Multi-Temporal and Multi-Frequency SAR Analysis for Forest Land Cover Mapping of the Mai-Ndombe District (Democratic Republic of Congo)

: The European Space Agency’s (ESA) “SAR for REDD” project aims to support complementing optical remote sensing capacities in Africa with synthetic aperture radar (SAR) for Reducing Emissions from Deforestation and Forest Degradation (REDD). The aim of this study is to assess and compare Sentinel-1 C-band, ALOS-2 PALSAR-2 L-band and combined C / L-band SAR-based land cover mapping over a large tropical area in the Democratic Republic of Congo (DRC). The overall approach is to beneﬁt from multi-temporal observations acquired from 2015 to 2017 to extract statistical parameters and seasonality of backscatters to improve forest land cover (FLC) classiﬁcation. We investigate whether and to what extent the denser time series of C- band SAR can compensate for the L-band’s deeper vegetation penetration depth and known better FLC mapping performance. The supervised classiﬁcation di ﬀ erentiates into forest, inundated forest, woody savannah, dry and wet grassland, and river swamps. Several feature combinations of statistical parameters from both, single and multi-frequency observations in a multivariate maximum-likelihood classiﬁcation are compared. The FLC maps are reclassiﬁed into forest, savannah, and grassland (FSG) and validated with a systematic sampling grid of manual interpretations of very-high-resolution optical satellite data. Using the temporal variability of the dual-polarized backscatters, in the form of either wet / dry seasonal averages or using the statistical variance, in addition to the average backscatter, increased the classiﬁcation accuracies by 4–5 percent points and 1–2 percent points for C- and L-band, respectively. For the FSG validation overall accuracies of 84.4%, 89.1%, and 90.0% were achieved for single frequency C- and L-band, and C / L-band combined, respectively. The resulting forest / non-forest (FNF) maps with accuracies of 90.3%, 92.2%, and 93.3%, respectively, are then compared to the Landsat-based Global Forest Change program’s and JAXA’s ALOS-1 / 2 based global FNF maps.


Introduction
Tropical forest represents the most important above-ground carbon pool and plays a crucial role in biodiversity, hydrological and biochemical cycles, and socio-economics for local communities. Deforestation and forest degradation are estimated to account for up to 17% of the global anthropogenic greenhouse gas emissions [1]. The forest sector is therefore an important part in climate policies [2] and the negotiations of the United Nations Framework Convention on Climate Change (UNFCCC) as stated in Article 5 of the Paris Agreement [3]. The UN initiative, Reducing Emissions from Deforestation and Forest Degradation, including conservation, sustainable management of forests, and enhancement of deeper vegetation penetration depth. The sub-objectives are to test different combinations of statistical parameters, the seasonality and different time periods in the supervised multi-variate classification, and to compare the overall performance with global FNF maps.

Region of Interest (ROI): Mai-Ndombe District in DRC
The Mai-Ndombe district is located in the Bandundu Province in the west of DRC, bordering the Congo River and the Republic of Congo. The area is very rich in biodiversity and endemic species (Bonobo). However, for many years this area has been facing high rates of deforestation and forest degradation caused by charcoal production for cities, especially due to the proximity of the capital Kinshasa, slash and burn agriculture, and industrial logging. The district covers an area of 128,789 km 2 with its center geographical coordinates around [18°31' E; 2°42" S].
It is part of the humid tropics with primary and secondary forests, interspersed by swamp forest in humid lowlands and wet meadows, inundated forests in proximity of waterbodies, especially in the north, and savannahs predominantly in upland areas. There is a distinctive dry season from June to September, followed by the rainy season until December. The region has a north-south humidity gradient with the south being drier. Figure 1 shows the location of the Mai-Ndombe district and an overview of the very-high-resolution (VHR) data for validation.

Satellite Data
The study is based on C-and L-band space-borne SAR data. C-band SAR data is from the CSAR sensor on the operational satellites Sentinel-1A and 1B (S1A and S1B) [34] from the European Copernicus program. L-band SAR is from the Phased Array type L-band SAR (PALSAR-1 and -2) [35,36] on the Advanced Land Observing Satellites (ALOS-1 and -2) from the Japan Aerospace Exploration Agency (JAXA).

