class: center, middle, inverse, title-slide # FSU activity report: 2015 - 2019 ## <html> <div style="float:left"> </div> <hr color='#EB811B' size=1px width=796px> </html> ### Alexander Brenning, Patrick Schratz, Jannes Muenchow ### LIFE Healthy Forest Meeting, Vitoria, 26-27 Mar 2019 --- layout: true <div class="my-header"><img src="../figs/life.jpg" style="width = 5%;" /></div> --- # Outline ### 1) Action A2 deliverables: "*Optimization of the integrated systems*" - Assistance for acquistion of hyperspectral imagery - Database of plot characterization - Database of possible predisposing factors for spatial modelling ### 2) Action B1.1 deliverables: "*Spatial mapping using statistical and machine-learning data analysis*" - Remotely sensed forest health map - Maps of forest disease potential ### 3) Action D3: "*Complementary Actions*" - Summer school Jena March 2017 --- # A2.1.1: Field monitoring and sampling ## Pathogen presence/absence .font140[ NEIKER provided field sampling data for the following pathogens: - *Armillaria mellea*: 1016 obs. (395/621) (after cleaning) - *Diplodia sapinea*: 922 obs. (700/222) (after cleaning) - *Fusarium circinatum*: 922 obs. (781/141) (after cleaning) - *Heterobasidion annosum*: 1016 obs. (785/231) (after cleaning) ] --- # A2.1.1: Field monitoring and sampling ## Defoliation data NEIKER gathered in-situ information in 2016 and 2017. ### 2016 - Demonstration plots (Luiando, Oiartzun, Laukiz 1-3, Hernani) - Sampling of all trees in the plot - Total amount of sampled trees: ~ 1800 ### 2017 - Sampling of all 28 plots (Total # of obs.: 1400) - 50 trees per plot - Sampling scheme from FSU --- # A2.1.4: Envir. condi. as predisp. factors .font140[ The following variables were collected: - Long-term **precipitation** (1950 - 2000) - Long-term **temperature** (1950 - 2000) - Long-term **PISR** (1950 - 2000) - **Soil** type (at 250 m res.) - **Lithology** type (at 1 km res.) - **pH** value (at 1 km res.) - Probability of **hail damage** at trees (200 m res.) ] --- # A2.1.6: Hyperspectral img acquisition .font160[ * Hyperspectral imagery was acquired by HAZI in September 2016 for all plots (28 in total) * Unfortunately, one of the five "demonstration plots" ("hernani") was not covered by the flight mission * FSU helped with the planning of the flight routes (Marco Pena) ] --- # A2.1.6: Hyperspectral img acquisition .font160[ * Hyperspectral satellite Hyperion-1 was decomissioned in January 2017 without prior notice * Hyperspectral Airborne AVIRIS data as a replacement not suitable (price, flights only in US) * No other spaceborne hyperspectral sensor available ] --- # A2.1.6: Hyperspectral img acquisition .font160[ <i class="fas fa-arrow-right "></i> We acquired spaceborne multispectral **Sentinel-2** data as an alternative for Hyperion-1 data. Available cloud-free mosaics of the Basque Country for the vegetation period (April - September) - 2017: 3 - 2018: 8 ] --- # A2 Deliverables .font160[ <i class="fas fa-check-square "></i> Acquisition of hyperspectral imagery <i class="fas fa-check-square "></i> Database of plot characterization <i class="fas fa-check-square "></i> Database of possible predisposing factors for spatial modelling ] --- # Outline .color-grey[ 1) Action A2 deliverables: "*Optimization of the integrated systems*" - Assistance for acquistion of hyperspectral imagery - Database of plot characterization - Database of possible predisposing factors for spatial modelling ] ### 2) Action B1.1 deliverables: "*Spatial mapping using statistical and machine-learning data analysis*" - Remotely sensed forest health map - Maps of forest disease potential ### 3) Action