This study aims at mapping wall-to-wall forest aboveground biomass (AGB) of Canada by directly upscaling the national forest inventory (NFI) plot measurements with machine learning method and satellite observations. We used the geolocated ground plots provided by NFI project from 10 provinces over the period 1992 to 2018. This dataset contained ground plots with measurements that were performed up to three times since 1992. We cleaned the data based on age and historical disturbance information to retain as many plots as possible for model training, while ensuring that the AGB in the used plots did not vary greatly or affected by disturbance from the date of measurement up to 2020. Finally, if there were repeat measurements in the remaining plots, we only kept the latest measurement records. The input features for estimation model were extracted from seasonal composited Sentinel 1 spectral images, Sentinel 2 L band SAR images and ALOS PALSAR yearly mosaic data. The Machine learning method - Random Forest Regression was used for AGB estimation. We trained the RF model locally and uploaded the model to the GEE platform to predict a wall-to-wall AGB map for Canada. To train and select the best performing model, we employed three categories of training and validation methods including random split (RS, repeated 100 times), simple 10-fold cross-validation (S10C, repeated 10 times) and stratified 10-fold cross-validation (ST10C, repeated 10 times). The prediction uncertainty of the model was determined by the Quantile Regression (QR at 5%,50% and 95%) equations between the mean bias and the mean prediction of 100 model. The bias of the model showed a characteristic V-shape pattern when compared to the predicted AGB values, which showed the range of bias value widened as the predicted AGB values increased. This distribution of bias can be described by the 5%, 50% and 95% QR line equation response to the lower, median and upper bounds of model prediction bias. With those equations, we can generate bias variation range for all predicted pixels.

How to cite: Qin, S., Bermúdez, J., A. Rogers, C., So, K., Gonsamo, A., and Wang, H.: Direct upscaling of national forest inventory aboveground biomass of Canada with Sentinel and ALOS PALSAR observations, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9996, https://doi.org/10.5194/egusphere-egu23-9996, 2023.

How to cite: Bermudez Castro, J. D., Qin, S., Sothe, C., and Gonsamo, A.: Convolutional Neural Networks Regression Model with Uncertainty Estimates to predict GEDI Canopy height at 30m resolution using multisource SAR and optical observations , EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10091, https://doi.org/10.5194/egusphere-egu23-10091, 2023.

The role of remote sensing observations in quantifying the biomass of forests is frequently debated because of both their strengths and limitations. Satellite remote sensing is nowadays standard in research activities thanks to missions designed to last over decades. Nonetheless, satellites cannot measure the organic mass stored in trees. As such, indirect approaches are developed that combine multiple observations and mathematical models together with ground-based observations to provide a set of estimates presented in the form of a map.

While small-scale studies profit from a strategy that collects observations best suited to estimate biomass, continental and global mapping efforts need to restrict to datasets that have been collected following observation plans and are free of charge. In turn, this increases the demand on the performance of the models selected to link the predictor metrics derived from remote sensing and the response variable biomass. A map of biomass is eventually the result of an interplay between sensitivity of the remote sensing data to response forest variables, the spatial resolution of the sensors, the number of remote sensing observations and the capability of the models to reproduce the relationship between predictors and response variables. A consequence of such interplay is the level of accuracy affecting the biomass estimate, which ultimately is a key parameter to inform user communities on the reliability and efficiency of biomass maps. A comparison of biomass estimates obtained with different predictors and models for the same region provides additional measures to increase our understanding of the uncertainty affecting current biomass maps derived from satellite data.

In this presentation, we explore such uncertainties by comparing four maps of forest aboveground biomass (AGB) based on satellite images acquired in 2020 and covering Europe. The maps were based on different predictors (Sentinel-1 and ALOS-2 PALSAR-2, ASCAT, SMOS as well as spaceborne LiDAR metrics) but share the same modelling framework for biomass retrieval. Depending on the spatial resolution of the satellite data, spatial scales ranging between 100 m and 25 km were covered.

Validation of each of the datasets indicates that the overall spatial distribution of AGB is well captured even in regions with dense mature forests. However, the maps show substantial discrepancies at the level of individual pixels, regardless of the set of predictors. In addition, the precision of individual AGB estimates is rather low, between 30 and 50% of the estimated value. AGB biases were identified in specific regions and were mostly explained as imperfect modelling of the relationship between predictors and forest variables. The maps’ precision increases with spatial averaging; nonetheless, the spatial correlation of errors implies that the resulting estimates can still be affected by non-negligible uncertainty. These results in turn explain why AGB values from the different maps are highly correlated although the magnitudes can be substantially different. In conclusion, the reliability of biomass maps from satellite data is questionable at the scale of the spatial resolution; their use is instead advised at the landscape scale and for understanding broad spatial patterns.

How to cite: Santoro, M., Cartus, O., Miettinen, J., Antropov, O., Araza, A., and Herold, M.: Understanding the uncertainty of forest aboveground biomass maps derived from satellite observations, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9562, https://doi.org/10.5194/egusphere-egu23-9562, 2023.

