This article describes a data-driven framework based on spatiotemporal machine learning to produce distribution maps for 16 tree species (Abies alba Mill., Castanea sativa Mill., Corylus avellana L., Fagus sylvatica L., Olea europaea L., Picea abies L. H. Karst., Pinus halepensis Mill., Pinus nigra J. F. Arnold, Pinus pinea L., Pinus sylvestris L., Prunus avium L., Quercus cerris L., Quercus ilex L., Quercus robur L., Quercus suber L. and Salix caprea L.) at high spatial resolution (30 m). Tree occurrence data for a total of three million of points was used to train different algorithms: random forest, gradient-boosted trees, generalized linear models, k-nearest neighbors, CART and an artificial neural network. A stack of 305 coarse and high resolution covariates representing spectral reflectance, different biophysical conditions and biotic competition was used as predictors for realized distributions, while potential distribution was modelled with environmental predictors only. Logloss and computing time were used to select the three best algorithms to tune and train an ensemble model based on stacking with a logistic regressor as a meta-learner. An ensemble model was trained for each species: probability and model uncertainty maps of realized distribution were produced for each species using a time window of 4 years for a total of six distribution maps per species, while for potential distributions only one map per species was produced. Results of spatial cross validation show that the ensemble model consistently outperformed or performed as good as the best individual model in both potential and realized distribution tasks, with potential distribution models achieving higher predictive performances (TSS = 0.898, R²_logloss = 0.857) than realized distribution ones on average (TSS = 0.874, R²_logloss = 0.839). Ensemble models for Q. suber achieved the best performances in both potential (TSS = 0.968, R²_logloss = 0.952) and realized (TSS = 0.959, R²_logloss = 0.949) distribution, while P. sylvestris (TSS = 0.731, 0.785, R²_logloss = 0.585, 0.670, respectively, for potential and realized distribution) and P. nigra (TSS = 0.658, 0.686, R²_logloss = 0.623, 0.664) achieved the worst. Importance of predictor variables differed across species and models, with the green band for summer and the Normalized Difference Vegetation Index (NDVI) for fall for realized distribution and the diffuse irradiation and precipitation of the driest quarter (BIO17) being the most frequent and important for potential distribution. On average, fine-resolution models outperformed coarse resolution models (250 m) for realized distribution (TSS = +6.5%, R²_logloss = +7.5%). The framework shows how combining continuous and consistent Earth Observation time series data with state of the art machine learning can be used to derive dynamic distribution maps. The produced predictions can be used to quantify temporal trends of potential forest degradation and species composition change.

Climate change causes more extreme droughts and heat waves in Central Europe, affecting vegetative fuels and altering the local fire regime. Wildfire is projected to expand into the temperate zone, a region traditionally not concerned by fire. To mitigate this new threat, local forest management will require spatial fire hazard information. We present a holistic and comprehensible workflow for quantifying fuels and wildfire hazard through fire spread simulations. Surface and canopy fuels characteristics were sampled in a small managed temperate forest in Northern Germany. Custom fuel models were created for each dominant species (Pinus sylvestris, Fagus sylvatica, and Quercus rubra). Canopy cover, canopy height, and crown base height were directly derived from airborne LiDAR point clouds. Surface fuel types and crown bulk density (CBD) were predicted using random forest and ridge regression, respectively. Modeling was supported by 119 predictors extracted from LiDAR, Sentinel-1, and Sentinel-2 data. We simulated fire spread from random ignitions, considering eight environmental scenarios to calculate fire behavior and hazard. Fuel type classification scored an overall accuracy of 0.971 (Kappa = 0.967), whereas CBD regression performed notably weaker (RMSE = 0.069; R² = 0.73). Higher fire hazard was identified for strong winds, low fuel moisture, and on slopes. Fires burned fastest and most frequently on slopes in large homogeneous pine stands. These should be the focus of preventive management actions.

