Deep learning models for estimating volume and Lorey's height across Nordic countries using optical and SAR satellite images

Deep learning models for estimating volume and Lorey's 1 height across Nordic countries using optical and SAR 2 satellite images 3 Zsofia Koma1 Oleg Antropov2 Oliver Cartus3 Jukka Miettinen2 and Johannes Breidenbach1 4 1 Norwegian Institute for Bioeconomy (NIBIO), Division of Forest and Forest Resources, National 5 Forest Inventory, Høgskoleveien 7, 1433 Ås, Norway 6 2 VTT Technical Research Centre of Finland, P.O. Box 1000, 02044 Espoo, Finland 7 3Gamma Remote Sensing, Worbstrasse 225, 3073 Gümligen, Switzerland 8 Target journal: International Journal of Applied Earth Observation and Geoinformation 9 Highlights 10 - A large-scale, fine-resolution forest resource map was produced in Norway from multisource 11 optical and SAR imagery, trained on an ALS-based reference map using a UNet deep learning 12 framework. 13 - The best UNet model achieved coefficients of determination (R²) of 0.63 and 0.55 for timber 14 volume and Lorey’s height, respectively, compared to 0.39 and 0.37 from the reference 15 approach (k-nearest neighbours with k = 7). 16 - We also analyse the geographical transferability of the pre-trained model on Finnish data to 17 fine-tune for Norwegian conditions. 18 Abstract 19 Spatially explicit information on forest resources and structure is essential for sustainable forest 20 management and evidence-based policy-making. In the Nordic region, large-scale forest mapping often 21 relies on integrating National Forest Inventory (NFI) field plots with airborne laser scanning (ALS) data. 22 However, the infrequent coverage of nationwide ALS campaigns limits their use for continuous 23 monitoring. Satellite imagery, with its high temporal and spatial resolution, provides a promising 24 alternative. We evaluate UNet-based deep learning models trained on wall-to-wall ALS-derived forest 25 resource maps for predicting volume and Lorey’s height in Norway using optical (Sentinel-2) and SAR 26 (Sentinel-1, PALSAR-2) data. The UNet models, trained on both Finnish and Norwegian ALS maps, are 27 benchmarked against kNN models based on Norwegian NFI data. Transfer learning is further explored 28 by fine-tuning models using Norwegian NFI plots. Model uncertainties are assessed using 541 reserved 29 NFI plots and 44 independent validation stands mainly including high-volume boreal forests (>200 m³ 30 ha⁻¹). The best-performing UNet model, trained on Norwegian ALS data, achieved an R² of 0.63 for 31 volume and 0.55 for Lorey’s height, consistently outperforming kNN. Fine-tuning improved model 32 transferability, with gains in R² of up to 0.13 for volume and 0.46 for Lorey’s height, when adapting the 33 Finnish model to Norwegian conditions. Utilizing SAR data alongside optical data enhanced model 34 accuracy, with R2 improvement ranging from 0.03 to 0.19. Overall, our findings demonstrate the 35 potential of UNet models trained on wall-to-wall ALS maps, offering a transferable approach for forest 36 resource mapping across Nordic countries. 37 Keywords: UNet, deep learning, forest resource mapping, satellite imagery, SAR, multisource 38 39 1. Introduction 40 Quantifying the status and trends of forest resources and structure—such as timber volume and Lorey’s 41 height—is crucial for informing evidence-based policy and management decisions across local, 42 national, and international levels (Tomppo et al., 2010; Vidal et al., 2016a). Sampling-based National 43 Forest Inventories (NFIs) provide this information at national and regional scales, forming the 44 foundation for strategic decision and policy-making regarding key ecosystem services, including 45 bioenergy potential, biodiversity, and changes in carbon stock (Vidal et al., 2016b; Breidenbach et al., 46 2020). However, because NFIs are sampling-based surveys, their capacity to produce reliable estimates 47 at local scales is restricted (Breidenbach and Astrup, 2012; Astrup et al., 2019). To bridge this gap, NFI 48 field plots are often combined with remotely sensed data to model relationships between field49 observed forest attributes and remote sensing-based predictor variables, enabling mapping between 50 NFI plot locations (McRoberts and Tomppo, 2007; Kangas et al., 2018). This approach aligns with the 51 requirements of several international policy frameworks—including the EU Forest Strategy for 2030, 52 the EU Biodiversity Strategy, and the EU Deforestation Regulation—which emphasize the need for 53 spatially explicit forest monitoring at both national and local scales. Consequently, integrating remotely 54 sensed data has become essential for producing detailed forest resource maps and information. 