Forest canopy disturbances such as caused by bark beetle, fire, windthrow or harvest have increased in the past three decades and are expected to increase further in response to climate and land use change. Consistent information on forest canopy disturbances is therefore essential to understanding changes in forest dynamics, structure and demography over time and space. As part of the ForestPaths Horizon project we aim to create the next generation forest disturbance maps, extending both the time frame and context of existing pan-European forest disturbance assessments. Disturbances are mapped using the Landsat archive at 30 meters resolution for 1984-2021. A new machine-learning based approach trained on manually labelled reference pixels is applied to the time series, estimating forest disturbances annually and accounting for stand-replacing and non-stand replacing disturbances, as well as different causal agents (i.e., bark beetle, fire, windthrow or harvest). Summarising annual disturbance maps over time ultimately allows to detect multiple disturbance events and recovery signals per pixel and thus for the characterization of complex disturbance trajectories (e.g., multiple fires, thinnings before final harvest). We test our approach at national levels for three countries accounting for three forest biomes: boreal (Finland), temperate (Germany) and Mediterranean (Spain); covering a total land area of 1,194,526 km² and a total of 65,623 Landsat images. The results from those initial tests will provide information on the accuracy and precision of the annual, wall-to-wall maps of forest disturbances and pave the road for a consistent disturbance monitoring system of all of Europe’s forests.

How to cite: Viana Soto, A. and Senf, C.: Next generation of European forest disturbance maps based on the Landsat archive, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2769, https://doi.org/10.5194/egusphere-egu23-2769, 2023.

Forest disturbance detection studies in European temperate forests are currently largely based on optical imagery, often using fixed thresholds on vegetation indices (Francini et al. 2020; Thonfeld et al. 2022) to distinguish between disturbed and non-disturbed forest. Such approaches are limited by data availability, (especially in winter due to persistent cloud cover), and do not take natural seasonal variability as a result of forest phenology into account in the signal. Radar-based disturbance monitoring has been successfully applied over wet tropical forests (Reiche et al. 2021), but implementation in Europe is challenging due to seasonal signal variability and heterogenous forest composition. In addition, the detection of low-intensity disturbances has not been widely studied. This study will explore the capability of dense Sentinel-1 C-band time series to track disturbances of varying intensities in temperate European forests, using a set of 14 experimental sites in the Netherlands as a case study. These sites contain homogeneous forest cover (Beech, Douglas Fir, and Scots Pine) and four disturbance intensities per site which were carried out at a known date. They simulated clearcut, shelterwood, high-thinning and control management regimes, with 100%, 80%, 20%, and 0% basal area removed in each regime respectively, see figure. High-resolution Lidar and drone data were used to derive the canopy cover fraction at a 10m resolution pixel level, which were then compared with Sentinel 1 backscatter timeseries. The results indicate that at a canopy cover loss of 30-40% (of total pixel area), 75% (+-15%) of pixels are detectable as ‘disturbed’ on average. In addition, geometric effects related to radar viewing geometry such as layover and shadow affect the detection potential. Shadow effects ‘pull’ backscatter values down, while layover effects ‘push’ backscatter values up, resulting in lower detection potential at equal canopy cover loss values. Finally, it was found that using the information contained in opposing orbit directions can increase detection potential at all canopy cover loss values by mitigating inaccuracies introduced by geometric effects. Overall, these results could be of great importance in the development of a radar-based system for large scale (near-real time) disturbance detection in European temperate forest.

References

Francini, Saverio et al. 2020. “Near-Real Time Forest Change Detection Using PlanetScope Imagery.” European Journal of Remote Sensing 53(1): 233–44.

Reiche, Johannes et al. 2021. “Forest Disturbance Alerts for the Congo Basin Using Sentinel-1.” Environmental Research Letters 16(2).

Thonfeld, Frank et al. 2022. “A First Assessment of Canopy Cover Loss in Germany’s Forests after the 2018–2020 Drought Years.” Remote Sensing 14(3): 562.

How to cite: van der Woude, S., Reiche, J., Herold, M., Sterck, F., and Nabuurs, G.-J.: Towards radar-based disturbance detection in temperate forest: Testing the limits of Sentinel-1 C-band backscatter in the detection of canopy cover loss using experimental sites, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15039, https://doi.org/10.5194/egusphere-egu23-15039, 2023.

