Detecting bark beetle damage with Sentinel-2 multi-temporal data in Sweden

Abstract: The European spruce bark beetle is considered as one of the most destructive forest insects to Norway spruce trees in Europe. Climate change may increase the frequency and intensity of bark beetle outbreaks. It is therefore of vital importance to detect the bark beetle outbreaks and take it under control to prevent further damages. Remote sensing techniques may provide a cost-efficient solution to the detection of bark beetle outbreaks. In the past years, the detection of bark beetle outbreaks in Northern America has achieved success with the aid of the long time series of LANDSAT satellite images. Sentinel-2 provides satellite images of high spatial and temporal resolution which may be suitable for bark beetle detection in Europe. The extreme drought and heat in the summer in 2018 favored the outbreaks of bark beetles in central and southern Sweden. In this project, detection of two stages (gray-attack and green-attack stage) of bark beetle outbreaks in southern and central Sweden was carried out separately with Sentinel-2 level 2A satellite multi-temporal images. In bark beetle gray-attack stage detection, the two most commonly used methods: maximum likelihood and random forest classification, were performed and compared on different combinations of Sentinel-2 10m resolution raw bands sensed in March-April and VIs derived from them. Maximum likelihood classification method with EVI and GNDVI gave the highest accuracy: total accuracy of 89% and Kappa of 0.74 (substantial agreement). Random forest classification method with all variables achieved the second best result: total accuracy of 85% and Kappa of 0.62 (substantial agreement). The two best methods were thereafter applied to two test areas in southern (test area 1) and central Sweden (test area 2). Random forest classification method with all variables obtained higher accuracy: total accuracy of 76% and Kappa of 0.53 (moderate agreement) in test area 1 and total accuracy of 71% and Kappa of 0.39 (fair agreement) in test area 2. Based on detection result from the first part, random forest classification method was employed for bark beetle green-attack stage detection. A series of VIs derived from Sentinel-2 20m resolution bands sensed in the summer in 2018 were calculated and the importance of the VIs and raw bands were ranked with random forest algorithm. The first 13 or 14 most important variables were used for classification. Results show that water content related raw bands and VIs, red-edge VIs and the NIR band are the most sensitive variables to bark beetle green-attack. Bark beetle green-attack stage detection obtained high accuracy in study area 1: total accuracy of 88% and Kappa of 0.67 (substantial agreement) on July 26th and total accuracy of 84% and Kappa of 0.58 (moderate agreement) on October 12th. Relatively low accuracy were achieved in test area 1: total accuracy of 53% and Kappa of 0.03 (no or rarely any agreement). Moderate accuracy were achieved in test area 2: total accuracy of 64% and Kappa of 0.27 (fair agreement) on July 8th, and total accuracy of 71% and Kappa of 0.42 (moderate agreement) on July 31st.