Over the last decades, climate change has triggered an increase in the frequency of spruce bark beetle (Ips typographus L.) in Central Europe. More than 50% of forests in the Czech Republic are seriously threatened by this pest, leading to high ecological and economic losses. The exponential increase of bark beetle infestation hinders the implementation of costly field campaigns to prevent and mitigate its effects. Remote sensing may help to overcome such limitations as it provides frequent and spatially continuous data on vegetation condition. Using Sentinel-2 images as main input, two models have been developed to test the ability of this data source to map bark beetle damage and severity. All models were based on a change detection approach, and required the generation of previous forest mask and dominant species maps. The first damage mapping model was developed for 2019 and 2020, and it was based on bi-temporal regressions in spruce areas to estimate forest vitality and bark beetle damage. A second model was developed for 2020 considering all forest area, but excluding clear-cuts and completely dead areas, in order to map only changes in stands dominated by alive trees. The three products were validated with in situ data. All the maps showed high accuracies (acc > 0.80). Accuracy was higher than 0.95 and F1-score was higher than 0.88 for areas with high severity, with omission errors under 0.09 in all cases. This confirmed the ability of all the models to detect bark beetle attack at the last phases. Areas with no damage or low severity showed more complex results. The no damage category yielded greater commission errors and relative bias (CEs = 0.30–0.42, relB = 0.42–0.51). The similar results obtained for 2020 leaving out clear-cuts and dead trees proved that the proposed methods could be used to help forest managers fight bark beetle pests. These biotic damage products based on Sentinel-2 can be set up for any location to derive regular forest vitality maps and inform of early damage.
Bark beetle outbreaks and forest mortality have rocketed in European forests because of warmer winters, intense droughts, and poor management. The methods developed to predict a bark beetle outbreak have three main limitations: (i) a small-spatial-scale implementation; (ii) specific field-based input datasets that are usually hard to obtain at large scales; and (iii) predictive models constrained by coarse climatic factors. Therefore, a methodological approach accounting for a comprehensive set of environmental traits that can predict a bark beetle outbreak accurately is needed. In particular, we aimed to (i) analyze the influence of environmental traits that cause bark beetle outbreaks; (ii) compare different machine learning architectures for predicting bark beetle attacks; and (iii) map the attack probability before the start of the bark beetle life cycle. Random Forest regression achieved the best-performing results. The predicted bark beetle damage reached a high robustness in the test area (F1 = 96.9, OA = 94.4) and showed low errors (CE = 2.0, OE = 4.2). Future improvements should focus on including additional variables, e.g., forest age and validation sites. Remote sensing-based methods contributed to detecting bark beetle outbreaks in large extensive forested areas in a cost-effective and robust manner.
The spatial and temporal dynamics of the forest cover can be captured using remote sensing data. Forest masks are a valuable tool to monitor forest characteristics, such as biomass, deforestation, health condition and disturbances. This study was carried out under the umbrella of the EC H2020 MySustainableForest (MSF) project. A key achievement has been the development of supervised classification methods for delineating forest cover. The forest masks presented here are binary forest/non-forest classification maps obtained using Sentinel-2 data for 16 study areas across Europe with different forest types. Performance metrics can be selected to measure accuracy of forest mask. However, large-scale reference datasets are scarce and typically cannot be considered as ground truth. In this study, we implemented a stratified random sampling system and the generation of a reference dataset based on visual interpretation of satellite images. This dataset was used for validation of the forest masks, MSF and two other similar products: HRL by Copernicus and FNF by the DLR. MSF forest masks showed a good performance (OAMSF = 96.3%; DCMSF = 96.5), with high overall accuracy (88.7–99.5%) across all the areas, and omission and commission errors were low and balanced (OEMSF = 2.4%; CEMSF = 4.5%; relBMSF = 2%), while the other products showed on average lower accuracies (OAHRL = 89.2%; OAFNF = 76%). However, for all three products, the Mediterranean areas were challenging to model, where the complexity of forest structure led to relatively high omission errors (OEMSF = 9.5%; OEHRL = 59.5%; OEFNF = 71.4%). Comparing these results with the vision from external local stakeholders highlighted the need of establishing clear large-scale validation datasets and protocols for remote sensing-based forest products. Future research will be done to test the MSF mask in forest types not present in Europe and compare new outputs to available reference datasets.
Forests cover one third of Europe’s land and significantly contribute to the regional economy. Moreover, they play an essential role in climate regulation. Traditional inventory-based forest data update is often much lower than required. Remote Sensing is a valuable source for forest monitoring, as it provides periodic data on vegetation status. In this context, EU Horizon-2020 MySustainableForest project (MSF, Grant Agreement nº 776045) aims at developing remote sensingderived geo-information services for integrated forest management through a web service platform. An unsupervised method to obtain a forest mask over European forests using optical Sentinel-2 data was implemented. Kmeans algorithm was used for segmenting the images in clusters, which were subsequently assigned to a forest class depending on its overlap with the forest classes of ancillary land cover data. The resulting classification was refined applying a filter and a vegetation mask. The algorithm was tested over 16 sites representing Europe’s main biogeographic regions. A confusion matrix was built using points selected via photointerpretation. Validation metrics were computed from the confusion matrix. The results showed that it is possible to develop an automatic forest mask for Europe, (overall accuracy above 90%). Accuracies varied depending on forest characteristics. Best results were achieved in Boreal and Continental forests. Although the algorithm was tuned to consider the diversity of European forests, there is scope for improving the adaptability of MSF Forest Mask, mainly in the southern Mediterranean region, where the mixed effect of tree-grass formations hindered a better forest discrimination. These results may be of interest to forest and land managers and climate modellers.
Forests cover 36.5% of Spanish land. Natural and man-induced disturbances are causing important changes in species distribution. As Spanish National Forest Inventory is updated every 10 years, a more recurrent periodic data source providing information on species distribution is needed in order to predict changes in forest area and composition. Remote Sensing meets this demand, as it provides periodic and spatially continuous data on forest status. In this context, MySustainableForest (MSF) H2020 project aims at providing remote sensing-based geo-information services through a web service platform. One of MSF products is a classification of main forest types, whose preliminary development was tested over a 950 km2 area located in Northern Spain. A Random Forest model was trained with data delineated with the help of local forest data. The output was validated using stratified k-fold cross-validation. Validation metrics were computed from the confusion matrix for each class separately and for the total set of classes. Although overall metrics were high (OA = 95%; DC = 85.1%), they varied significantly for different classes (e.g., Fagus sylvatica was classified with higher accuracy than Pinus nigra, which was mainly confused with other Pinus species), showing that species with higher seasonal variations were easier to identify. Random Forest feature importance ranking showed that bands in the near-infrared (NIR) and shortwave-infrared (SWIR) wavelengths were essential to discriminate forest species, since they explained more than 40% of the variations alone and 82% in combination with Red wavelength.