According to recent studies, many semi-arid forests are rapidly declining, which necessitates a profound understanding of the processes and causes of degradation. The Zagros Forest in Iran has been degraded during the past few decades. The analysis of forest degradation in this region is still in its initial phases, with no detailed investigation of the underlying causes. Understanding forest degradation is crucial for more effective forest management, particularly in arid and semi-arid regions.
Since one core principle of remote sensing is to identify changes in signals from multiple acquisitions that relate to the status of vegetation in a specific area, they offer efficient techniques for assessing forest degradation across large and rarely accessible forests. Frequently used remote sensing metrics to assess vegetation health are measures of vegetation greenness. For example, vegetation indices may be computed from optical remote sensing data and used to quantify forest degradation over time. Time series of vegetation indices can track forest degradation across large areas by identifying the decrease in photosynthetic activity caused by leaf loss, defoliation, and structural changes in trees.
However, numerous studies examining forest degradation either focus on dense forests or use very high-resolution remote sensing data, which is often expensive and generally difficult to obtain for large regions. Furthermore, most forest monitoring studies using remote sensing have focused on deforestation rather than forest degradation. Detecting forest degradation is challenging compared with detecting tree mortality induced by abrupt disturbances because degradation processes last longer and have a more subtle signal. 
Because of the free accessibility, relatively high spatial resolution, and long and consistent acquisition record, Landsat time series are a viable source of data for monitoring and assessing forest degradation and disturbances, as well as providing continuous reporting on forest changes. There are several methods to monitor forest disturbances, but most of these are better suited to monitoring large-scale deforestation than subtle changes in forest status. These include Landsat-based detection of trends in disturbance and recovery (LandTrendr) and breaks for additive season and trend (BFAST).
The aim of the study is to compare the mentioned algorithms with other methods, such as random forest classification, anomaly analysis, and Sen's slope. We applied the aforementioned methodologies to Landsat time series data from 1986 to 2021 to separate healthy from declining forest patches in a representative portion of the Zagros.
The highest random forest accuracy result returned an overall accuracy and kappa value of 0.77 and 0.54, respectively. The most accurate results of the anomaly analysis were an overall accuracy and kappa value of 0.58 and 0.005, respectively. Sen's slope had the lowest accuracy among the applied methods, with the highest overall accuracy and kappa values of 0.53 and 0.0039, respectively. These results indicate that the detection of degraded forest regions using Landsat data is challenging and may only be possible if additional information is added to the analysis. We hypothesize that a particularly weak vegetation signal of sparse canopy cover before the bright soil background hampers the detectability of subtle degradation processes.

How to cite: Shafeian, E., Fassnacht, F., and Latifi, H.: Using Landsat Time Series to detect forest degradation in semi-arid areas, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17409, https://doi.org/10.5194/egusphere-egu23-17409, 2023.

The construction of logging roads has a major ecological impact on tropical forests and leads to large carbon emissions (Kleinschroth & Healy, 2017, Umunay et al. 2019). Negative impacts and emissions of logging roads could potentially be drastically lowered with the adoption of reduced-impact logging practices (Umunay et al. 2019). Accurate, timely, and dynamic logging road maps would help quantify and prioritize opportunities for improved road management and forest conservation across the globe. However, to-date, the limited mapping of logging roads has required time-consuming field data collection or manual digitization from satellite images. The open availability of Sentinel-1 radar and Sentinel-2 optical satellite imagery at high spatiotemporal resolutions now offers a unique opportunity for better automated logging road monitoring in the tropics.

In this study, we employ Sentinel-1 and Sentinel-2 data for near real-time mapping of logging roads in the Congo Basin tropical forests. We monitor newly constructed roads based on Sentinel-1 change ratio composites and cloud-masked Sentinel-2 composites. We acquired an extensive reference dataset of manually digitized logging roads to train and test a convolutional neural network for road/non-road classifications.

First results indicate promising capacities of Sentinel-1 and -2 data to monitor logging roads especially in forest types in the Republic of Congo and the Democratic Republic of Congo. Forest landscapes in Gabon, Equatorial Guinea and Cameroon appeared to be more challenging for logging road monitoring due to effects of cloud-cover and elevation. Near-future work includes model refinements, the acquisition of more reference data, and a Google Earth Engine-based wall-to-wall application of our model to produce a dynamic Congo Basin logging road dataset.

References:

Kleinschroth, Fritz, Healy, John R. (2017), Impacts of logging roads on tropical forests, Biotropica 49(5): 620–635 2017

Umunay, Peter M., Gregoire, Timothy G., Gopalakrishna, Trisha, Ellis, Peter W., Putz, Francis E. (2019) Selective logging emissions and potential emission reductions from reduced-impact logging in the Congo Basin, Forest Ecology and Management 437 360-371

How to cite: Slagter, B., Fesenmyer, K., and Reiche, J.: Rapid monitoring of Congo Basin logging roads with Sentinel-1 and Sentinel-2 data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11391, https://doi.org/10.5194/egusphere-egu23-11391, 2023.

Canopy spatial structure plays an essential role in ecosystem function and the carbon cycle. The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) provided continuous three-dimensional sampling observation that can be used to derive canopy structure parameters. Although ICESat-2 data is delivering global estimates of forest structure, analysis of the performance of ICESat-2 data across a range of forest conditions remains limited. Therefore, the overall goal of this study was to evaluate the structural estimates of plant area index (PAI) from ICESat-2 data over temperate deciduous forest structural types. The PAI was derived using the geolocated photon data (ATL03) and the segment-based path length distribution method based on 100-m ICESat-2 vegetation product data (ATL08) segments. The ground-measured data used to evaluate the accuracy of PAI inversion at 100-m ATL08 segments was collected in the Saihanba forest reservation, northern China, which was covered by temperate deciduous needle-leaved forest. The results showed that the ICESat-2 PAI was in good agreement with ground-measured data, which indicated that the method had a better performance in retrieving PAI with ICESat-2 data. Moreover, we compared the effects of the characteristic of signal photons in the segments on the accuracy of PAI inversion and found that the accuracy of PAI inversion was limited by the quality of signal photons. Findings from this study highlight the method for estimating PAI with ICESat-2 data that may be suitable for a range of cover types.

How to cite: Guo, D., Hu, R., and Song, X.: Performance evaluation of ICESat-2 laser altimeter data for retrieving plant area index, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12493, https://doi.org/10.5194/egusphere-egu23-12493, 2023.