Sentinel-1 A and B CSAR (2015-2017)
The two identical S1A and S1B have been launched on 3 April 2014 and 25 April 2016, respectively, circumnavigating the whole earth on a 12 day repeat cycle. Data over the ROI are a b

Satellite Data
The study is based on C-and L-band space-borne SAR data. C-band SAR data is from the CSAR sensor on the operational satellites Sentinel-1A and 1B (S1A and S1B) [34] from the European Copernicus program. L-band SAR is from the Phased Array type L-band SAR (PALSAR-1 and -2) [35,36] on the Advanced Land Observing Satellites (ALOS-1 and -2) from the Japan Aerospace Exploration Agency (JAXA). The two identical S1A and S1B have been launched on 3 April 2014 and 25 April 2016, respectively, circumnavigating the whole earth on a 12 day repeat cycle. Data over the ROI are covered only by descending paths 109, 036, and 138 and acquired only by S1A on a 12 day revisiting cycle (Figure 2). Prior to October 2016, paths 036 and 138 were only acquired sporadically. The Sentinel-1 data set processed in this study covers the period April 2015-December 2017 with acquisitions in dual polarized interferometric wide swath mode (IW) processed at Ground Range Detected (GRD) level-1 freely available from the Copernicus Open Access Hub (https://scihub.copernicus.eu/). The dual polarization is in VV and VH, i.e. emitted in vertical (first V) and received (second V/H) in vertical and horizontal (H) polarization. The processed mode in this study acquired over the ROI is the fine beam dual (FBD) strip map mode in HH and HV polarization. PALSAR-2 data is used for direct comparison and combined used with Sentinel-1 data, whereas PALSAR is used in comparison as a historical reference as it turned out to be better calibrated than PALSAR-2. PALSAR-2 calibration has been updated on 28 March 2017 [37]. However, 2015 and 2016 data from ALOS-2 have been acquired prior to the update. JAXA's correction values have been applied during the pre-processing stage but did not fully correct for the calibration error. The processed mode in this study acquired over the ROI is the fine beam dual (FBD) strip map mode in HH and HV polarization. PALSAR-2 data is used for direct comparison and combined used with Sentinel-1 data, whereas PALSAR is used in comparison as a historical reference as it turned out to be better calibrated than PALSAR-2. PALSAR-2 calibration has been updated on 28 March 2017 [37]. However, 2015 and 2016 data from ALOS-2 have been acquired prior to the update. JAXA's correction values have been applied during the pre-processing stage but did not fully correct for the calibration error.
In all, 347 PALSAR and 196 PALSAR-2 FBD scenes have been acquired, covering each pixel 5-7 times and 3-6 times, respectively (Figure 3), not considering overlap between neighboring scenes. PALSAR data have been provided in level 1.1 by both ESA and JAXA, whereas PALSAR-2 level-1.1 data has been provided only by JAXA. Both PALSAR (/-2) acquisition plans gave priority to the dry season, but for each sensor at least one measurement per pixel has been acquired over the wet season. Although ALOS and ALOS-2 are commercial satellite programs, this project benefited from several research data grants (see Acknowledgements) and nearly all PALSAR (/-2) FBD acquisitions from ascending orbits were available in this study. mode in HH and HV polarization. PALSAR-2 data is used for direct comparison and combined used with Sentinel-1 data, whereas PALSAR is used in comparison as a historical reference as it turned out to be better calibrated than PALSAR-2. PALSAR-2 calibration has been updated on 28 March 2017 [37]. However, 2015 and 2016 data from ALOS-2 have been acquired prior to the update. JAXA's correction values have been applied during the pre-processing stage but did not fully correct for the calibration error.