55 Some of the earliest national forest maps were based on a combination of NFI and Landsat data 56 (Gjertsen, 2007; McRoberts and Tomppo, 2007). In many countries, national forest resource maps are 57 today based on integrating field plots from NFIs with country-wide Airborne Laser Scanning (ALS) data 58 (Kangas et al., 2018). ALS offers detailed 3D information on forest structure, enabling accurate 59 estimation of key attributes such as timber volume and Lorey’s height. However, national ALS 60 campaigns lack the temporal frequency needed for effective forest monitoring. For example, Norway’s 61 nationwide ALS campaign required approximately ten years to complete full coverage (Hauglin et al., 62 2021), compared to a seven-year timeframe in Sweden (Nilsson et al., 2017). In Finland, the future ALS 63 campaigns are planned to be conducted in nine-year cycles (NLS, 2025). In contrast, Earth Observation 64 (EO) datasets—such as optical imagery from Sentinel-2 and Synthetic Aperture Radar (SAR) data from 65 Sentinel-1 or ALOS-2 PALSAR-2—offer both high temporal and spatial coverage. Several studies have 66 utilized optical remote sensing for forest resource mapping, such as Sentinel-2 or LANDSAT (Astola et 67 al., 2019; Lang et al., 2019; Miettinen et al., 2021) or have combined EO with ALS data (Esteban et al., 68 2019; Astola et al., 2021). SAR data with clear sensitivity to forest structure have also been used for 69 forest mapping, primarily at stand-level or coarser mapping units and in wide-area studies (Santoro et 70 al., 2011; Antropov et al., 2017; Ge et al., 2023b). Recent studies have shown that the combination of 71 optical and SAR data improves the quantification of forest structural characteristics (Wittke et al., 2019; 72 Persson et al., 2021; Ge et al., 2022, 2023a), especially if SAR operates in lower frequencies (e.g., L73 band used on board ALOS-2), which opens new possibilities for improving the accuracy of maps for 74 estimating forest attributes. 75 EO-based forest resource maps have been created using statistical methods such as linear regression, 76 k-nearest neighbours (kNN), and ensemble-based machine learning techniques such as random forests. 77 Studies using only Sentinel-2 data with these approaches have reported RMSEs ranging from 27.2% to 78 59.3% for estimates of volume and Lorey’s height in boreal forest contexts (Astola et al., 2019; 79 Miettinen et al., 2021; Persson et al., 2021). When using SAR data alone, RMSEs for volume range from 80 32% to 54.3% (Wittke et al., 2019; Ge et al., 2023b; Persson et al., 2021). One study systematically 81 tested the integration of optical and SAR datasets using kNN method and reported improved accuracies 82 in Sweden, with volume RMSE reduced from 37.9% to 30.2% in boreal forests (Persson et al., 2021). 83 However, a recurring limitation is, for instance, a saturation effect in timber volume predictions— 84 particularly above 300 m³/ha (Santoro et al., 2011) —which contributes to lower overall accuracy 85 compared to models based on airborne laser scanning (ALS) data (Wittke et al., 2019). Recent advances 86 in Deep Learning (DL) show promise for improving EO-based forest resource mapping. Ge et al. (2023a) 87 demonstrated that the prediction of Lorey’s height using Sentinel-2 data alone, as well as combined 88 Sentinel-2 and SAR data, improved by 0.3% to 1.6% RMSE when applying the UNet DL method 89 compared to traditional approaches such as kNN, random forest, or linear regression. Similarly, Astola 90 et al. (2021) reported RMSE improvements of 7.9% for volume and 2.5% for height when using DL 91 models compared to reference methods. Nevertheless, DL remains underutilized in regression tasks for 92 EO-based forest attribute estimation, especially when compared to its more widespread application in 93 classification tasks such as disturbance mapping (Yuan et al., 2020). 94 One of the key challenges in applying DL to forest resource mapping is the limited availability and 95 quality of suitable training data (Ball et al., 2017; Li et al., 2019). NFI datasets are often considered weak 96 labels in the context of DL because they lack spatial context due to a relatively small plot size. To address 97 this challenge, transfer learning has emerged as a promising strategy in various remote sensing tasks 98 e.g. (Wurm et al., 2019; Zhou et al., 2023; Kuzu et al., 2024), but only a few studies have been used for 99 in the context of forest resource mapping (Astola et al., 2021; Ge et al., 2023a). Instead of training a 100 model solely on data from the target region, transfer learning allows the use of a pre-trained model— 101 developed on a related task —which can then be fine-tuned using a smaller set of local training data. 102 This approach leverages the generalizable features learned during pre-training, enabling effective 103 model adaptation even with limited labelled data. In addition, ALS-based forest resource maps can 104 offer high-quality, wall-to-wall training data that can complement or substitute traditional plot-based 105 labels (Ge et al., 2022, 2023a). For example, Ge et al. (2023a) demonstrated the successful training of 106 models in Finnish Lapland using ALS-derived maps, along with optical and SAR satellite imagery, 107 followed by fine-tuning with plot-level measurement data in southern Finland. To further evaluate the 108 effectiveness of this approach, applying transfer learning to neighbouring countries—such as Norway— 109 provides an ideal test case. 110 In this study, we develop and evaluate the potential of UNet-based deep learning models in multiple 111 EO data combinations to estimate timber volume and Lorey’s height using optical (Sentinel-2) and SAR 112 (Sentinel-1 and PALSAR-2) satellite imagery. ALS-based forest resource maps serve as wall-to-wall 113 training data for models trained in both Finland and Norway. We investigate transfer learning by fine114 tuning ALS-based pre-trained models with