D3: "*Complementary Actions*" - Summer school Jena March 2017 --- # B1.1: Spatial mapping ## Defoliation mapping .font150[ 1. Training of an **Extreme Gradient Boosting** (xgboost) model - ~ 7500 Variables - 90 Vegetation Indices (VI) - 7400 Normalized Ratio Indices (NRI) 2. Extraction of the most important variables <i class="fas fa-arrow-right "></i> 7 3. Prediction of defoliation (Basque Country; 2017 + 2018) ] --- # B1.1: Spatial mapping ## Defoliation mapping .font150[ ### Model Performance RMSE: ~ 40 (defoliation) ### Most Important Vegetation Indices - "EVI", - "GDVI_2", "GDVI_3", "GDVI_4", - "mNDVI", - "mSR", - "D1" ] --- # B1.1: Spatial mapping ## Defoliation mapping (Prediction 2017)  --- # B1.1: Spatial mapping ## Defoliation mapping (Prediction 2018)  --- # B1.1: Spatial mapping ## Defoliation mapping (Histograms) <img src="2019-03-Vitoria_files/figure-html/unnamed-chunk-1-1.png" width="100%" style="display: block; margin: auto;" /> --- # B1.1: Spatial mapping ## Pathogen modeling .font160[ Tuning and training of 7 classifiers for each pathogen - Random Forest (RF) - Support Vector Machine (SVM) - Boosted Regression Trees (BRT) - Generalized Additive Model (GAM) - Generalized Linear Model (GLM) - Extreme Gradient Boosting (XGBOOST) - k-Nearest Neighbor (KNN) ] --- # B1.1: Spatial mapping ## Pathogen modeling .font150[ - Performance evaluation for all possible combinations (26/28)<sup>1</sup> - Creation of prediction maps for all possible combinations (24/28)<sup>2</sup> ] .footnote[ [1] XGBOOST and GAM did not converge on Armillaria. [2] XGBOOST cannot predict to data with new factor levels. ] --- # B1.1: Spatial mapping ## Best classifier (Performance) ### Brier Score .font150[ - Armillaria: Random Forest (0.243) - Diplodia: Random Forest (0.165) - Fusarium: Random Forest (0.128) - Heterobasidion: Random Forest (0.165) ] --- # B1.1: Spatial mapping ## Pathogen modeling (prediction maps)  --- # B1.1: Spatial mapping ## Pathogen modeling (prediction maps)  --- # B1.1: Spatial mapping ## Pathogen modeling (prediction maps)  --- # B1.1: Spatial mapping ## Pathogen modeling (prediction maps)  --- # B1.1: Deliverables .font160[ <i class="fas fa-check-square "></i> Remotely-sensed forest health map <i class="fas fa-check-square "></i> Map of forest disease potential ] --- # Outline .color-grey[ 1) Action A2 deliverables: "*Optimization of the integrated systems*" - Assistance for acquistion of hyperspectral imagery - Database of plot characterization - Database of possible predisposing factors for spatial modelling 2) Action B1.1 deliverables: "*Spatial mapping using statistical and machine-learning data analysis*" - Remotely sensed forest health map - Maps of forest disease potential ] ### 3) Action D3: "*Complementary Actions*" - Summer school Jena March 2017 --- # D3: Complementary Actions .font160[ FSU hosted a training school on "**statistical analysis of hyperspectral data**". Date: March 2017 Duration: 5 days (full-time) ### Statistics - 54 applications - 28 (international) participants Participant report: https://www.r-spatial.org/r/2017/03/25/spring-school-jena.html ] --- # D3: Complementary Actions .font150[ ### Speaker list - Marco Peña (Alberto Hurtado University, Chile) - Aneta Modzelewska (Forest Research Institute, Raszyn, Poland) - Dr. Henning Buddenbaum (University of Trier, Germany) - Dr. Tim Appelhans (University of Marburg/GfK Geomarketing Nürnberg, Germany) - Dr. Thomas Bocklitz (IPC Jena, Germany) - Prof. Dr. Alexander Brenning (FSU) - Dr. Jannes Muenchow (FSU) - Patrick Schratz (FSU) ] --- # D3: Complementary Actions  <i class="fas fa-copyright "></i> H. Goetz