SAR backscatter across the radar incidence range between 15˝to 45˝covering both forest types and bare-ground were extracted and analysed. Different algorithms have been developed for land cover classification particularly related to forest cover types. Studies have shown that the MLC algorithm can be considered more reliable than other classifier approaches [49]. The algorithm estimates class probability density functions assuming a multivariate normal distribution of the spectral signatures and often achieves better accuracy with SAR data than other classifiers [50].
The changing information content of S-band radar backscatter was examined for different spatial resolutions by re-sampling from the original 0.75 m to different spatial resolutions and repeating the classification. The spatial aggregation was performed for each polarisation band using nearest neighbour resampling with the aid of the Ordnance survey master map. The spatial scales of the forthcoming NovaSAR-S imaging modes of 6 m and 20 m were studied. For the assessment of classification accuracy, 20 non-forest and 21 forest independent points based on the ancient tree database and aerial photos were incorporated along with the training plots, providing 126 forest and 130 non-forest points to provide cross-validation (Figure 1). These sample points were collected at the boundaries of classes (forest to clear-felled, forest to bare-ground, forest to grassland) where mixed pixels with different land cover types exist and are often associated with high error [51]. Therefore, an attempt was made to highlight the accuracies of forest cover and its change maps focusing on the mixed pixels in addition to the homogenous classes. A confusion matrix was computed and the Overall accuracy, User's and Producer's accuracies and Kappa coefficient (κ) were calculated. A cross-comparison with the ALOS PALSAR-based global forest cover map provided by the Japan Aerospace Exploration Agency (JAXA) at 25 m scales was also carried out [52] (Figure 6a,b). Different algorithms have been developed for land cover classification particularly related to forest cover types. Studies have shown that the MLC algorithm can be considered more reliable than other classifier approaches [49]. The algorithm estimates class probability density functions assuming a multivariate normal distribution of the spectral signatures and often achieves better accuracy with SAR data than other classifiers [50].
The changing information content of S-band radar backscatter was examined for different spatial resolutions by re-sampling from the original 0.75 m to different spatial resolutions and repeating the classification. The spatial aggregation was performed for each polarisation band using nearest neighbour resampling with the aid of the Ordnance survey master map. The spatial scales of the forthcoming NovaSAR-S imaging modes of 6 m and 20 m were studied. For the assessment of classification accuracy, 20 non-forest and 21 forest independent points based on the ancient tree database and aerial photos were incorporated along with the training plots, providing 126 forest and 130 non-forest points to provide cross-validation (Figure 1). These sample points were collected at the boundaries of classes (forest to clear-felled, forest to bare-ground, forest to grassland) where mixed pixels with different land cover types exist and are often associated with high error [51]. Therefore, an attempt was made to highlight the accuracies of forest cover and its change maps focusing on the mixed pixels in addition to the homogenous classes. A confusion matrix was computed and the Overall accuracy, User's and Producer's accuracies and Kappa coefficient (κ) were calculated. A cross-comparison with the ALOS PALSAR-based global forest cover map provided by the Japan Aerospace Exploration Agency (JAXA) at 25 m scales was also carried out [52] (Figure 6a,b).  [53] was also utilized for cross-comparison. Both PALSAR and Landsat-based products are intended to support the interpretation of the S-band derived maps.

MIMICS-I Model Simulation Experiment
The forest canopies simulations with MIMICS-I suggest strong S-band radar total backscatter returns particularly in deciduous stands and slightly weaker scattering from coniferous canopies ( Figure 6). Total backscatter at HH-polarisation is high due to ground/trunk interactions from the deciduous canopy, higher than for a coniferous canopy. This trend is also evident for VV-polarisation. In all polarisations, backscatter from bare-ground (a function of soil moisture and roughness) is lower across the whole incidence angle range than forest canopy backscatter. This lower backscatter from soil could be due to single scattering from the soil surface. The simulation suggests that S-band backscatter can differentiate forest and non-forest due to the loss of the double-bounce scattering from ground/trunk interaction when the canopy is removed.
In comparison to the model predictions, the observed S-band radar backscatter from the airborne data reveals a similar trend to the model output but with lower sensitivity for both forest types and soil characteristics. These differences could be due to the different levels of branch (primary, secondary) and leaf density assumed in the model. The differences between simulated and observed backscatter from soil could also be due to the single scattering from soil surface as a function of soil parameters assumed in the model with limited field data on soil properties. Additionally, the observed soil backscatter could also have some influence from neighbouring classes (i.e., forest) while extracting the pixel information due to the trade-off between the size of the bare-ground (roads and pathways) and pixel spacing. The radiative scattering mechanisms at S-band simulated by MIMICS-I show a similar behaviour to the longer wavelength at L-band for forested areas [54].