Norwegian NFI plots, both by transferring the Finnish model 115 to Norwegian conditions and exploring further options for improving the Norwegian model. The 116 performance and uncertainty of the UNet models trained on existing forest resource maps are 117 compared with kNN models trained on NFI field plots. Model uncertainty is assessed using both a test 118 dataset of NFI plots, which was not used in modelling, and an independent validation dataset of forest 119 stands. Our research questions are: (1) Which modelling method -- UNet or kNN -- achieves the highest 120 accuracy? (2) Can transfer learning, where NFI field plots are incorporated in the model, improve the 121 UNet accuracy, especially in the case of applying the Finnish model to Norwegian conditions? (3) Can 122 the inclusion of SAR imagery enhance the model’s prediction accuracy? 123 2. Materials and Methods 124 The overall workflow developed for this study (Fig. 1) consists of four main processing steps: (a) kNN 125 modelling; (b) pretraining UNet models using ALS-based forest resource maps from Finland and 126 Norway; (c) fine-tuning UNet models from Finland and Norway for the Norwegian study site; and (d) 127 assessing the uncertainty associated with each model type (k-NN, pretrained UNet, and finetuned 128 UNet). 129 130 Figure 1. Overview of the workflow used for assessing the accuracy of different models (kNN, UNet, finetuned UNet). The 131 input datasets are marked with rectangles (satellite imagery data, forest resource maps, Norwegian NFI plots, and 132 Norwegian validation stands), dark grey indicating input for training the models and light green for validation models. The 133 intermediate models (kNN, UNet, and finetuned UNet) are marked with a snipped rectangle. The main processing steps are 134 shown with rounded rectangles (a-d corresponding to the main text). V is volume, and H is Lorey’s height, S2 is Sentinel-2 135 optical images, S1 is Sentinel-1 SAR images, and P2 is PALSAR-2 SAR images. FI = Finland and NO = Norway. 136 2.1. Datasets 137 2.1.1. Satellite imagery data and pre-processing 138 ESA Copernicus Sentinel-2 satellites operating with the Multi-Spectral Instrument (MSI) provide optical 139 imagery data on 13 spectral bands at a 2–3-day revisit frequency in the Nordic countries, in visible, 140 near-infrared, and the short-wave infrared part of the electromagnetic spectrum. In this study, we used 141 the Level-2 surface reflectance products provided by ESA at the visible and near infrared bands (B2, B3, 142 B4, B5, B8) and short-wave infrared bands (B11, B12). The 20 m resolution bands (B5, B11, B12) were 143 resampled into 10 m resolution using the bilinear interpolation method for matching the resolution 144 with the other bands (10 m). The Sentinel-2 data were further processed with the Terramonitor 145 compositing script (Miettinen et al., 2021), producing growing season composites by weighting based 146 on cloudiness, haze, and shadows. The compositing process also delivered a quality flag that was used 147 for filtering cloud-affected or bad-quality pixels. 148 In this study, we have used two types of SAR datasets. The first consisted of C-band backscatter imagery 149 acquired by Sentinel-1A in Interferometric Wide (IW) swath mode, using VV and VH polarisation. The 150 images in Ground-Range-Detected (GRD) format were pre-processed using the GAMMA Remote 151 Sensing software (Werner et al., 2000), which resulted in geocoded and radiometric terrain corrected 152 g0 backscatter images. The images were subsequently resampled to 10 m to match the Sentinel-2 153 imagery. The Sentinel-1 data acquired within a year (30 to 31 images) were then used to calculate the 154 annual mean backscatter per polarization. Maps depicting areas of layover or shadow were applied to 155 mask such areas. 156 The second type of SAR dataset was based on the Japan Aerospace Exploration Agency (JAXA), which 157 releases free and open annual mosaics of the L-band data acquired by the Phased-Array type L-band 158 SAR (PALSAR-2) onboard the Advanced Land Observing Satellite-2 (ALOS-2).The backscatter mosaics, 159 with a resolution of ~ 25 × 25 m², are produced from imagery acquired in Fine Beam Dual Polarization 160 mode (HH and HV), using stripmap mode under JAXA’s global basic observation scenario. The images 161 used to generate the mosaics were radiometric terrain-corrected, and geocoded to a geographic spatial 162 reference system, as with (Shimada and Ohtaki, 2010). In this study, the HH and HV mosaics were 163 resampled to a 10 m grid and masked for layover and shadow using the layover/shadow maps produced 164 by JAXA. 165 All satellite imagery datasets were processed for the study areas in Finland and Norway. In Finland, the 166 datasets were processed for the year 2020, and in Norway for 2021, to align as closely as possible with 167 the timing of the ALS-based forest resource maps. 168 2.1.2. Finnish reference datasets 169 The Finnish UNet model was trained based on the wall-to-wall Finnish ALS-based forest resource map 170 produced by the Finnish Forest Centre using ALS acquisitions from 2020 and associated field inventory 171 plots as training for k-NN based approaches (Maltamo and Packalen, 2014). The map covers the core 172 central and southern regions of Finland (Fig. 2), capturing boreal forests in 16 m spatial