Pre-Processing and Mosaicking
Since the 1990s, Norut (now NORCE) has developed its in-house Generic SAR processing system, called GSAR [38], that has been used for pre-processing as well as to extract the statistical parameters for each pixel over a given data stack/time period.
All SAR data were first geocoded, radiometrically calibrated, terrain and slope corrected [39] on the same 30 m Universal Transverse Mercator (UTM) grid of zone 34S into gamma naught (γ • ) backscatter for both, co-and cross-polarized bands (γ • [copol] and γ • [xpol]), i.e., VV and VH for Sentinel-1 and HH and HV for PALSAR and PALSAR-2. The 1 arc-second (~30 m resolution) digital surface model (DSM) from the Shuttle Radar Topography Mission (SRTM) [40,41] from the year 2000 was used for the pre-processing. Since the topography of the Mai-Ndombe district does not include high mountains and only a few steep slopes, the pixels affected by topographic-induced SAR shadow or overlay are negligible. Furthermore, ALOS-2 data from 2015 and 2016 have been corrected for JAXA's initial error in radiometric calibration factors given in the data according to [37].
Following the pre-processing, all SAR data were statistically analyzed with the GSAR software suite establishing mean and variance images of γ • for each polarization band and each single-year and multi-year period in 2007-2010 for PALSAR and 2015-2017 for Sentinel-1 and PALSAR-2. Mean γ • images were also established for the wet and dry seasons using all October-May and June-September scenes, respectively. The normalized difference index (NDI) bands were calculated with the mean images by Figure 4 illustrates the five image bands that are available as features in a multi-variate classification for each sensor and time period.
Mean dual-polarized mosaics are represented in red-green-blue channels as RGB = [γ • [copol]; γ • [xpol]; NDI]. Figure 5 shows the multi-year averaged mosaics for PALSAR (2007-2010), PALSAR-2 (2015-2017), and Sentinel-1 (2015-2017) along with an optical Landsat-8 mosaic as reference. Averages based on data from the wet season show significantly different backscatter signatures compared to averages based on the dry season in some areas, especially those with less vegetation. This has been shown to be the case even on a global scale on a 0.05 decimal degree grid by [27]. The cause of this is due to the differences in ground humidity, flooding, and vegetation foliage. Examples of such different signatures for both C-and L-band SAR are shown in Figure 6, which is a detailed view of the red rectangle in Figure 5. Although the dry season is in general very distinctive, the wetness and soil humidity of the rest of year is hard to quantify because of different acquisition times and specifically few acquisitions during the wet season from PALSAR and PALSAR-2. Furthermore, climate change Remote Sens. 2019, 11, 2999 6 of 19 makes the historically established seasonal periods less reliably defined. A different approach to use the seasonal variability is therefore to use the variance of each pixel of the whole time series instead of specific seasons. Figure 7 shows the yearly averaged mosaics and RGB composites representing the variance of co-and x-pol backscatters and γ • [co-pol]. These variables are used in different multi-variate combinations in a maximum-likelihood classification (MLC) [39] for forest land cover.
or overlay are negligible. Furthermore, ALOS-2 data from 2015 and 2016 have been corrected for JAXA's initial error in radiometric calibration factors given in the data according to [37].
Following the pre-processing, all SAR data were statistically analyzed with the GSAR software suite establishing mean and variance images of γ° for each polarization band and each single-year and multi-year period in 2007-2010 for PALSAR and 2015-2017 for Sentinel-1 and PALSAR-2. Mean γ° images were also established for the wet and dry seasons using all October-May and June-September scenes, respectively. The normalized difference index (NDI) bands were calculated with the mean images by (1) Figure 4 illustrates the five image bands that are available as features in a multi-variate classification for each sensor and time period. Averages based on data from the wet season show significantly different backscatter signatures compared to averages based on the dry season in some areas, especially those with less vegetation. This has been shown to be the case even on a global scale on a 0.05 decimal degree grid by [27]. The cause of this is due to the differences in ground humidity, flooding, and vegetation foliage. Examples of such different signatures for both C-and L-band SAR are shown in Figure 6, which is a detailed view of the red rectangle in Figure 5. Although the dry season is in general very distinctive, the wetness and soil humidity of the rest of year is hard to quantify because of different acquisition times and specifically few acquisitions during the wet season from PALSAR and PALSAR-2. Furthermore, climate change makes the historically established seasonal periods less reliably defined. A different approach to use the seasonal variability is therefore to use the variance of each pixel of the whole time series instead of specific seasons. Figure 7 shows the yearly averaged mosaics and RGB composites representing the variance of co-and x-pol backscatters and γ°[co-pol]. These variables are used in different multi-variate combinations in a maximum-likelihood classification (MLC) [39] for forest land cover.