Forest fuels are essential for wildfire behavior modeling and risk assessments but difficult to quantify accurately. An increase in fire frequency in recent years, particularly in regions traditionally not prone to fire, such as central Europe, has increased demands for large-scale remote sensing fuel information. This study develops a methodology for mapping canopy fuels over large areas (Germany) at high spatial resolution, exclusively relying on open remote sensing data. We propose a two-step approach where we first use measurements from NASA’s Global Ecosystem Dynamics Investigation (GEDI) instrument to estimate canopy fuel variables at the footprint level, before predicting high-resolution raster maps. Instead of using field measurements, we generate (GEDI-) footprint-level estimates for canopy (Base) height (CH, CBH), cover (CC), bulk density (CBD), and fuel load (CFL) by segmenting airborne Light Detection and Ranging point clouds and processing tree-level metrics with allometric crown biomass models. To predict footprint-level canopy fuels we fit and tune Random Forest models, which are cross-validated using k-fold nearest neighbor distance matching. Predictions at >1.6 M GEDI footprints and biophysical raster covariates are combined with a universal Kriging method to produce countrywide maps at 20 m resolution. Agreement (RMSE/R ²) with validation data (from the same population) was strong for footprint-level predictions and moderate for map predictions. A validation with estimates based on National Forest Inventory data revealed low to modest agreement. Better accuracy was achieved for variables related to height (CH, CBH) rather than to cover or biomass (CBD, CFL). Error analysis pointed towards a mixture of biases in model predictions and validation data, as well as underestimation of model prediction standard errors. Contributing factors may be simplification through allometric equations and spatial and temporal mismatch of data inputs. The proposed workflow has the potential to support regions where wildfire is an emerging issue, and fuel and field information is scarce or unavailable.

Wildfires pose a significant threat to ecosystems, human life, and infrastructure, particularly in South America, where diverse climatic and environmental factors contribute to their occurrence. Climate change has exacerbated extreme weather conditions such as intense heat and drought, leading to a global increase in the frequency and intensity of wildfires. Countries like Brazil have experienced significant rises in wildfire damage, highlighting the urgent need for predictive models that accurately assess future wildfire risks to mitigate their impact effectively. This thesis addresses this need by developing a wildfire risk prediction system leveraging deep learning methods and remote sensing data.Using Earth Observation (EO) APIs, the system avoids downloading and storing vast amounts of satellite imagery, enabling efficient data acquisition and preprocessing. The study focuses on key variables that influence wildfire activity, including dynamic variables such as Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), radiation, Leaf Area Index (LAI), evapotranspiration (ET), wind speed, and temperature, as well as static variables like land cover, Digital Elevation Model (DEM), and population density. The system is designed to predict wildfire risk for the next day and up to eight days, offering a robust tool for proactive wildfire management.Given the stochastic and nonlinear nature of wildfire phenomena, this research employs advanced deep learning techniques, including Random Forests (RF), Long Short-Term Memory networks (LSTM), and Convolutional LSTM (ConvLSTM) models, to predict wildfire risk in near real-time. Active fire data from MODIS products, along with their burn dates, serve as the basis for training datasets. Non-fire points are generated by mapping the land cover distribution of fire points, ensuring balanced datasets for model training. Variables are extracted and classified into dynamic and static categories to capture both temporal variability and fixed geographical characteristics.The objectives of this research are threefold: (1) to investigate existing remote sensing-based wildfire management methodologies and identify enhancements through the integration of data cubes and deep learning; (2) to develop a scalable platform for efficient data acquisition, preprocessing, and risk prediction using deep learning algorithms; and (3) to evaluate the system’s accuracy, efficiency, and scalability with real-world datasets and disaster scenarios.Preliminary results highlight the effectiveness of integrating remote sensing data with deep learning models for wildfire risk prediction. Dynamic variables such as EVI, LST, and NDVI, along with human influence factors like Global Human Modification Index (gHM), emerged as key predictors, demonstrating the interplay of environmental and anthropogenic drivers in wildfire occurrences. Seasonal analysis from 2021 to 2024 revealed a strong correlation between fire activity, elevated temperatures, and declining vegetation indices from November to April. The Random Forest model achieved 83% accuracy, while the LSTM model showed promise with 75% accuracy, emphasizing the potential of both static and temporal data. These findings lay a robust foundation for enhancing wildfire risk management through advanced machine-learning approaches.