High-elevation trees cannot always reach the thermal treeline, the potential upper range limit set by growing-season temperature. But delineation of the realized upper range limit of trees and quantification of the drivers, which lead to trees being absent from the treeline, is lacking. Here, we used 30 m resolution satellite tree-cover data, validated by more than 0.7 million visual interpretations from Google Earth images, to map the realized range limit of trees along the Himalaya which harbours one of the world’s richest alpine endemic flora. The realized range limit of trees is ~800 m higher in the eastern Himalaya than in the western and central Himalaya. Trees had reached their thermal treeline positions in more than 80% of the cases over eastern Himalaya but are absent from the treeline position in western and central Himalaya, due to anthropogenic disturbance and/or pre-monsoon drought. By combining projections of the deviation of trees from the treeline position due to regional environmental stresses with warming-induced treeline shift, we predict that trees will migrate upslope by ~140 m by the end of the twenty-first century in the eastern Himalaya. This shift will cause the endemic flora to lose at least ~20% of its current habitats, highlighting the necessity to reassess the effectiveness of current conservation networks and policies over the Himalaya.

How to cite: Wang, X., Wang, T., Xu, J., Shen, Z., Yang, Y., Chen, A., Wang, S., Liang, E., and Piao, S.: Enhanced habitat loss of the Himalayan endemic flora driven by warming-forced upslope tree expansion, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13836, https://doi.org/10.5194/egusphere-egu23-13836, 2023.

Circumboreal forests represent close to 30% of all forested areas and are changing in response to climate, with potentially important feedback mechanisms to regional and global climate through altered carbon cycles and albedo dynamics (e.g., Loranty et al., 2018). A large portion of these boreal forests are located in Siberia, Russia. Here the forests are made up of mainly two types: evergreen (coniferous i.e., Pine, Picea) and summergreen (deciduous i.e., Larix) needle-leaf.  The evergreen–summergreen forest zone stretching across Western Central Yakutia is a dynamic vegetation transition zone with high disturbance due to forest fire and potential invasion of evergreen forest taxa to the east into the summergreen dominated forest zone that needs mapping and monitoring.

Sentinel-2 based Remote Sensing offers the opportunity to obtain forest type maps on a 10-20 m spatial scale. We provide in the SiDroForest (Siberian drone-mapped forest inventory) data collection (https://doi.org/10.1594/PANGAEA.933268), a Sentinel-2 data set containing Level-2 Bottom of Atmosphere labelled image patches for the early (April-May), peak (June-July) and late (August-September) summer seasons (van Geffen et al., 2022). This dataset contains 63 30 by 30-meter labelled patches with vegetation labels assigned derived from fieldwork measurements taken by the Alfred Wegener Institute in Siberia, Russia in 2018.

Building on the SiDroForest dataset, we used K-means clustering to perform an unsupervised classification of Sentinel-2 for five  locations from the SiDroForest set.  We then assigned two broad forest classes in the Sentinel-2 images, summergreen and evergreen. We used the SiDroForest Sentinel-2 patches as validation data for the K-means generated classes in addition to the fieldwork and expert knowledge.

The new dataset contains 100,000 labelled pixels, distributed over  the two classes. We created the dataset for three time stamps to include different forest phenophases in the classification. The phenophases make it easier to distinguish between the two types of forests as summergreen’s spectral signal changes significantly over the seasons.

We trained a Gaussian Naïve Bayes (GNB), a Random Forest (RF) and a Decision Tree (DT) classifier on three-time stamps separately. A combination of Sentinel-2 bands and the NDVI were evaluated with the different classifiers. The highest average accuracy score was achieved with a DT classifier and a balanced set for the two classes and the early summer time stamp and the NDVI band (82%). The peak summer also performed decently with 74%, but the accuracy dropped to 60% for the late summer time stamp. 

We used the trained DT to classify Sentinel-2 data at two locations in Siberia ; Lake Khamra and Nyurba. We masked out all non-forest data and created a forest map to measure the distribution of evergreen and summergreen over the larger areas. With our analyses we will improve the understanding of satellite data for monitoring remote places. The insights from the Siberian boreal forests are valuable in analyses of the boreal forests located in other parts of the world as well in these times or rapidly changing climate.

How to cite: van Geffen, F., Hänsch, R., Demir, B., Kruse, S., Herzschuh, U., and Heim, B.: Evergreen and summergreen classification with Sentinel-2 data, K-means clustering derived labels and Machine learning methods, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16242, https://doi.org/10.5194/egusphere-egu23-16242, 2023.

The availability of high-resolution remote sensing imagery has enabled precise mapping of the forest cover at large scale. Since forest cover evolves due to land use change, climate and extreme events, understanding its past dynamics becomes crucial in a changing climate context. In this work we analyze historical aerial imagery acquired in Switzerland since 1946 [2] for high-resolution forest mapping. We focus on the 1500–2500m a.s.l. altitude range in the Valais and Vaud Alps, where agricultural land abandonment and climate change have caused forest cover changes.

The times series are composed of single-band, panchromatic images until year 1998, then RGB images up to the year 2020, the last acquisition date over our study area. As a reference for the forest cover in 2020, we use the Topographic Landscape Model SwissTLM3D [1]. For previous years, we plan to manually generate labels to evaluate our results.