Maximum-Likelihood Classification into Forest and Land Covers (FLC)
During fieldwork in March 2013 northwest of Lake Mai-Ndombe [42], several areas that could be directly compared to SAR signatures were identified. Based on these observations and VHR optical data from GoogleEarth and RapidEye satellites, shapefile polygons of homogeneous training areas for six different land cover classes were assigned and used for a supervised MLC [43]. The six forest land cover (FLC) classes identified are forest, inundated forest, savannahs, predominantly dry grassland, predominantly wet grassland, and river swamps. Figure 8 shows examples of such areas that were recognized. A new water mask was constructed by using existing water mask shapefiles of lakes and rivers, the water mask extracted from the Facet Atlas [44], and by thresholding the cross-polarized SAR data.

Maximum-Likelihood Classification into Forest and Land Covers (FLC)
During fieldwork in March 2013 northwest of Lake Mai-Ndombe [42], several areas that could be directly compared to SAR signatures were identified. Based on these observations and VHR optical data from GoogleEarth and RapidEye satellites, shapefile polygons of homogeneous training areas for six different land cover classes were assigned and used for a supervised MLC [43]. The six forest land cover (FLC) classes identified are forest, inundated forest, savannahs, predominantly dry grassland, predominantly wet grassland, and river swamps. Figure 8 shows examples of such areas that were recognized. A new water mask was constructed by using existing water mask shapefiles of lakes and rivers, the water mask extracted from the Facet Atlas [44], and by thresholding the crosspolarized SAR data. The MLC has been chosen for classification because of several reasons; among those are: MLC is based on probability calculations assuming a Gaussian distribution of the signature for each class, which simplifies future operational implementation and allows for automatic use once the signatures are known for each class and variable; MLC is a general and widely distributed tool in most geographic information system (GIS) and image processing software and the results should therefore be easily reproducible; MLC [45] performed clearly better compared to all, but neural network classifier (NNC), supervised classification methods from the ENVI software (https://www.harrisgeospatial.com/Software-Technology/ENVI) suit that have been test run for this study. The NNC results were very similar to MLC results in accuracy. The calculated probabilities easily allow assigning super classes when the reference validation data does not include all six classes. All classification results have then been filtered with a 3 × 3 pixel majority window, so that less than five adjacent forest pixels, a minimum of 0.45 ha, would not be considered as forest according to the 0.5 ha forest area definition The MLC has been chosen for classification because of several reasons; among those are: MLC is based on probability calculations assuming a Gaussian distribution of the signature for each class, which simplifies future operational implementation and allows for automatic use once the signatures are known for each class and variable; MLC is a general and widely distributed tool in most geographic information system (GIS) and image processing software and the results should therefore be easily reproducible; MLC [45] performed clearly better compared to all, but neural network classifier (NNC), supervised classification methods from the ENVI software (https://www.harrisgeospatial.com/Software-Technology/ENVI) suit that have been test run for this study. The NNC results were very similar to MLC results in accuracy. The calculated probabilities easily allow assigning super classes when the reference validation data does not include all six classes. All classification results have then been filtered with a 3 × 3 pixel majority window, so that less than five adjacent forest pixels, a minimum of 0.45 ha, would not be considered as forest according to the 0.5 ha forest area definition The aim of the study is to investigate different combination of variables in a multi-variate MLC classification from single C-/L-band and multi-frequency, C-, and L-band combined. The different combinations of variables are based on:

Validation and Inter-Comparison Approach
The general approach to estimate the accuracy and quantifying uncertainty of land cover maps is done by comparing the produced maps with a reference sample data set via a confusion matrix [46]. The main purpose in our case is however to inter-compare the different combinations of variables in the MLC. As reference data set, three Pléiades images, two from 19 and one from 21 November 2016 (© Centre National D'Etudes Spatiales (CNES)/AirbusDS) in 50 cm resolution and a SPOT-5 scene (©CNES) dated 25 June 2015 in 5 m resolution from the SPOT5-Take5 program were made available through ESA and the Centre National D'Etudes Spatiales (CNES) (Figure 9). We follow the approach of [47] based on a systematic sampling grid. As we have trained the MLC with in situ knowledge, RapidEye imagery from 2013, and GoogleEarth, the SPOT5 and Pléiades VHR data are independent and solely used for validation. The Pléiades images cover an area of about 1000 km 2 each and have been sampled on a regular 2.1 km grid for visual interpretation. The SPOT-5 scene covers an area of about 3600 km 2 and has been sampled on a 4.2 km grid. The FLC maps include six land cover/vegetation classes. A reliable manual interpretation from optical satellite images did in general not allow a better differentiation than into forest, savannah, and grassland (FSG) as the other classes were not abundant in the VHR images and difficult to interpret visually. The samples where therefore interpreted manually into FSG samples considering the majority of each land cover in a square area of 0.5 ha, i.e., a 70 m × 70 m square, following the forest definition. In total, 924 samples were interpreted: 709 forest, 144 savannah, and 71 grassland samples.
Hence, we first reclassified the FLC land cover maps of the MLC into three-classes FSG maps; forest therefore includes the original "forest" and "inundated forest" class, grassland includes the dry and wet grassland classes and to reclassify "river swamp", we choose the forest, savannah, or grassland class with the highest probability from the MLC. For the final forest/non-forest (FNF) validation, savannah and grassland are considered as non-forest. As a final step, all classification in 30 m resolution are filtered with a 3 × 3-pixel majority window.