Forest/Non-Forest Classification Results and Accuracy Assessment
Using the S-band radar backscatter coefficient at HH-and VV-polarisation and the RFDI in the MLC algorithm, a map of F/NF classes over Savernake forest was produced at the NovaSAR-S 6 m and 20 m resolutions (Figure 7). For the 6 m resolution, the forest class is dominant in the Savernake forest area and covers 648 ha. The forest class covers 86% of the classified image, in comparison to  [53] was also utilized for cross-comparison. Both PALSAR and Landsat-based products are intended to support the interpretation of the S-band derived maps.

MIMICS-I Model Simulation Experiment
The forest canopies simulations with MIMICS-I suggest strong S-band radar total backscatter returns particularly in deciduous stands and slightly weaker scattering from coniferous canopies ( Figure 6). Total backscatter at HH-polarisation is high due to ground/trunk interactions from the deciduous canopy, higher than for a coniferous canopy. This trend is also evident for VV-polarisation. In all polarisations, backscatter from bare-ground (a function of soil moisture and roughness) is lower across the whole incidence angle range than forest canopy backscatter. This lower backscatter from soil could be due to single scattering from the soil surface. The simulation suggests that S-band backscatter can differentiate forest and non-forest due to the loss of the double-bounce scattering from ground/trunk interaction when the canopy is removed.
In comparison to the model predictions, the observed S-band radar backscatter from the airborne data reveals a similar trend to the model output but with lower sensitivity for both forest types and soil characteristics. These differences could be due to the different levels of branch (primary, secondary) and leaf density assumed in the model. The differences between simulated and observed backscatter from soil could also be due to the single scattering from soil surface as a function of soil parameters assumed in the model with limited field data on soil properties. Additionally, the observed soil backscatter could also have some influence from neighbouring classes (i.e., forest) while extracting the pixel information due to the trade-off between the size of the bare-ground (roads and pathways) and pixel spacing. The radiative scattering mechanisms at S-band simulated by MIMICS-I show a similar behaviour to the longer wavelength at L-band for forested areas [54].

Forest/Non-Forest Classification Results and Accuracy Assessment
Using the S-band radar backscatter coefficient at HH-and VV-polarisation and the RFDI in the MLC algorithm, a map of F/NF classes over Savernake forest was produced at the NovaSAR-S 6 m and 20 m resolutions (Figure 7). For the 6 m resolution, the forest class is dominant in the Savernake forest area and covers 648 ha. The forest class covers 86% of the classified image, in comparison to 88.26% based on the FC sub-compartment database. At the spatial resolution of the 20 m resolution case, an over-estimation of the forest class was observed, resulting in an estimated 696 ha of forest (Table 4).
Remote Sens. 2016, 8, 577 12 of 21 88.26% based on the FC sub-compartment database. At the spatial resolution of the 20 m resolution case, an over-estimation of the forest class was observed, resulting in an estimated 696 ha of forest (Table 4).   The ability to discriminate forest from non-forest at S-band is due to the strong double-bounce scattering as a function of ground/trunk interaction in forest stands. Similar results have been found in the Baginton site of Southern England using the polarimetric scattering entropy H/average alpha angle α and Pauli's polarimetric decompositions [37]. This strong scattering arises predominantly from deciduous forest species due to the complex structural behaviour in comparison to the homogenous conifers. For coniferous species, a lower volume scattering was observed at S-band frequency due to the high needles and branch densities [35].
In the classified map at 20 m resolution, many non-forested areas are misclassified as forest, possibly due to the inability to capture the finer spatial details of varying surface roughness of the non-forested areas at the coarser spatial scale. The overall accuracy of the classified FNF map at 6 m resolution is 70% (κ = 0.4). Users' accuracy exceeded 70% for the non-forest class but was lower for the forest (65%). The producers' accuracies for forest and non-forest were higher 71% and 68% respectively (Table 5). At the 20 m resolution, the classified FNF map has a lower overall accuracy of 63.67% (κ = 0.27) ( Table 6). Reduced users' and producers' accuracies are evident.