We frame forest mapping as a multi-temporal semantic segmentation task: given a time series of images, we predict a map for each image
attributing every pixel to the class "forest" or "non-forest". To solve this task, we develop a deep learning model composed of:

• a segmentation module, trained with the images and labels from the year 2020;
• a temporal module, which takes consecutive features generated by the segmentation module and outputs a multi-temporal segmentation map. This module is trained using a Mean Squared Error (MSE) loss enforcing temporal consistency.

We analyze predictions obtained with three models, each one containing one or two of the modules described above. We observe that using the full spectral information of the input images leads to a better delineation of forest borders for both old and recent images (Table 1, Figure 1). By adding the temporal module, the accuracy on the last image is practically unchanged (Table 1), while temporal consistency along the time series is improved (Figure 2).

Our method is currently not suited for abrupt forest loss, and is prone to error spreading from previous predictions. Future work will consist in designing a temporal consistency loss that better reflects known dynamics of the forest cover, in order to obtain a more accurate segmentation for the oldest images and encourage physical consistency across time.

References
[1] Swisstopo. SwissTLM3D. https://www.swisstopo.admin.ch/en/geodata/landscape/tlm3d.html [Online; accessed 06.01.2023].
[2] Swisstopo. Orthoimages. https://www.swisstopo.admin.ch/en/geodata/images/ortho.html [Online; accessed 06.01.2023].

How to cite: Nguyen, T.-A., Rußwurm, M., Kellenberger, B., and Tuia, D.: Mapping forest cover dynamics in the Swiss Alps using 70 years of aerial imagery, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8798, https://doi.org/10.5194/egusphere-egu23-8798, 2023.

Airborne laser scanning (ALS) data has been widely used for the assessment of various forest inventory parameters, such as forest stand height, biomass, etc. However, the spatial distribution of the ALS point cloud can be affected by various factors related to the survey methodology and forest stand characteristics. This study uses national coverage high-resolution ALS data with minimum point density of 4 points per square meter in combination with National forest inventory (NFI) field data to construct forest stand height models for forest stands dominated by 6 most common tree species in Latvia in mixed forest stand conditions- Pinus sylvestris L., Betula pendula Roth, Picea abies (L.) Karst, Populus tremula L., Alnus incana (L.) Moench and Alnus glutinosa (L.) Gaertn. We also take into account the ALS technology used and variations in the growing season. The ALS point cloud data was cut along the borders of the NFI plots and a statistical analysis of the spatial distribution of points within the borders of the NFI plots was performed. The results show that the RMSE value of the linear model using all NFI plot data is 1.91m, while the data sets divided by different tree species and seasonality reach the RMSE value in the range of 1.4m to 3.8m for Scots pine and Birch respectively.

Key words: Forest inventory, airborne laser scanning, phenology, large scale forest mapping

How to cite: Ivanovs, J. and Lang, M.: Impact of different tree species composition and seasonality on forest stand height predictions using airborne laser scanning and National forest inventory data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17260, https://doi.org/10.5194/egusphere-egu23-17260, 2023.

Forest edges represent the transition zone between the open country side and the forest interior. Ideally, the edge zone consists of a shrub belt (vegetation height < 4 m) and a shelterbelt with a distinct height gradient towards the forest interior. Forest edges provide several ecological functions. They offer habitats for plant and animal species, regulate fluxes of nutrients and pollutants between surrounding agricultural areas and the forest, or regulate the microclimate. Repeated assessment of forest edge conditions will help forest owners the maintenance of these edge functions considering the increased pressure due to e.g., intensified agriculture.

The aim of the ongoing project is to provide a map of forest edge structure characterization for entire Switzerland (total forest edge length 186’773 km). We used the latest freely available full coverage airborne laser scanning (ALS) data, which provides point densities of 15-20 points/m² and a forest mask provided by the Swiss National Forest Inventory (NFI). The high point densities allowed to assess both the horizontal and the vertical structure of the forest edges. On the one hand, we extracted information on the edge composition which is closely related to parameters extracted within NFIs. These comprise detailed information on the shelterbelt composition including its slope, presence or absence of the shrub layer and detection of overhanging trees. Furthermore, we computed the vegetation height distribution and the number of vegetation layers within the edge zone. On the other hand, ALS data was used to compute additional features such as the horizontal canopy cover and canopy gaps, the vertical canopy density, and the 3D light availability within the forest edge zone.

We followed a sampling-based approach and characterized the forest edge structure for discrete sampling points at the edge of the forest mask. The forest edge zone covers an area of +/- 25 m along the forest mask from the sample point and +/- 30 m into and outside of the forest, respectively, measured perpendicular to the forest mask edge. Validation of the derived parameters is based on more than 300 terrestrial NFI plots comprising forest edges. This discrete sampling-based forest edge characterization map could potentially be transferred into a quasi-continuous forest edge description by increasing the number of sampling-points on the forest edge.

Repeated six-yearly ALS acquisitions by the Federal Office of Topography swisstopo will enable to produce regular forest edge characterization data sets as a basis for the monitoring of the forest edge development at a countrywide extent. This will help to identify forest edge areas which are at severe threat of degradation and thus require treatment intervention. Thereby, the quality of forest edges can be preserved or improved for important ecosystem services.