Results
The MLC was applied on each SAR sensor dataset alone, PALSAR (2007-2010) as a historic reference, PALSAR-2, Sentinel-1, and on a combined PALSAR-2/Sentinel-1 data set (2015-2017) with the four combinations of variables described in Section 2.4. Figure 10 shows a comparison of the results from Sentinel-1, ALOS-2, and Sentinel-1/ALOS-2 combined using the multi-year statistical parameters, mean and variance, of the co-and cross-polarization backscatters γ°. As Table 1 shows, this feature combination gave in general the highest accuracies.

Results
The MLC was applied on each SAR sensor dataset alone, PALSAR (2007-2010) as a historic reference, PALSAR-2, Sentinel-1, and on a combined PALSAR-2/Sentinel-1 data set (2015-2017) with the four combinations of variables described in Section 2.4. Figure 10 shows a comparison of the results from Sentinel-1, ALOS-2, and Sentinel-1/ALOS-2 combined using the multi-year statistical parameters, mean and variance, of the co-and cross-polarization backscatters γ • . As Table 1 shows, this feature combination gave in general the highest accuracies.
The left panel of Figure 10 shows that Sentinel-1 results can be quite noisy in the forested areas. The reason for this could be that C-band SAR penetrates only a little in the canopy and its signature therefore reflects also height differences in the upper canopy surface in inhomogeneous mixed forest that L-band will not see as it penetrated deeper in the canopy and will smooth out the canopy surface effects. C-band however seems to better distinguish lower biomass land covers, such as savannah, wet, and dry grassland up to a certain biomass level, above which C-band will saturate and classify as forest. Somewhat surprisingly, it seems that C-band also distinguishes inundated forest from forest better than L-band does. This could be due to this specific type of inundated forest, which might be a much lower biomass forest than the surrounding tropical forest. This still needs further investigation. The black arrow in the PALSAR-2 result indicates a vast misclassification area into river swamp in a particular satellite path which we anticipate is mainly due to the calibration factor error in 2015 and 2016 ALOS-2 data or do to the low ALOS-2 coverage during the wet seasons. The use of both sensors, ALOS-2 and Sentinel-1 (right panel in Figure 10), clearly seems to better distinguish the land cover areas, smoothing out the noise in forest areas of Sentinel-1 and still keeping a better distinction between savannah, wet, and dry grassland. Figure 11 shows the derived FSG and FNF results from the same MLC variable combination and enlarged area as in Figure 10. The left panel of Figure 10 shows that Sentinel-1 results can be quite noisy in the forested areas. The reason for this could be that C-band SAR penetrates only a little in the canopy and its signature therefore reflects also height differences in the upper canopy surface in inhomogeneous mixed forest that L-band will not see as it penetrated deeper in the canopy and will smooth out the canopy surface effects. C-band however seems to better distinguish lower biomass land covers, such as savannah, wet, and dry grassland up to a certain biomass level, above which C-band will saturate and classify as forest. Somewhat surprisingly, it seems that C-band also distinguishes inundated forest from forest better than L-band does. This could be due to this specific type of inundated forest, which might be a much lower biomass forest than the surrounding tropical forest. This still needs further investigation. The black arrow in the PALSAR-2 result indicates a vast misclassification area into river swamp in a particular satellite path which we anticipate is mainly due to the calibration factor error in 2015 