Copernicus Sentinel-1 images are widely used for forest mapping and predicting forest growing stock volume (GSV) due to their accessibility. However, certain important aspects related to the use of Sentinel-1 time series have not been thoroughly explored in the literature. These include the impact of image time series length on prediction accuracy, the optimal feature selection approaches, and the best prediction methods. In this study, we conduct an in-depth exploration of the potential of long time series of Sentinel-1 SAR data to predict forest GSV and evaluate the temporal dynamics of the predictions using extensive reference data. Our boreal coniferous forests study site is located near the Hyytiälä forest station in central Finland and covers an area of 2500 km2 with nearly 17,000 stands. We considered several prediction approaches and fine-tuned them to predict GSV in various evaluation scenarios. Our analyses used 96 Sentinel-1 images acquired over three years. Different approaches for aggregating SAR images and choosing feature (predictor) variables were evaluated. Our results demonstrate a considerable decrease in the root mean squared errors (RMSEs) of GSV predictions as the number of images increases. While prediction accuracy using individual Sentinel-1 images varied from 85 to 91 m3/ha RMSE, prediction accuracy with combined images decreased to 75.6 m3/ha. Feature extraction and dimension reduction techniques facilitated the achievement of near-optimal prediction accuracy using only 8–10 images. Examined methods included radiometric contrast, mutual information, improved k-Nearest Neighbors, random forests selection, Lasso, and Wrapper approaches. Lasso was the most optimal, with RMSE reaching 77.1 m3/ha. Finally, we found that using assemblages of eight consecutive images resulted in the greatest accuracy in predicting GSV when initial acquisitions started between September and January.

Our objective was to develop a method for the assessment of forest area and structural variables for cases in which the availability of representative ground reference data is poor and these data are not collected from the whole area of interest. We implemented two independent approaches to the estimation of the forest variables of a European boreal forest: (i) the computation of wall-to-wall estimates using moderate- to low-resolution VIIRS imagery from the Suomi NPP mission; and (ii) the visual interpretation of plots of samples from very high resolution (VHR) satellite data obtained via a two-stage design. Our focus was on the statistical comparison of forest resources at a country or larger level. The study area was boreal forest ranging from Norway to the Ural Mountains in Russia. We computed a seamless mosaic from 111 VIIRS images. From the mosaic, we computed predictions for the forest area, growing stock volume, height of the dominating tree layer, proportion of conifers and broadleaved trees, site fertility class, and leaf area index. The reference data for the VIIRS imagery were national forest inventory (NFI)-based raster maps from Finland. The first stage sample of VHR data included 42 images; of these, a second stage sample of 2690 plots was visually interpreted for the same variables. The forest area prediction from VIIRS for the whole study area was 1.2% higher than the VHR-based result. All other structural variable predictions using VIIRS fitted within the 95% confidence intervals computed from the VHR sample except for estimates of the main tree species groups, which were outside the limits. A comparison of VIIRS-based forest area estimates using Finnish and Swedish NFI data indicated overestimations of 10.0% points and 4.6% points, whereas the total growing stock volumes were overestimated by 8% and underestimated by 3.4%, respectively. The correlation coefficients between the VIIRS and VHR image predictions at the 42 VHR image locations varied from 0.70 to 0.85. The VIIRS maps strongly averaged the local predictions due to their coarse spatial resolutions. Based on our findings, the approach using two independent estimations yielded similar figures for the central forest variables for the European boreal forest. A model computed using reference data from a small part of the area of interest can provide satisfactory predictions for a much larger area with a similar biome. Therefore, our concept is applicable to the estimation and overall mapping of a forest area and central structural variables at regional to national levels.

ESA Forest Carbon Monitoring project (FCM) is developing Earth Observation based, user-centric approaches for forest carbon monitoring. Forest carbon accounting based on forest inventory requires precise and timely estimation of forest variables at various spatial levels accompanied by verifiable uncertainty information. In this paper, we present the algorithm trade-off and selection approach and preliminary results of the algorithm intercomparison exercise in the FCM project. The studies were performed over 7 European test sites located in Finland, Ireland, Romania, Spain and Switzerland, and one tropical forest site in Peru. EO datasets were represented by Sentinel-1, Sentinel-2, TanDEM-X and ALOS-2 PALSAR-2 imagery. Examined approaches include popular parametric and SAR/InSAR scattering physics based approaches, and nonparametric and machine learning approaches such as k-NN, random forests, support vector regression.