Detailed Land Cover Classification Results and Accuracy Assessment
The different backscatter responses in the various non-forest land cover types in the study area provide an opportunity to investigate the sensitivity of S-band backscatter to different non-forested surface types (Figure 8 and Table 7). At 6 m resolution, grassland covers 50 ha of the study site. In general, grass leaves have erect inclination with intermediate soil brightness. The lower backscatter returns from grasses at S-band compared to soil could be due to weaker radar backscatter returns as a result of specular scattering from a smooth surface in comparison to the rougher forest canopies [34,36].  Bare-ground and clear-felled areas occupy 29 ha and 13 ha of the classified map at 6 m resolution. The distinctive backscatter signatures of bare-ground and clear-felled areas at 6 m resolution are likely due to the sensitivity of S-band backscatter to surface roughness within these areas. On the basis of the observation data, both bare-ground and clear-felled areas show similar scattering  Bare-ground and clear-felled areas occupy 29 ha and 13 ha of the classified map at 6 m resolution. The distinctive backscatter signatures of bare-ground and clear-felled areas at 6 m resolution are likely due to the sensitivity of S-band backscatter to surface roughness within these areas. On the basis of the observation data, both bare-ground and clear-felled areas show similar scattering mechanism ( Figure 5). Clear-felled areas are often covered with dead leaves and grasses, giving the area a very rough surface after logged operations. The areas in the study site were planted with Norway spruce and oak seedlings, further increasing the surface roughness. In contrast, bare-ground has no vegetation cover and appears as a relatively smooth surface to the radar.
Similarly, the 20 m resolution classification also shows little grassland (23 ha) and bare-ground (20 ha). In this classification map, clear-felled areas were misclassified as bare-ground, possibly due to the different classes being shared as mixed pixels at 20 m resolution which the algorithm could not reasonably map. This is a possible reason for the misclassification of non-forested areas as forest whereas the finer details related to bare-ground and clear-felled were captured by 6 m resolution. The different moisture content of grasses, clear-felled areas and forest classes could be another reason for misclassification because S-band backscatter is also sensitive to moisture content of the soil layer and forest canopies [30]. Closer observation of the classified maps reveals that S-band SAR differentiates forested areas with high backscatter very accurately.
The overall accuracy of the more detailed land cover maps (forest and a variety of different non-forest classes) at both spatial scales revealed a lower overall accuracy than the classified forest/non-forest maps. An overall accuracy of 52.34% (κ = 0.31) for 6 m and 36.72% overall accuracy (κ = 0.07) for 20 m are found (Tables 8 and 9). However, somewhat higher producers' accuracies are observed for all of the non-forest classes at 6 m resolution, particularly for clear-felled areas (71%).  Closer observation between the F/NF and detailed non-forest classified maps at both 6 m and 20 m resolutions confirms that the spatial resolution influences the discrimination of forested and non-forested areas from S-band backscatter signatures. When the S-band derived F/NF maps at 6 m and 20 m resolutions were compared to the global ALOS PALSAR derived F/NF map for 2010 (at 25 m resolution), a fair agreement between these maps was found. The capacity of the S-band data to detect forest cover is similar to L-band. We now conclude that S-band backscatter can be used to produce forest cover maps, similarly to other wavelengths [22]. This is validated using the global forest cover product of PALSAR at 25 m, though S-band derived maps have lower accuracy level. There is a fair agreement between S-band derived F/NF maps and the PALSAR product. However, the PALSAR forest cover map of 2015 missed out a portion of non-forest (i.e., grassland) while the majority of the forested area is correctly mapped. This may be due to the coarse resolution size of 25 m PALSAR data which missed out in comparison to the airborne datasets. Comparatively, the PALSAR based 2010 data reproduced a combination of both forest and grassland (non-forest) almost similar to the S-band derived 20 m map. This shows that S-band data, similar to L-band data, has a robust capacity to map forest and different non-forest classes as shown above, which may be due to the radiative nature of the S-band backscatter and higher spatial resolution of airborne data. The present findings also support previous results of [36][37][38].

Forest Cover Change Analysis
Forest cover change was observed in the S-band backscatter-derived forest cover maps between 2010 and 2014 ( Figure 9). Clear-felled areas in two sub-compartments (3.45 ha and 5.2 ha) were confirmed by a field visit in 2015 and the FC database. Thus, clear-felled areas are correctly classified as forest cover change in the FNF map as areas that were NF in 2014 and F in 2010. Figure 9c depicts the change map showing areas undergoing changes from forest to non-forest (red colour) between 2010 and 2014 at 6 m resolution.
A comparison with the forest loss map [53] from 2013 to 2014 from multi-temporal Landsat data at 30 m spatial resolution shows a good correspondence with the forest change between 2010 and 2014 from the S-band SAR (Figure 9d). According to the Landsat-derived forest loss data, the areas of the two clear-cut sub-compartments are 2.3 ha and 4.9 ha. The areas that have changed from forest to non-forest are somewhat different particularly along the class boundaries, where the pixels have mixed cover types (Table 10). Table 10. Area distribution of FNF cover for Savernake forest between 2010 and 2014 at 6 m resolution.