How to cite: Bruggisser, M., Wang, Z., and Waser, L. T.: Countrywide characterization of forest edge structure from airborne laser scanning data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2288, https://doi.org/10.5194/egusphere-egu23-2288, 2023.

This work aims to develop a highly automated workflow for generating a forest cover map and detecting forest gaps at the countrywide level (i.e. Switzerland) using the alpha shape approach (Edelsbrunner et al. 1983).

Forest provides society with several functions. In Switzerland e.g., more than 50% of the forests have a protection function and mitigate or prevent the impact of a natural hazard. The accurate detection of forest gaps (openings in the forest canopy) is crucial for properly managing and planning protection forests. In addition, knowledge of the distribution of forest gaps is a useful indicator to assess forest structure and biodiversity. Although the required information is collected at the plot level within the framework of the National Forest Inventory (NFI), remote sensing allows us to derive spatially explicit and accurate products at the pixel level for the entire country.

The countrywide available 1 m spatial resolution Vegetation Height Model (VHM) (Ginzler & Hobi, 2015) serves as a basis to extract forest cover and forest gaps. The VHM was generated from image-based point clouds acquired between 2013 and 2021 for the full coverage of Switzerland. In the first step, a forest cover map was derived using the VHM. In a second step, a dense forest cover map was generated and forest gaps were delineated taking into account the Swiss NFI forest definition criteria comprising minimum tree height and width, crown coverage, and land use. In summary, the overall workflow consists of extracting the tree top points from the VHM (FINT software). Erroneous tree tops were removed using the probability forest mask derived from Sentinel-1/-2 data (Rüetschi et al. 2021). We then derived forest area and non-forest area polygons from the filtered tree top points using alpha shape (lasboundary, LAStools from rapidlasso) that computes a boundary polygon that encloses the points.

A dense forest cover map is calculated using a moving window approach and forest areas greater than 60% are extracted. The forest gaps detection within the dense forest cover map follows a similar approach adopted for the forest cover map, but the alpha shape polygons are extracted from the VHM which is converted to the las format. The entire workflow is developed in Python.

Accuracy assessments of forest cover boundary and forest gaps based on terrestrial and stereo image-interpreted NFI plots are promising and reveal an overall agreement of more than 95% over the entire country.

Reference

Edelsbrunner, H., Kirkpatrick, D.G., Seidel, R., 1983. On the shape of a set of points in the plane. IEEE Transactions on Information Theory, 29(4), pp.551-559.

Ginzler, C. and Hobi, M.L., 2015. Countrywide stereo-image matching for updating digital surface models in the framework of the Swiss National Forest Inventory. Remote Sensing, 7(4), pp.4343-4370.

Rüetschi, M., Weber, D., Koch, T.L., Waser, L.T., Small, D. and Ginzler, C., 2021. Countrywide mapping of shrub forest using multi-sensor data and bias correction techniques. International Journal of Applied Earth Observation and Geoinformation, 105, 102613.

How to cite: Rüetschi, M., Piermattei, L., Marty, M., and Waser, L. T.: Automated detection of countrywide forest cover and forest gaps using alpha shape, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11458, https://doi.org/10.5194/egusphere-egu23-11458, 2023.

The recent explosion in availability of high resolution remote sensing technologies and, crucially, the tools to analyse the 3D data they produce is leading to substantial interest in using them for widespread forest structural monitoring. The level of detail contained in the entire 3D shape of trees, fully captured within these data, can generate a wide range of metrics of interest to ecologists, but the potential metrics of interest and their uncertainties have not been fully explored. In particular, the value of different technologies - whether passive or active sensors, and from the ground or the air - for accurately deriving different metrics is not well known.

Working across a range of European forest ecosystems, we have constructed a unique 3D dataset of European forest structural properties from passive and active sensors. We segment individual trees from concurrent and co-located Structure from Motion photogrammetry (SfM) (passive sensor), and UAV LiDAR, and terrestrial laser scanning (active sensors) campaigns, and use these to compute tree structural metrics. We compare the ability of these different technologies to accurately measure key tree properties across a diversity gradient in multiple biomes.

How to cite: Lines, E., Flynn, W., Grieve, S., Owen, H., and Ruiz-Benito, P.: Comparison of extracted ecological features of forests from multiple 3D technologies, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8254, https://doi.org/10.5194/egusphere-egu23-8254, 2023.

This work reports on a novel C-band monostatic UAV-radar system deployed over two forested wetlands in arctic Sweden, near to the Abisko research station. A Videodrone X4S drone acted as the carrying body, allowing programmable and repeatable flight paths. The radar system is multi-polarized (VV, VH, HV, HH), using one transmitter and optionally one or two receivers. The radar operates in a sawtooth FMCW mode, monotonically stepping in frequency across 5.2 GHz to 5.6 GHz. The choice of sweep time (1 to 8 ms) and number of data points (128 to 2048) are programmable and selected before a flight. The radar is triggered in flight manually from the ground using a Wi-Fi link, and which then repeatedly loops over a pre-set number of 10 s scans. Here, the choice of a drone speed of 5 m s-1 meant that each scan covered a 50 m flight line. There is a re-setting time of 0.3 s between scans. The wetlands are covered by a sparse forest, primarily of birch typically 3 to 7 m tall. We used the tomographic profiling (TP) scheme to collect high-resolution maps of the vertical scattering through the forest canopy. Such information is not available from the coarser satellite imagery, which provides no information on the vertical distribution of the backscatter, not even on the relative strengths of the ground and canopy returns. As the TP scheme has the antennas forward facing, only a narrow image transect beneath the flight path is collected. A synthetic aperture technique is used both to sharpen the real beam in the along track direction, and additionally steer it in angle. Thus, post-measurement, a single flight can be processed to capture the incidence angle response of the whole scene at a single incidence angle, selectable over a ~40 degree range. The results show how the forest backscatter response changes from one dominated by a ground return close to nadir viewing, to one dominated by the canopy above 20 degrees incidence angle. Comparisons and comment will also be provided of the differing responses with polarisation. 