Time series of SAR imagery combined with reference ground data can be suitable for producing forest inventories. Copernicus Sentinel-1 imagery is particularly interesting for forest mapping because of its free availability to data users; however, temporal dependencies within SAR time series that can potentially improve mapping accuracy are rarely explored. In this study, we introduce a novel semi-supervised Long Short-Term Memory (LSTM) model, CrsHelix-LSTM, and demonstrate its utility for predicting forest tree height using time series of Sentinel-1 images. The model brings three important modifications to the conventional LSTM model. Firstly, it uses a Helix-Elapse (HE) projection to capture the relationship between forest temporal patterns and Sentinel-1 time series, when time intervals between datatakes are irregular. A skip-link based LSTM block is introduced and a novel backbone network, Helix-LSTM, is proposed to retrieve temporal features at different receptive scales. Finally, a novel semisupervised strategy, Cross-Pseudo Regression, is employed to achieve better model performance when reference training data are limited. CrsHelix-LSTM model is demonstrated over a representative boreal forest site located in Central Finland. A time series of 96 Sentinel-1 images are used in the study. The developed model is compared with basic LSTM model, attention-based bidirectional LSTM and several other established regression approaches used in forest variable mapping, demonstrating consistent improvement of forest height prediction accuracy. At best, the achieved accuracy of forest height mapping was 28.3% relative root mean squared error (rRMSE) for pixel-level predictions and 18.0% rRMSE on stand level. We expect that the developed model can also be used for modeling relationships between other forest variables and satellite image time series.

In this study, we evaluate the performance of several imaging modes of X-band satellite SAR data in mapping various forest logging operations. We use stripmap and spotlight data acquired by DLR's TanDEM-X mission and ICEYE SAR satellites. Also, the potential of InSAR coherence data is evaluated. Experiments are performed over the study site located in the central part of Finland. Reference data are represented by forest management plans and harvester logging tracking reports. Our results indicate that not only clear-cutting, but also forest thinning operations can be detected using advanced imaging modes of satellite SAR data even at X-band, primarily using high resolution imaging modes and interferometric SAR capability.

Land cover (LC) mapping is essential for monitoring the environment and understanding the effects of human activities on it. Recent studies demonstrated successful applications of specific deep learning models to small-scale LC mapping tasks (e.g., wetland mapping). However, it is not readily clear which of the existing state-of-the-art models for natural images are the best candidates to be taken for the particular remote sensing task and data. In this article, we answer that question for mapping the fundamental LC classes using the satellite imaging radar data. We took ESA Sentinel-1 C-band SAR images acquired during the whole summer season of 2018 in Finland, which are representative of the land cover in the country. CORINE LC map was used as a reference, and the models were trained to distinguish between the five major CORINE-based classes. We selected seven among the state-of-the-art semantic segmentation models so that they cover a diverse set of approaches: U-Net, DeepLabV3+, PSPNet, BiSeNet, SegNet, FC-DenseNet, and FRRN-B, and further fine-tuned them. Upon evaluation and benchmarking, all the models demonstrated solid performance with overall accuracy between 87.9% and 93.1%, with good to a very good agreement (Kappa statistic between 0.75 and 0.86). The two best models were fully convolutional DenseNets (FC-DenseNet) and SegNet (encoder-decoder-skip), with the latter having a much shorter inference time. Overall, our results indicate that the semantic segmentation models are suitable for efficient wide-area mapping using satellite SAR imagery and provide baseline accuracy against which the newly proposed models should be evaluated.