Class 6 m Area 2010 (%) 6 m Area 2014 (%)
Forest 88 80 Non-forest 12 20 Total (ha) 100 A small portion of non-forest in 2010 shows a transition to forest in 2014 (Figure 9a,b). This could be a false detection or a plantation that has grown sufficiently to change the radiometric signature at S-band. Previous studies [39,40] highlighted the good capability of S-band SAR in mapping clear-cuts and forested areas in the boreal forest of Canada and Sweden. Based on the results from the temperate forest site at Savernake forest presented here, the S-band SAR backscatter coefficient at all polarisations has shown suitable for monitoring forest cover change but lower than L-band SAR data [21,22]. The S-band SAR sensor on the NovaSAR satellite that is being built in the UK for launch in 2017 will provide image data that can be used for operational forest cover change detection and be used in regions with persistent cloud cover such as the tropics [37,55] and in the boreal winter with low sun angles [39,40].

Conclusions
High resolution S-band radar data in the context of the AirSAR campaign in Britain was utilised for assessing the mapping of forest/non-forest, different non-forest types and forest cover changes. To assist in identifying the potential of S-band radar backscatter sensitivity to forest canopies, the

Conclusions
High resolution S-band radar data in the context of the AirSAR campaign in Britain was utilised for assessing the mapping of forest/non-forest, different non-forest types and forest cover changes. To assist in identifying the potential of S-band radar backscatter sensitivity to forest canopies, the Michigan Microwave Canopy Scattering (MIMICS-I) radiative transfer model was utilised. This study was conducted in a temperate mixed forest of Savernake forest in England where different species of deciduous and coniferous stand grow. The mapping was conducted in two scenarios: one related to forest/non-forest and the second with different non-forest types (grassland, cleared-felled and bare-ground) that are found in the study area. Using the Maximum likelihood classification algorithm with HH-and VV-polarisation and the Radar Forest Degradation Index (RFDI) forest/non-forest and different non-forest types were mapped at 6 m and 20 m resolutions corresponding to StripMap and ScanSAR resolution modes of NovaSAR-S. MIMICS-I model reveals strong S-band backscatter sensitivity to the forest canopy in comparison to soil (bare-ground) across all polarisations and incidence angles. A greater sensitivity of S-band backscatter to the forest canopy was found with a broad-leaved canopy in comparison to coniferous canopy. The overall accuracy of the classification was high reaching up to 70% (κ = 0.41) for 6 m and 63% (κ = 0.27) for 20 m resolution modes. The reduction in accuracy with 20 m resolution map is due to the nature of the MLC algorithm which allocates every pixel to only one class rather than the actual mixed land cover and the size of the signal resolution resulting in mixed pixels [51,56]. This also means that the level of misclassification due to the pixels of multiple land cover classes is a function of spatial resolution. The coarser the spatial resolution, thus the lower is the overall accuracy level. However, such misclassification can be either minimized or improved by adopting an alternative mapping approach based on identifying the membership of each class to the pixel e.g., super-resolution analysis at the sub-pixel scale [57,58]. Additionally, S-band backscatter could also map different non-forest classes with overall accuracy 52% (κ = 0.31) particularly recently clear-felled areas at sub-hectare level (71% producers' accuracy). S-band backscatter between 2010 and 2014 data also found forest cover change detection for sub-compartments having 3-5 ha area due to the loss of the double-bounce scattering from ground/trunk interaction when the canopy is removed. This observation is supported by modelling predictions for deciduous species. Although these results indicate the considerable potential for mapping forest cover and non-forest types using MLC algorithm for S-band SAR backscatter, further research using the alternative classification algorithm is required.
The results show that S-band SAR data from NovaSAR-S could potentially be suitable for mapping forest/non-forest and different non-forest in details and monitor forest cover changes for temperate mixed forest considered in this paper. Current (e.g., Huan Jing-1 Constellation) and future space-borne (e.g., NovaSAR-S and NISAR) S-band SAR missions will continue to provide opportunity to generate forest cover extent including change maps at stand level from continental to global scale. This benefit can be acknowledged as the SAR system can provide continuous data in areas where cloud cover prevents observations from optical sensor data or can be integrated with optical data to provide better detection of forest dynamics such as regeneration, degraded areas and deforestation etc. However, the challenge will be the full integration of SAR data acquired in different frequencies and modes which holds enormous potential to expand on studies related to forest monitoring such as degraded canopy cover and biophysical retrieval.