Swiss forests get more than a dominant species: 30 m maps give basal-area shares for ten tree species
Synopsis
Swiss forest composition was previously shown as one dominant species per pixel, erasing the share of accompanying species. A deep-learning model anchored on 0.1 m aerial RGB imagery and combined with airborne laser scanning, annual AlphaEarth satellite embeddings and climate variables was trained on basal-area observations from 3,992 Swiss National Forest Inventory sites to output basal-area shares for ten main tree species and an Other class on a 30 m grid. Spatial cross-validation gave a macro coefficient of determination of 0.485; on forest-reserve data not used in model development, the macro value was 0.552 at 30 m pixel support and 0.733 at plot support, with dominant-class identification in 73.2% of pixels and 81.1% of plots.
Interpretation
Swiss forest composition can now be reported as the basal-area share of each species within every 30 m cell, not only as a dominant species. Earlier national products assigned each pixel a single dominant class, and at 10 to 30 m resolution one pixel can cover several crowns, so the mixture inside a pixel stayed unresolved. In five-fold spatial cross-validation the model reached a macro coefficient of determination of 0.485 across 11 classes, 0.105 above the strongest baseline, ResNet-50, and 0.172 above a random forest.
The model anchors on 0.1 m aerial RGB imagery and lets laser scanning, satellite embeddings and climate data each contribute a gated correction. The inputs differ in spatial support, from 0.1 m image pixels to 10 m satellite embeddings and 1 km climate grids, and in acquisition date, so simple concatenation is not appropriate. Training used 3,992 sites and 39,380 tree records surveyed in 2017 to 2022; removing satellite embeddings lowered the macro coefficient of determination from 0.485 to 0.464, removing climate variables to 0.465, and removing laser scanning only to 0.483.
Continuous shares make species such as maple and ash visible nationwide even though they rarely dominate. A dominant-class map marks maple and ash only where they hold the largest share, while across much of the Plateau and Jura they occur as minor components. The map covers about 13,516 square kilometres of Swiss forest and reproduces the elevational sequence from broadleaved lowland stands through spruce to larch and Arolla pine near the treeline.
Perspective
The demonstrated scope is the 30 m Swiss forest grid, with training labels from 3,992 National Forest Inventory sites and independent checks from 132 plots in 35 forest reserves; accuracy should be read at the spatial support where the product is used, with plot-level results for aggregated applications and pixel-level or cross-validation results for individual pixels.
The laser-scanning encoder was trained from random initialization on only 3,992 sites, national point density is about 15 to 20 points per square metre, and the first nationwide survey was flown between 2017 and 2024, so point clouds and aerial images of a site can be several years apart; the authors note these factors may partly explain confusion among spruce, fir and Arolla pine. Inventory labels come from concentric 200 and 500 square metre sample circles that cover only part of each 30 m cell, so a finer 10 m composition map would need reference labels that support that scale.