resolution. The 173 study area has relatively small elevation differences, with the highest points generally less than 300 m 174 above sea level. The forest in the reference areas has around 40% pine, more than 30% spruce, and the 175 rest deciduous trees. The average forest volume across the study sites was 120 m3 ha-1, with an average 176 height of 14 m. 177 2.1.3. Norwegian reference datasets 178 The Norwegian study site covers approximately 210,000 km² of land, encompassing the South-Eastern, 179 Central, Southern, and Western regions, extending from the county of Trøndelag and south (Fig. 2). It 180 is located in the boreal climate zone and contains most of the productive forest of Norway. The boreal 181 forests are dominated by three main species: Norway spruce, Scots pine, and deciduous trees, primarily 182 birch, which are often mixed with coniferous species. The study site has a diverse topographical 183 variation. 184 185 Figure 2 The Norwegian study area (highlighted in green on the Nordic map) is where the different model types—kNN, UNet, 186 and UNet fine-tuned —were evaluated. Green dots represent the NFI plots used for testing, while pink coloured dots indicate 187 the training NFI plots used for the kNN model and UNet fine-tuning. Brown areas mark the independent validation stands. 188 Panel A shows a zoomed-in view of the study area, with panels B and C highlighting two example validation stands used in 189 the accuracy assessment. The background maps are the ALS-based timber volume map from Kilden (NIBIO, 2025) and 190 Google satellite imagery in panel A. 191 Different field observation datasets were utilized in this study (Fig.2, Table 1). First, the Norwegian NFI 192 is a continuously operating inventory based on a permanent sample grid, where each field plot is 193 remeasured every five years (Breidenbach et al., 2020). The NFI field measurements are carried out 194 within circular 250 m2 plots where the diameter at breast height (dbh) and species of all trees with a 195 dbh >= 5 cm are recorded, and the height of ten trees is measured to calculate timber volume and 196 Lorey’s height. Only NFI plots measured in 2021 within a valid remote sensing-based forest mask were 197 included in the study, which resulted in 1566 NFI plots. 541 NFI plots were set aside for testing and not 198 used for kNN or DL modelling, and the remaining 1025 NFI plots were used in model training (Table 1). 199 Additional stand-level field measurements (Fig. 2, Table 1), independent of the NFI, were used for 200 validation (Koma and Breidenbach, 2025). The dataset includes two subsets: (a) mature stands in Asker 201 (eastern region) and Alver (western region), where 10–15 plots per stand were sampled following NFI 202 protocols. Mean timber volume and Lorey’s height were calculated to characterise each stand. (b) Long203 term field trials assessing forest management methods, consisting of 1–30 adjacent rectangular stands. 204 Within each stand, dbh and species were recorded for all trees, and height was measured for every 205 fourth tree. Stand-level timber volume and Lorey’s height were calculated and aggregated to trial-level 206 averages. After applying a remote sensing-based forest mask and excluding cloud-covered stands from 207 2021, 44 validation stands covering 62 ha were available for uncertainty assessment. 208 Table 1 Summary statistics of training NFI, testing NFI, and validation stand datasets. The table includes the number of 209 observations (N), minimum (Min), maximum (Max), mean, and standard deviation (SD) for each variable (Volume, Lorey’s 210 height, area of the plot, and the stands). 211 Dataset Variable Unit N Min Max Mean SD Training NFI plots Volume m3ha-1 1025 0.18 1157.59 118.13 127.99 Lorey's height m 2.40 28.18 11.96 5.11 Area m2 250 Testing NFI plots Volume m3ha-1 541 0.30 779.70 125.91 118.66 Lorey's height m 3.40 28.00 12.38 4.66 Area m2 250 Validation stands Volume m3ha-1 44 100.62 906.70 369.14 206.17 Lorey's height m 12.24 26.55 18.58 4.11 Area ha 0.11 4.32 1.40 1.06 212 Third, the Norwegian Forest Resource Map (SR16) was used as a wall-to-wall training dataset for the 213 UNet model. These spatially continuous maps provide estimates of timber volume, Lorey’s height, and 214 other forest attributes at a 16 m × 16 m resolution and are operationally updated each year following 215 the workflow described by Hauglin et al. (2021). They are produced by combining NFI plot data with 216 predictor variables derived from nationwide ALS acquisitions conducted by the Norwegian Mapping 217 Agency. Species-specific univariate linear mixed-effects regression models are used to account for 218 systematic differences across ALS projects and regions. To ensure consistency, NFI plot values are 219 projected to a common reference date, and a remote sensing-based forest mask is applied. 