In this study, we assess the potential of long time series of Sentinel-1 SAR data to predict forest growing stock volume and evaluate the temporal dynamics of the predictions. The boreal coniferous forests study site is located near the Hyytiälä forest station in central Finland and covers an area of 2,500 km2 with nearly 17,000 stands. We considered several prediction approaches (linear, support vector and random forests regression) and fine-tuned them to predict growing stock volume in several evaluation scenarios. The analyses used 96 Sentinel-1 images acquired over three years. Different approaches for aggregating SAR images and choosing feature (predictor) variables were evaluated. Our results demonstrate considerable decrease in RMSEs of growing stock volume as the number of images increases. While prediction accuracy using individual Sentinel-1 images varied from 85 to 91 m3/ha RMSE (relative RMSE 50-53%), RMSE with combined images decreased to 75.6 m3/ha (relative RMSE 44%). Feature extraction and dimension reduction techniques facilitated achieving the near-optimal prediction accuracy using only 8-10 images. When using assemblages of eight consecutive images, the GSV was predicted with the greatest accuracy when initial acquisitions started between September and January. Highlights ### Competing Interest Statement The authors have declared no competing interest.

Forest biomass and carbon monitoring play a key role in climate change mitigation. Operational large area monitoring approaches are needed to enable forestry stakeholders to meet the increasing monitoring and reporting requirements. Here, we demonstrate the functionality of a cloud-based approach utilizing Sentinel-2 composite imagery and process-based ecosystem model to produce large area forest volume and primary production estimates. We describe the main components of the approach and implementation of the processing pipeline into the Forestry TEP cloud processing platform and produce four large area output maps: (1) Growing stock volume (GSV), (2) Gross primary productivity (GPP), (3) Net primary productivity (NPP) and (4) Stem volume increment (SVI), covering Finland and the Russian boreal forests until the Ural Mountains in 10 m spatial resolution. The accuracy of the forest structural variables evaluated in Finland reach pixel level relative Root Mean Square Error (RMSE) values comparable to earlier studies (basal area 39.4%, growing stock volume 58.5%, diameter 35.5% and height 33.5%), although most of the earlier studies have concentrated on smaller study areas. This can be considered a positive sign for the feasibility of the approach for large area primary production modelling, since forest structural variables are the main input for the process-based ecosystem model used in the study. The full coverage output maps show consistent quality throughout the target area, with major regional variations clearly visible, and with noticeable fine details when zoomed into full resolution. The demonstration conducted in this study lays foundation for further development of an operational large area forest monitoring system that allows annual reporting of forest biomass and carbon balance from forest stand level to regional analyses. The system is seamlessly aligned with process based ecosystem modelling, enabling forecasting and future scenario simulation.

Dense time series of stripmap RADARSAT-2 data acquired in the Multilook Fine mode were used for detecting and mapping the extent of selective logging operations in the tropical forest area in the northern part of the Republic of the Congo. Due to limited radiometric sensitivity to forest biomass variation at C-band, basic multitemporal change detection approach was supplemented by spatial texture analysis to separate disturbed forest from intact. The developed technique primarily uses multi-temporal aggregation of orthorectified synthetic aperture radar (SAR) imagery that are acquired before and after the logging operations. The actual change analysis is based on textural features of the log-ratio image calculated using two SAR temporal composites compiled of SAR scenes acquired before and after the logging operations. Multitemporal aggregation and filtering of SAR scenes decreased speckle and made the extracted textural features more prominent. The overall detection accuracy was around 80%, with some underestimation of the area of forest disturbance compared to reference based on optical data. The user's accuracy for disturbed forest varied from 76.7% to 94.9% depending on the accuracy assessment approach. We conclude that change detection utilizing RADARSAT-2 time series represents a useful instrument to locate areas of selective logging in tropical forests.

In this study, we assess the potential of long time series of Sentinel-1 SAR data in forest growing stock volume (GSV) estimation. The study site with 17,762 forest stands is located near the Hyytiälä forestry field station in Finland and represents the boreal coniferous forest. Altogether 96 images spanning more than three years of observations have been studied using linear and random forest regression approaches. Our analysis demonstrates considerable decrease in the prediction errors of GSV as the the number of input scenes increases. The use of feature extraction and dimensionality reduction techniques allows to achieve to nearly optimal performance already with 10 scenes. While the GSV prediction errors using individual Sentinel-1 scenes varied considerably from 86 to 93 m 3 /ha, the prediction accuracy with combined scenes improved to 76 m 3 /ha (44.9%) RMSE.