and 2016 ALOS-2 data or do to the low ALOS-2 coverage during the wet seasons. The use of both sensors, ALOS-2 and Sentinel-1 (right panel in Figure 10), clearly seems to better distinguish the land cover areas, smoothing out the noise in forest areas of Sentinel-1 and still keeping a better distinction between savannah, wet, and dry grassland. Figure 11 shows the derived FSG and FNF results from the same MLC variable combination and enlarged area as in Figure 10. The FSG and FNF results shown in Figure 11 and the corresponding results from all different applied variable combinations have been validated according to the approach described in section 2.5. and are presented in Table 1. The different feature combinations using Sentinel-1 data have accuracies of 82% ± 3% with a kappa of 0.48% ± 0.07% and 87% ± 4% (kappa of 0.61 ± 0.1) for FSG and FNF, respectively. For the L-band SAR data, both from ALOS and ALOS-2, FSG accuracies are around 90% ± 1% (kappa 0.68 ± 0.03) and FNF accuracies are around 92% ± 1% (kappa 0.75 ± 0.02). For Sentinel-1 data, where we have continuous time series on a 12-day cycle, the use of temporal variability, either by dividing into a wet and dry data set or by using the variance over the whole data set improves the FSG and FNF classification results significantly by about 5-6 percent points and increasing kappa from 0.42 to 0.55 and from 0.52 to 0.70, respectively. For L-band SAR, which is  The FSG and FNF results shown in Figure 11 and the corresponding results from all different applied variable combinations have been validated according to the approach described in Section 2.5. and are presented in Table 1. The different feature combinations using Sentinel-1 data have accuracies of 82% ± 3% with a kappa of 0.48% ± 0.07% and 87% ± 4% (kappa of 0.61 ± 0.1) for FSG and FNF, respectively. For the L-band SAR data, both from ALOS and ALOS-2, FSG accuracies are around 90% ± 1% (kappa 0.68 ± 0.03) and FNF accuracies are around 92% ± 1% (kappa 0.75 ± 0.02). For Sentinel-1 data, where we have continuous time series on a 12-day cycle, the use of temporal variability, either by dividing into a wet and dry data set or by using the variance over the whole data set improves the FSG and FNF classification results significantly by about 5-6 percent points and increasing kappa from 0.42 to 0.55 and from 0.52 to 0.70, respectively. For L-band SAR, which is in general clearly better suited for forest and land cover classification, using the variability improves the results only by 1-2 percent points for ALOS. For ALOS-2, there is no improvement, i.e., an insignificant decrease and increase in accuracy and kappa, respectively. We assume that the main reason that there has been no improvement by using multi-year data including differentiating wet and dry seasons and statistics compared to only 2017 data is due to (1) the radiometric correction issue for 2015 and 2016 data and (2) that, also valid for ALOS, ALOS-2 acquisitions have a strong acquisition focus on the dry season, and much less acquisitions per pixel in general which could be the main reason for the quite stable results. This will be further discussed in Section 4. Tables 2 and 3 show the confusion matrixes for the FSG and FNF classification using the Sentinel-1 and ALOS-2 co-and cross polarized 2015-2017 average backscatters and their variances (last line in Table 1). They indicate that there is a tendency to classify into higher biomasses, i.e., grassland toward savannah and savannah toward forest) and the producer accuracy increases with higher biomass classes. However, there seems to be no strong bias in the user accuracies. Note that majority filtering took place after classifying FNF from FSG, and Table 3 is therefore not directly derived from Table 2.

