A novel approach for bilberry yield mapping integrating airborne laser scanning, map databases, and automatically interpreted terrestrial images

Picking of wild berries is a common activity in the Nordic countries, but finding productive areas can be challenging, especially for inexperienced pickers. This article presents a novel approach for bilberry ( Vaccinium myrtillus ) yield mapping based on random forest (RF) classification. A large variety of features for classification was derived from airborne laser scanning (ALS) data and existing map databases. Training data were obtained by automatic interpretation of terrestrial images. A web application was developed to visualise the berry map on a mobile phone and to combine it with real-time positioning in the forest. The complete mapping approach was tested in a forested area in southern Finland. The results suggest that the RF machine learning method has been able to identify useful mapping rules for determining growing sites with favourable conditions for bilberries. Results from a numerical quality evaluation and cumulative yield estimation suggest that the performance of the method is promising compared to previous yield modelling studies. The produced bilberry maps are not necessarily accurate at the individual pixel level, but according to the first results from user tests, the “big picture” of berry yield was correct in 97% of the tests, which is very promising from the practical point of view. The potential berry picking locations indicated by our approach represent areas where the growing conditions allow good bilberry yields, provided that weather conditions are favourable and pollination is successful. Among other topics, future research could include combination of weather observations in the mapping approach.