220 2.2. Modelling 221 2.2.1. kNN modeling 222 The kNN method has been chosen as the baseline method to be compared with the DL method, since 223 it has been successfully applied in estimating forest variables and it is operationally used, for example, 224 in Finland (McRoberts and Tomppo, 2007). It is a non-parametric algorithm, where the predictions of 225 the target variables are obtained from as the averages of the k neighbours with the smallest distances 226 to the target unit in the auxiliary space. The kNN model can be described as follows: 227 𝑦𝑝=∑𝑤𝑖,𝑝 𝑖 𝜖 𝐼 𝑦𝑖 228 where the predicted vector for each pixel is 𝑦𝑝 , 𝑦𝑖 is the vector of observations for the 𝑖-th contributing 229 unit in the reference set, 𝐼 is the subset of contributing units that are nearest with respect to the 230 distance metric, and 𝑤𝑖,𝑝 is the weight of i-th contributing unit calculated as 231 𝑤𝑖,𝑝= {𝑑𝑖,𝑝 −𝑡 ∑𝑑𝑖,𝑝 −𝑡 𝑖 𝜖 𝐼 0 𝑖 𝜖 𝐼 232 where 𝑡>0 weight units inversely to their feature space distance or distance squared from pixel 𝑝. - 233 In this study, 𝑡 = 1 was used, with Euclidean distance and seven nearest neighbours (k=7). 234 The kNN training database was constructed by intersecting training NFI plot locations (Fig.2, Table 1) 235 with the pre-processed EO datasets, including Sentinel-2, Sentinel-1, and PALSAR-2 bands. For each 236 plot, all pixels overlapping the 250 m² plot area were considered, and their contributions were 237 weighted by the fraction of pixel area falling within the plot. A weighted mean of the intersecting pixel 238 values was then computed, resulting in a training table that combined field-measured attributes with 239 corresponding EO features. Predictions were generated for each test NFI plot and validation stand 240 locations. 241 2.2.2. UNet model 242 A variant of a UNet model (Ronneberger et al., 2015) has been adapted for this study, which is a 243 convolutional neural network (CNN) widely used in biomedical image and other semantic segmentation 244 tasks. Compared to a basic CNN, UNet models allow extraction of deeper feature representations from 245 the input image datasets and keep the feature map size unchanged, which makes it suitable for pixel246 level regression tasks. 247 The UNet architecture applied (Fig.3.) in this study is a symmetric encoder–decoder network designed 248 for pixel-level predictions. The encoder contains a double-convolution structure, which includes a 2D 249 convolution, batch normalization, and ReLU activation. At each step, 2×2 max-pooling halves the 250 feature map. This allows the model to gradually transition from capturing fine, local details to learning 251 more global spatial patterns. The decoder mirrors this process by restoring the original resolution 252 through upsampling operations. To recover information lost during downsampling, skip connections 253 link corresponding encoder and decoder layers, as the pooling operation discarded some details. 254 Applying skip-connection, the shallow feature maps are concatenated to deep features recovered from 255 up-sampling. Finally, a 1×1 convolution projects the restored feature map into the target space, 256 producing the desired pixel-wise predictions without altering the spatial dimensions. 257 258 259 Figure 3 The UNet model architecture used in the study. Each box represents a feature map with its size indicated nearby. 260 2.2.3. UNet model training 261 First, all input datasets were prepared for training the UNet model (Fig. 4). During pre-training, the 262 Finnish and Norwegian wall-to-wall ALS-based forest resource maps, described in Sections 2.1.2–2.1.3, 263 were used as reference datasets. The pre-training dataset comprised 1,068 rectangular patches for 264 Finland and 1,025 patches for Norway, each with a size of 2.56 × 2.56 km² (256 × 256 pixels). These 265 patches were randomly distributed across the forested areas covered by the ALS-based forest resource 266 maps. The resolution of the forest resource maps was aligned with the satellite imagery datasets by 267 resampling the native spatial resolution of 16 × 16 m to 10 × 10 m. For fine-tuning, 1,025 Norwegian 268 NFI plots selected for training were used. When preparing image patches related to the training NFI 269 plots, only the 10 × 10 m pixels intersecting the field plot locations contained valid values; all other 270 pixels were masked out (set to no-data). 271 Predictor variables were extracted for the selected image patches (for both pre-training and fine272 tuning) based on the pre-processed and filtered EO bands from Sentinel-2 only (S2), and from the 273 multisource Sentinel-2, Sentinel-1, and PALSAR-2 datasets (S2S1P2), as described in Section 2.1.1. For 274 independent validation, EO image patches related to the NFI test set—which included 541 field plots 275 set aside at the beginning of the modelling process—and validation stands were prepared using the 276 same patch size, containing the relevant plots and stands. Within all image patches, non-forest pixels 277 were masked out. All EO patches were further normalised on a per-band basis using z-score 278 normalization, where each band’s mean was subtracted and the result divided by its standard 279 deviation. 280 281 Figure 4 Overview of the UNet model training workflow. Input patches are prepared for the Finnish and Norwegian datasets. 282 An independent test NFI plots were set aside before model training and used only for final uncertainty assessment. During 283 UNet pre-training and fine-tuning, the forest resource maps and training NFI plots used were randomly split 10 times into 284 training, validation and test subsets. The internal validation and test are used solely to run the DL model and are not used for 285 uncertainty assessment in this study. 