Here, we examine a deep learning approach to perform land cover classification using country-wide SAR mosaics compiled using multitemporal Sentinel-1 imagery. We capitalize on our earlier study [1], demonstrating the suitability of deep learning models for land cover mapping using satellite C-band SAR images A set of SAR mosaics compiled from consecutive Sentinel-1 IW mode acquisitions covering the whole territory of Finland was used in production of the whole-country land cover map. The imagery were used as an input to the state-of-the-art deep-learning model for semantic segmentation called FC-DenseNet. This model was pre-trained on the ImageNet dataset and further fine-tuned in this study. CORINE land cover map was used as a reference, and the model was trained to distinguish between 5 Level-1 CORINE classes. Upon the evaluation and benchmarking, we found that the FC-DenseNet model is able to achieve nearly 90% overall classification accuracy. These results indicate the suitability of deep learning approaches to support efficient operational wide-area mapping using satellite SAR imagery.

The amount and spatial distribution of forest aboveground biomass (AGB) were estimated using a range of regionally developed methods using Earth Observation data for Poland, Sweden and regions in Indonesia (Kalimantan), Mexico (Central Mexico and Yucatan peninsula), and South Africa (Eastern provinces) for the year 2010. These regions are representative of numerous forest biomes and biomass levels globally, from South African woodlands and savannas to the humid tropical forest of Kalimantan. AGB retrieval in each region relied on different sources of reference data, including forest inventory plot data and airborne LiDAR observations, and used a range of retrieval algorithms. This is the widest inter-comparison of regional-to-national AGB maps to date in terms of area, forest types, input datasets, and retrieval methods. The accuracy assessment of all regional maps using independent field data or LiDAR AGB maps resulted in an overall root mean square error (RMSE) ranging from 10 t ha(-1) to 55 t ha(-1) (37% to 67% relative RMSE), and an overall bias ranging from -1 t ha(-1) to +5 t ha(-1) at pixel level. The regional maps showed better agreement with field data than previously developed and widely used pan-tropical or northern hemisphere datasets. The comparison of accuracy assessments showed commonalities in error structures despite the variety of methods, input data, and forest biomes. All regional retrievals resulted in overestimation (up to 63 t ha(-1)) in the lower AGB classes, and underestimation (up to 85 t ha(-1)) in the higher AGB classes. Parametric model-based algorithms present advantages due to their low demand on in situ data compared to non-parametric algorithms, but there is a need for datasets and retrieval methods that can overcome the biases at both ends of the AGB range. The outcomes of this study should be considered when developing algorithms to estimate forest biomass at continental to global scale level.

Sirro L., Häme L., Pitkänen T.P., Hostikka A., Törmä M., Häme T., Kangas A., Rauste Y., Huotari R., Norppa J. (2019). Metsävaratiedon ajantasaistus – satelliittikuviin perustuva muutostulkinta. Metsätieteen aikakauskirja 2019-10260. Tieteen tori: Metsätieto ja sähköiset palvelut. 6 s. https:// doi.org/10.14214/ma.10260 Yhteystiedot 1Teknologian tutkimuskeskus VTT Oy, Espoo; 2 Satellio Oy (Terramonitor), Helsinki; 3 Luonnonvarakeskus, Biotalous ja ympäristö, Helsinki; 4 Suomen metsäkeskus, Tarkastukset ja valvonta, Lahti; 5 Suomen ympäristökeskus, Geoinformatiikkatutkimus, Helsinki; 6 Luonnonvarakeskus, Biotalous ja ympäristö, Joensuu Sähköposti laura.sirro@vtt.fi Hyväksytty 30.10.2019

Land use and land cover maps are vital sources of information for many applications. Recently, using high-resolution and open-access satellite images has become a preferred method for mapping land cover, especially over large areas. This study was designed to map the land cover and agricultural fields of a large area using Sentinel-1A synthetic aperture radar (SAR) images. Seven machine-learning methods were employed for image analyses. The Random Forest classifier algorithm outperformed the other machine-learning methods in the training step; thus, we selected it for further use and tuned its parameters. After several image processing steps, we classified the final image into 23 land cover classes and achieved an overall accuracy of 42% for all classes, and 57% for agricultural fields. This research note highlights some characteristics and advantages of using Sentinel-1A images and provides novel methods for nation-wide large-area mapping applications.