How to cite: Escobar-Ruiz, V., Morrison, K., Sjögersten, S., Siewert, M. B., and Fox, N.: A novel C-Band UAV-Radar for 3D characterisation of forest canopy backscatter profiles - Preliminary results, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8454, https://doi.org/10.5194/egusphere-egu23-8454, 2023.

Automated detection with deep learning opens the way for large-scale mapping of individual trees from aerial or satellite imagery. Convolutional neural networks offer unprecedented performance, under the condition that numerous and accurate labels are available to train and evaluate networks. Those two conditions are difficult to meet in the context of tree mapping, due to the high variability of tree shapes, species and environments, and to the lack of unambiguous ground truth data. Consequently, models learn on noisy data, do not reach optimality, and the errors seen during training are propagated to the predictions.

Here, we characterize and address the different types of noise in individual tree labels, notably comission/omission errors and positional errors. We propose a new method for tree detection, with an additional degree of freedom to account for annotation errors. We train and evaluate models on two large-scale datasets of aerial images in Denmark and France with manual annotations. Our approach, along with model ensembling, is able to learn from noisy point annotations and generalizes well to new areas, including dense forests.

How to cite: Gominski, D., Brandt, M., and Fensholt, R.: Deep learning for individual tree detection with noisy labels, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12859, https://doi.org/10.5194/egusphere-egu23-12859, 2023.

Close-range technologies capable of capturing forest ecosystems in three-dimensional space with great detail are revolutionising precision forestry research and practice, mainly by increasing the level of automation for data collection and processing. Furthermore, they provide options to measure some parameters directly, for example, volume or biomass. However, automatic tree species recognition still needs to be properly solved, which is a crucial and challenging task. A couple of approaches by different authors were done to overcome the challenge when data from close-range technologies are used. The authors mainly utilised 3D structures of whole trees or, in some cases, bark structures using point clouds. Or derived 2D blueprints of whole trees from point clouds to distinguish between tree species. In our approach, we are using images of bark. Usually, images are taken during the data acquisition by close-range technologies as a resource for photogrammetry or for colourising the point clouds in the case of terrestrial laser scanning, for example. Carpentier et al. (2018) did an experiment with 23 tree species in Canada and used convolutional neural networks to classify tree species with an accuracy of almost 94%. We focused on benchmarking multiple machine learning and deep learning algorithms in our experiment. Namely: Random forest; Decision tree; Support Vector Machine; Gradient boost; K-nearest Neighbors; Gaussian Naïve Bayes; Multilayer Perceptron; Convolutional neural networks.

In our first experiment, we collected two datasets of bark images using Sony alfa 7 and Canon EOS 4000D. We have collected 1755 images in Slovakia (1369) and Czechia (386); both datasets contain four tree species. The four species from Slovak datasets are European beech, sessile oak, Norway spruce, and European silver fir. Czechia data consists of the species European beech, large-leaved linden, Norway maple, and Scots pine. However, the bark images from Slovakia are from managed forests, and there is a variety of markings on bark; for that, images are cropped to small regions excluding the markings.

The most accurate results were achieved by CNN, which provides 94% accuracy on Slovak exact cropped dataset with a 50% dropout and 91% on an exact cropped dataset with a 50% dropout. When CNN is not considered, the most accurate algorithm was Multilayer perceptron with an accuracy of 92%.

The following research will focus on implementing such tree species classification within the point cloud processing workflow when close-range technologies are used. Secondly, Carpentier et al. (2018) created Barknet 1.0, where they stored 23,000 high-resolution bark images of 23 tree species in Canada. Our next goal is to develop a database of tree species across Europe. To achieve such a challenging task, we will do it within the 3DForEcoTech COST Action, a European collaborative project focusing on close-range technologies and their implementation for precision forestry and forest ecology.

References

Carpentier, M., Giguere, P. and Gaudreault, J., 2018, October. Tree species identification from bark images using convolutional neural networks. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 1075-1081). IEEE.

How to cite: Mokros, M. and Kottilapurath Surendran, G.: A Comparative Analysis of Machine Learning Algorithms for Tree Species Recognition Using An Image-Based Approach with Implementation Potential for Close-range Technologies, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14332, https://doi.org/10.5194/egusphere-egu23-14332, 2023.