286 During UNet model pre-training, the overall datasets were randomly split into training (70%), validation 287 (15%), and testing (15%) patches. The Adam optimiser was used with an initial learning rate of 1×10⁻⁴. 288 A ReduceLROnPlateau scheduler was applied to adaptively reduce the learning rate by half when the 289 validation loss plateaued, with a patience of ten epochs. The model was trained for 200 epochs with a 290 mild weight decay of 1×10⁻⁵. The best-performing model was selected based on the lowest validation 291 loss, determined using the internal validation data, and the corresponding weights were saved for 292 subsequent fine-tuning for each target variable. Training was conducted with a batch size of 16. The 293 final model was used to predict EO patches corresponding to the test NFI plot locations and validation 294 stands. 295 results also showed that in most cases bias was reduced, whereas models without fine-tunning often 439 exhibited higher bias, which aligns with Ge et al., (2023a). 440 Our results indicate that combining optical and SAR predictors generally improved model performance 441 compared to optical-only inputs (Tables 2–3). Similar trends have been reported in other boreal forest 442 studies. For example, Persson et al. (2021) demonstrated that including SAR data reduced the RMSE 443 for volume estimates in Sweden from 37.2% to 30.2%. Likewise, Ge et al. (2022, 2023a) reported an 444 improvement in Lorey’s height prediction in Finland, with RMSE decreasing from 30.3% to 18.1%. In 445 our study, the magnitude of improvement was smaller: the maximum reduction in RMSE was from 446 61.3% to 57.1% for volume and from 26.4% to 25.3% for Lorey’s height. Overall, we also found that the 447 addition of SAR, can improve capturing forest structure, which aligns with several studies (Wittke et al., 448 2019; Persson et al., 2021; Ge et al., 2022, 2023a). Two main factors can explain the minor 449 improvement in RMSE. First, the more complex topography in Norway compared to Sweden or Finland 450 can influence the SAR backscatter signal quality. Second, Persson et al. (2021) utilized TanDEM-X data, 451 which offers a finer spatial resolution (10 m) and a more frequent acquisition cycle (11-day repeat 452 cycle), and, most importantly, interferometric coherence that is inherently sensitive to vertical 453 structure of forests (Olesk et al., 2016). In contrast, our study also relied on PALSAR-2 data additionally 454 to S1 with coarser resolution (25 m) and used a single annual mosaic. 455 5. Conclusion 456 Our study demonstrates the potential of using spatially continuous ALS-based forest resource maps as 457 training data for UNet-based deep learning models. In contrast to most previous boreal forest mapping 458 approaches—which most commonly rely on end-to-end training using field plots of specified sizes and 459 satellite imagery within a single country—we show that ALS-derived maps can serve as an effective 460 training dataset that also leverages spatial context and can be applied across different Nordic countries. 461 We found that deep learning models outperform kNN method by reducing overestimation in low462 volume (<200 m³ ha⁻¹) forests and underestimation in high-volume (>200 m³ ha⁻¹) forests. A key 463 advantage of deep learning is its transferability with a limited set of local field plots. Our results indicate 464 that a model pre-trained on Finnish data performed comparably to kNN trained with Norwegian data. 465 With the development of more comprehensive foundation models trained on multi-country datasets, 466 this approach may enable robust, large-scale, cross-country forest resource mapping. 467 Acknowledgments 468 This research was funded by the European Space Agency (ESA) project on Forest Carbon Monitoring 469 under contract number 4000135015/21/I-NB—Forest Carbon Monitoring, and by the EU under 470 GA101056907 (PathFinder). The satellite data processing was supported by the ESA Network of 471 Resources Initiative. Microsoft Copilot assisted in enhancing the clarity and correctness of the English 472 text. 473 Author contributions 474 ZK: Conceptualization, Data curation, Formal analysis, Validation, Writing original draft, OA: 475 Conceptualization, Data curation, Formal analysis, Methodology, Writing review and editing, OC: Data 476 curation, Writing review and editing, JM: Formal analysis, Project administration, Funding acquisition, 477 Writing review and editing, JB: Project administration, Funding acquisition, Writing review and editing 478 Data and code statement 479 The pre-processed EO dataset is available upon request from the corresponding author. The used 480 methods are available within the Forestry-TEP (https://f-tep.com/) developed by VTT. 481 References 482 Antropov O, Rauste Y, Häme T, Praks J. 2017. Polarimetric ALOS PALSAR Time Series in Mapping 483 Biomass of Boreal Forests. Remote Sensing 9 : 999. DOI: 10.3390/rs9100999 484 Astola H, Häme T, Sirro L, Molinier M, Kilpi J. 2019. Comparison of Sentinel-2 and Landsat 8 imagery 485 for forest variable prediction in boreal region. Remote Sensing of Environment 223 : 257–273. DOI: 486 10.1016/j.rse.2019.01.019 487 Astola H, Seitsonen L, Halme E, Molinier M, Lönnqvist A. 2021. Deep Neural Networks with Transfer 488 Learning for Forest Variable Estimation Using Sentinel-2 Imagery in Boreal Forest. Remote Sensing 13 489 : 2392. DOI: 10.3390/rs13122392 490 Astrup R, Rahlf J, Bjørkelo K, Debella-Gilo M, Gjertsen A-K, Breidenbach J. 2019. Forest information at 491 multiple scales: development, evaluation and application of the Norwegian forest resources map 492 SR16. Scandinavian Journal of Forest Research 34 : 484–496. DOI: 10.1080/02827581.2019.1588989 493 Ball JE, Anderson DT, Sr CSC. 2017. Comprehensive survey of deep learning in remote sensing: 494 theories, tools, and challenges for the community. Journal of Applied Remote Sensing 11 : 042609. 