In this study, a set of bistatic interferometric SAR images acquired by the TanDEM-X mission are used along with fully polarimetric ALOS PALSAR data for the assessment of tropical forest properties in Nepal. Research to be presented at the conference concentrates on several scientific goals. First, location of interferometric scattering phase centre inside forest canopy is investigated using reference ALS measured canopy height model. Means for forest tree height extraction using both model based approaches (similar to Random Volume over Ground) and semi-empirical models are investigated and reported. It is shown, that forest tree height retrieval is possible with RMSE around 2.8 meters (R-2=0.68). Secondly, correlations between ALOS PALSAR and tropical forest data are analysed, and AGB estimation using fully polarimetric SAR features is performed. Several statistical and non-parametric methodologies were tested and compared. Thirdly, forest AGB estimation is done using both TanDEM-X based tree height and L-band PolSAR features using statistical inversion framework.

The applicability of optical and synthetic aperture radar (SAR) data for land cover classification to support REDD+ (Reducing Emissions from Deforestation and Forest Degradation) MRV (measuring, reporting and verification) services was tested on a tropical to sub-tropical test site. The 100 km by 100 km test site was situated in the State of Chiapas in Mexico. Land cover classifications were computed using RapidEye and Landsat TM optical satellite images and ALOS PALSAR L-band and Envisat ASAR C-band images. Identical sample plot data from Kompsat-2 imagery of one-metre spatial resolution were used for the accuracy assessment. The overall accuracy for forest and non-forest classification varied between 95% for the RapidEye classification and 74% for the Envisat ASAR classification. For more detailed land cover classification, the accuracies varied between 89% and 70%, respectively. A combination of Landsat TM and ALOS PALSAR data sets provided only 1% improvement in the overall accuracy. The biases were small in most classifications, varying from practically zero for the Landsat TM based classification to a 7% overestimation of forest area in the Envisat ASAR classification. Considering the pros and cons of the data types, we recommend optical data of 10 m spatial resolution as the primary data source for REDD MRV purposes. The results with L-band SAR data were nearly as accurate as the optical data but considering the present maturity of the imaging systems and image analysis methods, the L-band SAR is recommended as a secondary data source. The C-band SAR clearly has poorer potential than the L-band but it is applicable in stratification for a statistical sampling when other image types are unavailable.

Here, we examined multitemporal behavior of fully polarimetric SAR (PolSAR) parameters at L-band in relation to the stem volume of boreal forests. The PolSAR parameters were evaluated in terms of their temporal consistency, inter-dependence and suitability for forest stem volume estimation across several seasonal conditions (frozen, thaw and unfrozen). The satellite SAR data were represented by a time series of PolSAR images acquired during several seasons in the years 2006 to 2009 by the ALOS PALSAR sensor. The study area was in central Finland, and represented a managed area in typical boreal mixed forest land. Utility of different PolSAR parameters, their temporal stability and cross-correlations were studied along with reference stand-level stem volume data from forest inventory. Further, two polarimetric parameters, cross-polarization backscatter and co-polarization coherence, were chosen for further investigation and stem volume retrieval. A relationship between forest stem volume and PolSAR parameters was established using the kNN regression approach. Ways of optimally combining PolSAR images were evaluated as well. For a single scene, best results were observed with polarimetric coherence (RMSE ≈ 38.8 m3/ha) for scene acquired in frozen conditions. An RMSE of 40.8 m3/ha (42.9%, R2 = 0.66) was achieved for cross-polarization backscatter in the best case. Cross-polarization backscatter was a better predictor than polarimetric coherence for few summer scenes. Multitemporal aggregation of selected PolSAR scenes improved estimates for both studied PolSAR parameters. Stronger improvement was observed for coherence with RMSE down to 34 m3/ha (35.8%, R2 = 0.77) compared to 38.8–51.6 m3/ha (40.8–54.3%) from separate scenes. Finally, the accuracy statistics reached RMSE of 32.2 m3/ha (34%, R2 = 0.79) when multitemporal HHVV coherence was combined with multitemporal HV-backscatter.