Mapping within-pixel tree species proportions in boreal forest with Sentinel-2 time series data

More diverse forest management practices increase the demand for detailed information on species distribution in boreal forests. We used weekly time-series of Sentinel-2 satellite images over Norway in convolutional neural network (CNN) and multivariate random forest (RF) models trained on >10,000 National Forest Inventory field plots to map within-pixel proportions of spruce, pine and deciduous trees. The models were validated using an independent dataset with >2000 field reference plots. We also predicted the dominant tree species and analyzed model accuracy under different forest conditions. The CNN models performed marginally better than the RF models, and for species proportions, the predicted R2 values were in the range 0.39–0.52. Predictions of dominant species resulted in an overall accuracy of 71%. The accuracy of the models was better for mature and older forest with a closed canopy cover than for younger or sparsely stocked forest. Overall, the prediction accuracies were moderate, and it is evident that accurate predictions of within-pixel tree species proportions remain a challenging task. The results show that the class probabilities, as obtained from CNN models, can be useful to indicate the accuracy of single predictions.