495 DOI: 10.1117/1.JRS.11.042609 496 Becker A, Russo S, Puliti S, Lang N, Schindler K, Wegner JD. 2023. Country-wide retrieval of forest 497 structure from optical and SAR satellite imagery with deep ensembles. ISPRS Journal of 498 Photogrammetry and Remote Sensing 195 : 269–286. DOI: 10.1016/j.isprsjprs.2022.11.011 499 Breidenbach J, Astrup R. 2012. Small area estimation of forest attributes in the Norwegian National 500 Forest Inventory. European Journal of Forest Research 131 : 1255–1267. DOI: 10.1007/s10342-012501 0596-7 502 Breidenbach J, Granhus A, Hylen G, Eriksen R, Astrup R. 2020. A century of National Forest Inventory 503 in Norway – informing past, present, and future decisions. Forest Ecosystems 7 : 46. DOI: 504 10.1186/s40663-020-00261-0 505 Esteban J, McRoberts RE, Fernández-Landa A, Tomé JL, Nӕsset E. 2019. Estimating Forest Volume and 506 Biomass and Their Changes Using Random Forests and Remotely Sensed Data. Remote Sensing 11 : 507 1944. DOI: 10.3390/rs11161944 508 Ge S, Antropov O, Häme T, McRoberts RE, Miettinen J. 2023a. Deep Learning Model Transfer in Forest 509 Mapping Using Multi-Source Satellite SAR and Optical Images. Remote Sensing 15 : 5152. DOI: 510 10.3390/rs15215152 511 Ge S, Gu H, Su W, Praks J, Antropov O. 2022. Improved Semisupervised UNet Deep Learning Model for 512 Forest Height Mapping With Satellite SAR and Optical Data. IEEE Journal of Selected Topics in Applied 513 Earth Observations and Remote Sensing 15 : 5776–5787. DOI: 10.1109/JSTARS.2022.3188201 514 Ge S, Tomppo E, Rauste Y, McRoberts RE, Praks J, Gu H, Su W, Antropov O. 2023b. Sentinel-1 Time 515 Series for Predicting Growing Stock Volume of Boreal Forest: Multitemporal Analysis and Feature 516 Selection. Remote Sensing 15 : 3489. DOI: 10.3390/rs15143489 517 Gjertsen AK. 2007. Accuracy of forest mapping based on Landsat TM data and a kNN-based method. 518 Remote Sensing of Environment 110 : 420–430. DOI: 10.1016/j.rse.2006.08.018 519 Hauglin M, Rahlf J, Schumacher J, Astrup R, Breidenbach J. 2021. Large scale mapping of forest 520 attributes using heterogeneous sets of airborne laser scanning and National Forest Inventory data. 521 Forest Ecosystems 8 : 65. DOI: 10.1186/s40663-021-00338-4 522 Kangas A et al. 2018. Remote sensing and forest inventories in Nordic countries – roadmap for the 523 future. Scandinavian Journal of Forest Research 33 : 397–412. DOI: 10.1080/02827581.2017.1416666 524 Kattenborn T, Leitloff J, Schiefer F, Hinz S. 2021. Review on Convolutional Neural Networks (CNN) in 525 vegetation remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing 173 : 24–49. DOI: 526 10.1016/j.isprsjprs.2020.12.010 527 Koma Z, Breidenbach J. 2025. Large-scale validation of forest attribute maps across different spatial 528 resolutions. Silva Fennica 59 [online] Available from: https://www.silvafennica.fi/article/24061 529 (Accessed 16 September 2025) 530 Kuzu RS, Antropov O, Molinier M, Dumitru CO, Saha S, Zhu XX. 2024. Forest Disturbance Detection via 531 Self-Supervised and Transfer Learning With Sentinel-1&2 Images. IEEE Journal of Selected Topics in 532 Applied Earth Observations and Remote Sensing 17 : 4751–4767. DOI: 10.1109/JSTARS.2024.3361183 533 Lang N, Schindler K, Wegner JD. 2019. Country-wide high-resolution vegetation height mapping with 534 Sentinel-2. Remote Sensing of Environment 233 : 111347. DOI: 10.1016/j.rse.2019.111347 535 Li J, Huang X, Gong J. 2019. Deep neural network for remote-sensing image interpretation: status and 536 perspectives. National Science Review 6 : 1082–1086. DOI: 10.1093/nsr/nwz058 537 Maltamo M, Packalen P. 2014. Species-Specific Management Inventory in Finland. In Forestry 538 Applications of Airborne Laser Scanning: Concepts and Case Studies , Maltamo M, Næsset E, and 539 Vauhkonen J (eds). Springer Netherlands: Dordrecht; 241–252. [online] Available from: 540 https://doi.org/10.1007/978-94-017-8663-8_12 541 McRoberts RE, Tomppo EO. 2007. Remote sensing support for national forest inventories. Remote 542 Sensing of Environment 110 : 412–419. DOI: 10.1016/j.rse.2006.09.034 543 Miettinen J et al. 2021. Demonstration of large area forest volume and primary production estimation 544 approach based on Sentinel-2 imagery and process based ecosystem modelling. International Journal 545 of Remote Sensing 42 : 9467–9489. DOI: 10.1080/01431161.2021.1998715 546 Nilsson M et al. 2017. A nationwide forest attribute map of Sweden predicted using airborne laser 547 scanning data and field data from the National Forest Inventory. Remote Sensing of Environment 194 548 : 447–454. DOI: 10.1016/j.rse.2016.10.022 549 NLS. 2025. National Land Survey (NLS) Finland. 