Inter-Comparison between Single and Multi-Frequency SAR Results
In agreement with earlier studies, it is no surprise that the results reflect that L-band SAR is in general better suited for forest mapping than C-band SAR as C-band SAR saturates at much lower biomass vegetation than L-band. The results however show that the use of statistics of the better temporal resolution of C-band Sentinel-1 time series acquisition compared to less frequent L-band PALSAR can make up for about half of the difference in accuracies.
The general impression for all SAR sensors is that temporal filtering to reduce speckle and single acquisition conditions improves classification results. As long as we can also neglect inter-annual forest change compared to the mapping errors for example when building a baseline forest map, integration over a multi-year time period for data sets with very few yearly acquisitions, like PALSAR-1 and -2, also improves the results. This is especially the case if the temporal resolution of the data series does not even resolve the seasonality between dry and wet periods. Dividing a data stack into specific seasonal averaged products or extracting the statistics, even if that requires a multi-year period, clearly increases the accuracy and kappa value. This shows clearly in the PALSAR-1 and Sentinel-1 results, with improved accuracies of 1-2 percent points and 4-5 percent point, and improved kappa values by about 0.05 and 0.1, respectively. We assume that the reason why this does not show in the PALSAR-2 results is partly because of the calibration issue of 2015 and 2016 PALSAR-2 level 1.1 products, as well as the fact that PALSAR-2 acquisitions are too few (3-6 acquisitions per pixel) to define significant statistical values. Unfortunately, our available PALSAR-2 acquisitions in 2017 did not allow for a division into wet and dry mosaics over the whole area. All four PALSAR-2 feature combinations show therefore very similar results.
The temporal resolution of Sentinel-1 acquisitions of 12 days, i.e., 30 acquisitions per year, would probably allow integrating the data set also over a single year instead of a multi-year period to clearly detect seasonal differences and extract significant statistical values. For comparability with the L-band data set however, this has not been studied in detail, but the slightly better results of the 2017 averaged mosaics versus the 2015-2017 averaged mosaics clearly support this assumption.
In general, our approach to either use seasonal averages or statistics (mean and variance) is very similar as we assume that the main variance of SAR signatures is due to differences in humidity and foliage, especially in low vegetated areas, and results therefore should be and are very similar. Although, it seems that using the yearly statistical values, mean backscatter and variance, gives slightly better results. As the locals of the region reported to us during the field campaign that the rainy and dry season patterns have changed in their occurrence during the year, probably due to climate change, our definition of wet and dry seasons, based on historic weather observations might not correctly apply and even less in the future. The variability between wet and dry conditions might therefore be better reflected with statistical parameters instead, or each observation would need to be confirmed individually by meteorological observation.
In this study, we have only taken temporal statistics into account and not textural statistics in space which would have reduced our spatial resolution. The temporal resolution of Sentinel-1 could even allow to consider other statistical parameters, for example different percentiles and the temporal distribution of SAR signatures. Such dense time series are of course also better suited to clearly detect and confirm changes with time series analysis tools such as the Bfast algorithm [19,48]. A more specific study of the backscatter time series and seasonality could also be of particular interest in regard to forest ecology or physiology but is beyond the scope of this paper.
The highest accuracies are obtained when combining C-and L-band in a multi-frequency approach. Nevertheless, the results show that C-band does not seem to improve the L-band results when only using yearly or multi-year mosaics without using the seasonality or statistics. When C-and L-band are combined, it is clear that the Sentinel-1's strength lies in the dense time series, and that their temporal statistics and the multi-seasonal division are the main reasons of the improvements in accuracies compared to L-band ALOS-2 alone. Both the FSG and the FNF maps improve by about one percent point combining C-and L-band SAR compared to L-band alone.

Comparision with Global Forest Maps
A main reference data set for forest and forest change maps is the GFC program [6] based on the Landsat archive. It is also the data set that is widely used to establish baseline forest reference maps for the carbon emission reporting in REDD+ programs. JAXA provides yearly forest maps based on ALOS and ALOS-2 [31,32]. Both data sets have been validated with the same VHR data set and compared to our SAR results (Table 4). According to the final DRC Emission Reductions Program Document from 2016 [49], the forest definition has been set to a minimum crown cover of 30%, minimum land area 0.5 ha, and minimum tree height of 3 m. However, in the Emission Reductions Program Idea Note (ERPIN) [50] and based on [6], forest (primary and secondary) is defined with a crown cover higher than 50% and woodland, defined as crown cover of 26%-50%, is defined as non-forest. Our validation shows clearly that a crown cover higher than 50% corresponds better to a forest definition than a minimum crown cover of 30% in GFC data. Table 4 summarizes the accuracy assessment of these two global forest maps compared to our SAR-based results. Shimada [51] reported global accuracies for ALOS-2 forest maps of 88.21% and 87.77% for the years 2015 and 2016, respectively, which are in good agreement with our validation results. Table 4. Summary of the accuracy assessment of global forest maps from GFC (Landsat) and JAXA