Restoration and conservation efforts in critical regions affecting large populations with adverse climatic conditions, such as the Sahel, in Africa, also provide the grounds for ecosystem services in these areas. Accurate quantification and monitoring of trees in this context are essential for effectively implementing climate mitigation strategies and supporting local communities. Satellite technologies have emerged as powerful tools to obtain carbon stock estimates. However, tree count and coverage are underestimated in these semi-arid and dryland regions, and fine-grained estimates of carbon stocks can unlock tailored management and action and generate a deeper understanding of the distribution of these stocks. We present the first high-resolution, tree-level validated approach to estimate Above Ground Carbon stock leveraging Very High-Resolution imagery (0.5m), field-collected data, and Machine Learning algorithms. Local experts and youth and women communities participating in the Great Green Wall Initiative collected individual tree geolocation in 8 sites within the drylands of the Sahel region (Burkina Faso and Niger). We built a database of tree-level aboveground carbon (AGC) based on field measurements by using allometric equations and carbon conversion factors, and we trained and validated an Artificial Neural Network to predict AGC based on remote sensing imagery variables processed on individual segmented tree crowns. The validation resulted in a R² of 0.69, a Root Mean Square Error (RMSE) of 355.6 kg and a relative RMSE of 51%. When aggregating results at coarser spatial resolutions (plot and site), the relative RMSE decreased below 20% for all areas. AGC density (AGCd) errors remained under 6 Mg ha^-1 on ranges of AGCd up to 26 Mg ha^-1, reaching errors of less than a ton of carbon per hectare for half the study sites. A comparison with other methodologies in the recent literature was carried out and showed a competitive performance of this approach in these regions, with R² of other similar studies being between 0.6 and 0.95, and RMSE ranging from 0.25 to 100 Mg ha^-1. Model results confirm the current trend of underestimating the AGC stocks in drylands using coarser resolution data. Most of the available data in the region estimated the total AGC stocks of the 8 study sites to be less than half compared to the validated model results. The only map that predicted an overshot AGC stock compared to our study was a SAR-based approach at 25-meter resolution by Bouvet et al. 2018, in which the authors claimed more significant relative errors in dry regions. Our results confirm that most previous approaches implemented in drylands produce biased estimations of carbon. Our model exploiting VHR imagery offers the possibility to remedy the lack of resolution and then aggregate at the desired level of granularity. This first-of-its-kind validation at the individual tree level demonstrates the capability of very high-resolution models to correctly assess carbon stocks in the now underestimated drylands and semi-arid areas.

How to cite: Perpinyà-Vallès, M., Escorihuela, M. J., Ameztegui, A., and Romero, L.: Accurate quantification of carbon stocks at the individual tree level in semi-arid regions in Africa, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14479, https://doi.org/10.5194/egusphere-egu23-14479, 2023.

Nowadays data-integration opens up new possibilities for land surveys, involving both remote and proximal sensing devices. The fast advancement of both technology and devices allowed researchers to gather data from afar, making these acquisitions affordable and suitable even in locations with limited accessibility. We surveyed 3 veteran chestnut trees (Castanea sativa) by integration of Mobile Laser Scanner clouds with the top of the canopies reconstructed through photogrammetry, using an Unmanned Aerial Vehicle (UAV) equipped with RGB camera. These 3D models can be used to extract  precise tree metric data, compared with those collected in the field with traditional measurements, such as diameter at breast height (DBH), total height (TH), crown basal area (CBA) and crown volume (CV), providing valuable information on tree assessment and its potential carbon stock. Moreover, the veteran trees have exceptional genetic and cultural values and therefore must be properly inventoried, monitored and protected. We conducted our surveys during summer, when the trees had a crown full of leaves and in winter, when they were almost completely defoliated. We used a GNSS and a total station to collect ground control points, based on available satellites signal. We followed a circular path all around the three veteran chestnut trees with the MLS device, scanning the entire tree from multiple angles and thus obtaining detailed and accurate point clouds of the trees’ skeleton and including at least 3 highly reflective targets. With the UAV, we collected nadiral RGB images to reconstruct the upper part of the canopies and, using the same targets, we merged them with the MLS outputs. We used a Sony Alpha77 single-lens reflex camera to collect detailed, high-quality 3D data of our veteran trunks through the process of close-range photogrammetry. The latter have been merged with the previous 3D models obtained and thus completing the veteran trees reconstruction. Through manual segmentation, we split between trees skeleton and canopy. We extracted the TH and the crown basal area in both seasons using 3DForest software. DBH has been extracted by slicing the RGB trunks at 1.30m and creating a mesh of the sliced portions while the space occupied by the crowns has been computed through the volume obtained by the mesh created with the Alpha Shape algorithm. The volume of the canopies was determined in both the winter and summer seasons to compare the space they occupy when they are in vigor with the space they take up when there are no leaves. Our results, for the 3 individuals, appear to be concordant with the DBH and the TH obtained in the field by traditional measurements while the CBA and CV have not been measured in the field since they are challenging with these ancient trees. The DBH range values are between 150 - 190 cm, the TH is between 18 - 23 m, the CBA and CV range are respectively between 165 - 176 m² and 255 – 314 m³ in winter while 180 – 258 m² and 328 – 406 m³ in the summer surveys.

How to cite: Balestra, M., Pierdicca, R., Vitali, A., Tonelli, E., Chiappini, S., and Urbinati, C.: A geomatics data integration approach for veteran chestnut trees 3D modeling, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15779, https://doi.org/10.5194/egusphere-egu23-15779, 2023.