2025. Kansalliset ilmakuvausja laserkeilausohjelmat 550 saavat jatkoa – luvassa entistä tarkempaa kolmiulotteista tietoa Suomesta. Available at: 551 https://www.maanmittauslaitos.fi/ajankohtaista/kansalliset-ilmakuvaus-ja-laserkeilausohjelmat552 saavat-jatkoa-luvassa-entista (accessed 19 Sep 2025). [In Finnish and Swedish] 553 Olesk A, Praks J, Antropov O, Zalite K, Arumäe T, Voormansik K. 2016. Interferometric SAR Coherence 554 Models for Characterization of Hemiboreal Forests Using TanDEM-X Data. Remote Sensing 8 : 700. 555 DOI: 10.3390/rs8090700 556 Persson HJ, Jonzén J, Nilsson M. 2021. Combining TanDEM-X and Sentinel-2 for large-area species557 wise prediction of forest biomass and volume. International Journal of Applied Earth Observation and 558 Geoinformation 96 : 102275. DOI: 10.1016/j.jag.2020.102275 559 Ronneberger O, Fischer P, Brox T. 2015. U-Net: Convolutional Networks for Biomedical Image 560 Segmentation. Cham. 234–241 pp. 561 Santoro M, Beer C, Cartus O, Schmullius C, Shvidenko A, McCallum I, Wegmüller U, Wiesmann A. 562 2011. Retrieval of growing stock volume in boreal forest using hyper-temporal series of Envisat ASAR 563 ScanSAR backscatter measurements. Remote Sensing of Environment 115 : 490–507. DOI: 564 10.1016/j.rse.2010.09.018 565 Shimada M, Ohtaki T. 2010. Generating Large-Scale High-Quality SAR Mosaic Datasets: Application to 566 PALSAR Data for Global Monitoring. IEEE Journal of Selected Topics in Applied Earth Observations and 567 Remote Sensing 3 : 637–656. DOI: 10.1109/JSTARS.2010.2077619 568 Tomppo E, Gschwantner T, Lawrence M, McRoberts RE (eds). 2010. National Forest Inventories: 569 Pathways for Common Reporting . Springer Netherlands: Dordrecht [online] Available from: 570 https://link.springer.com/10.1007/978-90-481-3233-1 (Accessed 25 January 2024) 571 Vidal C, Alberdi I, Redmond J, Vestman M, Lanz A, Schadauer K. 2016a. The role of European National 572 Forest Inventories for international forestry reporting. Annals of Forest Science 73 : 793–806. DOI: 573 10.1007/s13595-016-0545-6 574 Vidal C, Alberdi IA, Hernández Mateo L, Redmond JJ (eds). 2016b. National Forest Inventories . 575 Springer International Publishing: Cham [online] Available from: 576 http://link.springer.com/10.1007/978-3-319-44015-6 (Accessed 13 August 2025) 577 Werner C, Wegmüller U, Strozzi T, Wiesmann A. 2000. GAMMA SAR AND INTERFEROMETRIC 578 PROCESSING SOFTWARE 579 Wittke S, Yu X, Karjalainen M, Hyyppä J, Puttonen E. 2019. Comparison of two-dimensional 580 multitemporal Sentinel-2 data with three-dimensional remote sensing data sources for forest 581 inventory parameter estimation over a boreal forest. International Journal of Applied Earth 582 Observation and Geoinformation 76 : 167–178. DOI: 10.1016/j.jag.2018.11.009 583 Wurm M, Stark T, Zhu XX, Weigand M, Taubenböck H. 2019. Semantic segmentation of slums in 584 satellite images using transfer learning on fully convolutional neural networks. ISPRS Journal of 585 Photogrammetry and Remote Sensing 150 : 59–69. DOI: 10.1016/j.isprsjprs.2019.02.006 586 Yuan Q et al. 2020. Deep learning in environmental remote sensing: Achievements and challenges. 587 Remote Sensing of Environment 241 : 111716. DOI: 10.1016/j.rse.2020.111716 588 Zhou J et al. 2023. A deep transfer learning framework for mapping high spatiotemporal resolution 589 LAI. ISPRS Journal of Photogrammetry and Remote Sensing 206 : 30–48. DOI: 590 10.1016/j.isprsjprs.2023.10.017 591 592 593 594 595 Appendix 596 Table S1 Overall uncertainty assessment for timber volume comparing kNN, UNet, and fine-tuned UNet models trained on 597 Finnish (FI) and Norwegian (NO) data. Results are shown for NFI plots, validation stands, and two types of predictor variable 598 combinations (Pred. type): S2 (Sentinel-2) and S2S1P2 (Sentinel-2 with Sentinel-1 and PALSAR ‑ 2). Fine-tuned models were 599 pre-trained on ALS-based forest resource maps from Finland (FI) or Norway (NO) and fine-tuned further using Norwegian 600 training NFI plots. 601 Test data type Pred. type Model type kNN UNet Finetuned UNet kNN UNet FineTuned UNet kNN UNet FineTuned UNet RMSE [m3ha-1] BIAS [m3ha-1] R2 NFI S2 FI 93.8 91.4 -20.4 -9.9 0.37 0.41 NFI S2 NO 94.0 77.2 77.2 9.9 3.1 2.7 0.37 0.58 0.58 NFI S2S1P2 FI 90.4 90.3 -5.1 -3.1 0.42 0.42 NFI S2S1P2 NO 92.8 72.6 72.0 7.0 10.2 3.4 0.39 0.62 0.63 Stand S2 FI 190.1 176.7 -116.6 -98.5 0.04 0.17 Stand S2 NO 148.6 128.1 128.5 -67.6 -45 -46.0 0.42 0.57 0.56 Stand S2S1P2 FI 170.6 168.8 -74.2 -70.8 0.23 0.25 Stand S2S1P2 NO 156.8 119.8 123.5 -67 -4.6 -23.3 0.35 0.62 0.60 602 Table S2 Overall uncertainty assessment for Lorey’s height comparing kNN, UNet, and fine-tuned UNet models trained on 603 Finnish (FI) and Norwegian (NO) data. Results are shown for NFI plots, validation stands, and two types of predictor variable 604 combinations (Pred. type): S2 (Sentinel-2) and S2S1P2 (Sentinel-2 with Sentinel-1 and PALSAR ‑ 2). Fine-tuned models were 605 pre-trained on ALS-based forest resource maps from Finland (FI) or Norway (NO) and fine-tuned further using Norwegian 606 training NFI plots. 607 Test data type Pred. type Model type kNN UNet FineTuned UNet kNN UNet FineTuned UNet kNN UNet FineTuned UNet RMSE [m] BIAS [m] R2 NFI S2 FI 37.3 36.6 3.0 -2.0 0.36 0.38 NFI S2 NO 36.9 33.0 32.6 3.4 5.2 1.7 0.37 0.50 0.51 NFI S2S1P2 FI 45.9 36.7 7.5 4.3 0.03 0.38 NFI S2S1P2 NO 37.0 31.9 31.3 3.2 6.2 1.8 0.37 0.53 0.55 Stand S2 FI 28.9 29.7 1.4 -5.6 0.48 0.46 Stand S2 NO 36.8 31.2 32.3 -15.1 -4.8 -9.2 0.17 0.40 0.36 Stand S2S1P2 FI 43.6 33.8 27.4 5.7 -0.17 0.29 Stand S2S1P2 NO 36.5 30.0 30.5 -13.7 0.8 -4.6 0.18 0.44 0.43 608 609 610