The REFEWOODLAND-project was initiated with the intention of
complementing existing estimations of above ground tree biomass (AGB) in
forest areas with comprehensive information collected in outside forest
areas, in order to calibrate a remote sensing data-based model to
seamlessly estimate existing AGB on areas outside forests.
With
this model a nationwide above ground tree biomass-estimation outside
forest areas becomes directly possible, supporting greenhouse gas
inventory reporting.
The base remote sensing data used derives from the first national airborne laser-scanning campaign (2018-2025) conducted by Swisstopo and resulting in the swissSURFACE3D (ALS) data. The timing of field recordings was set to be parallel with the existing flight schedule for swissSURFACE3D. For the field measurements, a stratified sampling-design was developed to account for regional differences in biomass-characteristics whilst ensuring a cost-effective sampling procedure. These sample-plots were probed in the following field-campaign for whole Switzerland.
After collecting data for all sample-plots (2019-2023), the above-ground biomass (AGB) was calculated after Chave et al. (2014) and an ALS-based mixed-effects model was set-up and calibrated, using vegetation-volume metrics as explanatory variable, allowing to predict above ground tree biomass-stocks outside forest areas.
At the beginning of the project, a stratified sampling design was chosen to account for consideration of homogeneous physiological appearance (allometry) of targeted vegetation, as well as providing a cost-effective sampling method rather than a biomass inventory. This sampling approach is documented in Gardi (2018). Price (2023), describes the data derived thresholds for the model application. It is important to point out, that this stratification process does not fulfill the requirements of an inventory design, but is designed, to achieve the best consideration of potential influences in the reference data, for improved calibration of the later created model.
The introduced stratification is based on Swiss Land Use Statistics which are enhanced with additional spatial data in its most recent form. As the most recent form of the Land Use Statistics (Arealstatistik (AREA) 4, 2014-2018) was not yet finalized at the beginning of the project, missing data was completed with the former classification of AREA version 3.
As additional data, a 25 m raster was calculated from the aerial image based vegetation height model (Ginzler and Hobi, 2016) to characterize a vegetation canopy cover (vegetation > 3 m) and the National Forest Inventory (NFI) forest area layer was used to mask our forest areas.
Combined, the AREA data was then filtered by canopy cover > 5 % (NFI vegetation height model (remaining n = 571’783 points)) and a distance to forest areas (NFI) of 30 m (remaining n = 360’334 points), as those regions are either out of interest or thematically covered by the NFI.
Afterwards, a list of parameters that potentially determine growth were selected based on expert input, leading to a stratification-clustering, which is needed for equal sample drawing:
| Parameter | Description | Thresholds for defined strata |
|---|---|---|
| Landuse | Swiss Greenhouse Gas Inventory Landuse/Landcover Combination Categories (CC) derived from the Swiss Land Use Statistics (AREA 3 & 4) | Settlement (L1) Agriculture (L2) Special crops and shrubs (L3) all others (L4) |
| Bioregion | Biogeogrpahical regions of Switzerland, defined by FOEN | Jura and Mittelland (B1) Alps (all others) (B2) |
| Elevation | DHM50 swisstopo | B1: x < 450 m (E1) 450 ≤ x < 570 m (E2) 570 m ≤ x (E3) B2: x < 1100 m (E1) 1100 ≤ x < 1’890 m (E2) 1’890 m ≤ x (E3) |
| Tree height | Mean per pixel (25 m) tree height from the VHM model NFI (Ginzler and Hobi, 2016) and considering the
following boundaries: x < p0.05 (H1) p0.05 ≤ x < p0.95 (H2) p0.95 < x (H3) |
H1: x < 6.86 (B1_E1) x < 6.43 (B1_E2) x < 7.08 (B1_E3) x < 6.7 (B2_E1) x < 6.57 (B2_E2) x < 5.01 (B2_E3) H2: 6.86 < x < 15.1 (B1_E1) 6.43 < x < 14.5 (B1_E2) 7.08 < x < 16.3 (B1_E3) 6.7 < x < 15.3 (B2_E1) 6.57 < x < 15.6 (B2_E2) 5.01 < x < 12.7 (B2_E3) H3: x ≥ 15.1 (B1_E1) x ≥ 014.5 (B1_E2) x ≥ 16.3 (B1_E3) x ≥ 15.3 (B2, E1) x ≥ 15.6 (B2_E2) x ≥ 12.7 (B2_E3) |
| Canopy cover | Mean per pixel (25 m) tree canopy cover from the VHM model NFI (Ginzler and Hobi, 2016) and considering the
following boundaries: x < p0.05 (C1) p0.05 ≤ x < p0.95 (C2) p0.95 < x (C3) |
C1: x < 0.174 (B1_E1) x < 0.158 (B1_E2) x < 0.162 (B1_E3) x < 0.152 (B2_E1) x < 0.14 (B2_E2) x < 0.11 (B2_E3) C2: 0.174 < x < 0.686 (B1_E1) 0.158 < x < 0.61 (B1_E2) 0.162 < x < 0.636 (B1_E3) 0.152 < x < 0.634 (B2_E1) 0.14 < x < 0.53 (B2_E2) 0.11 < x < 0.39 (B2_E3) C3: x ≥ 0.686 (B1_E1) x ≥ 0.61 (B1_E2) x ≥ 0.636 (B1_E3) x ≥ 0.634 (B2_E1) x ≥ 0.53 (B2_E2) x ≥ 0.39 (B2_E3) |
All factors combined, result in 217 strata in total considering all parameters with 165 strata being used for the model training. Missing strata were related to canopy cover and vegetation height, which could not be found in the fields. The stratification raster for the field-data collection is provided in the data-package as “strata_comb_code.tif” and the related coding table “strata_LUT_allcomb.csv”.
To support the economic feasibility of the project, the sample drawing is proceeded randomly, but with higher probability inclusion for positive economic aspects. This is done in a two-phase procedure:
This procedure is done to ensure, that each cluster can be handled by the field team in one day, which helps to minimize unnecessary travel and set-up times during field work.
For the first selection, at least one sample point should be included per stratum and panel, with panel 1 representing Eastern-, Western-, Central- and Southern Switzerland, and panel 2, representing Graubünden, Wallis, Northern Switzerland and Bern.
The distribution of sample plots was designed as random but considering a homogeneous distribution of points. This is achieved by dividing the same number of samples up in both panels, and at least assign one plot per stratum.
Additionally, for each point an inclusion probability P by accessibility, with
\[ P = \frac{1}{t_{ÖV} + 10 \cdot d_{ÖV} + 30 \cdot d_{Str}} \]
\(t_{ÖV}\) : Travel time from Bern,
by public transport (Bundesamt für
Landestopografie, 2011)
\(d_{ÖV}\) : Distance of the plot to the
nearest public transportation station (OSM)
\(d_{Str}\): Distance of the plot to the next
street (OSM)
was defined, ranking the plots for sample drawing.
In total, and when applying this method, n1 = 521 core-plots were selected for further processing.
For each identified cluster- center, 6-9 additional sample plots were drawn within a radius of 0.2 - 1 km. If no 6 additional sample points were found, the entire cluster was discarded.
This selection results in 466 clusters, with 4644 potential sample plots for recording (Fig. 1).