Central European forests face challenges with climate changing much faster than they can adapt. Extremely hot and dry summers like in 2018 deprive forests of soil moisture, leaving them with low ground water levels. While individuals with deep and well-established root systems survive, young individuals and shallow-rooted species perish.In southern Germany, die-off of single trees or small groups got noticeable recently. Such effects of harsher conditions rarely occur over large areas, but more in a spotted, irregular manner. This makes the phenomenon difficult to detect and to estimate its extent. The share of trees lately deteriorated may be larger than expected and represent a considerable portion of forests. Therefore, we see the great need for monitoring. Remote sensing data is suitable to examine inaccessible areas at a large scale. To quantify mortality of individual trees among a majority of vital ones, sensor platforms and respective data have to fulfill certain criteria regarding spatial, temporal and spectral resolution. Dead trees can be distinguished from others due to discoloration and defoliation. This change in appearance affects the spectral response, even in pixels larger than the tree’s extent.This study aims at recommending a suitable spatial scale for space-borne multispectral imagery products to achieve this task. We evaluate commercial and free remote sensing data products and their ability to estimate fractional cover of dead vegetation. Satellite data employed in this study comes from Landsat 8 (30 m), Sentinel-2 (10 m), RapidEye (6.5 m) and PlanetScope (3 m). Classification performance is tested against high-resolution multispectral aerial imagery (17 cm) acquired with a Micasense RedEdge-M camera.High-resolution Micasense images are capable of detecting single dead trees, even after downgrading the resolution from 17 cm to 3 m. For all data products tested, fraction of dead trees per pixel did not differ significantly among land cover types (dead vegetation, vital vegetation, pavement, open soil). This indicates that individual dead trees may not be detectable in vital forest stands. The finding even seems to be valid for a resolution of 3 m (PlanetScope), which is identical to the downgraded Micasense data. In the near future the detection of this phenomenon might profit from technical developments towards even higher spatial detail of space-borne sensors. Alternatively, high resolution images from aerial campaigns, manned or unmanned, could bridge this gap when flight time and spatial coverage are increased significantly and facilitating policies are in place.

This paper describes a data-driven framework based on spatiotemporal machine learning to produce distribution maps for 16 tree species (Abies alba Mill., Castanea sativa Mill., Corylus avellana L., Fagus sylvatica L., Olea europaea L., Picea abies L. H. Karst., Pinus halepensis Mill., Pinus nigra J. F. Arnold, Pinus pinea L., Pinus sylvestris L., Prunus avium L., Quercus cerris L., Quercus ilex L., Quercus robur L., Quercus suber L. and Salix caprea L.) at high spatial resolution (30 m). Tree occurrence data for a total of 3 million of points was used to train different algorithms: random forest, gradient-boosted trees, generalized linear models, k-nearest neighbors, CART and an artificial neural network. A stack of 305 coarse and high resolution covariates representing spectral reflectance, different biophysical conditions and biotic competition was used as predictors for realized distributions, while potential distribution was modelled with environmental predictors only. Logloss and computing time were used to select the three best algorithms to tune and train an ensemble model based on stacking with a logistic regressor as a meta-learner. An ensemble model was trained for each species: probability and model uncertainty maps of realized distribution were produced for each species using a time window of 4 years for a total of 6 distribution maps per species, while for potential distributions only one map per species was produced. Results of spatial cross validation show that the ensemble model consistently outperformed or performed as good as the best individual model in both potential and realized distribution tasks, with potential distribution models achieving higher predictive performances (TSS = 0.898, R2 logloss = 0.857) than realized distribution ones on average (TSS = 0.874, R 2 logloss = 0.839). Ensemble models for Q. suber achieved the best performances in both potential (TSS = 0.968, R2 logloss = 0.952) and realized (TSS = 0.959, R2 logloss = 0.949) distribution, while P. sylvestris (TSS = 0.731, 0.785, R2 logloss = 0.585, 0.670, respectively, for potential and realized distribution) and P. nigra (TSS = 0.658, 0.686, R2 logloss = 0.623, 0.664) achieved the worst. Importance of predictor variables differed across species and models, with the green band for summer and the Normalized Difference Vegetation Index (NDVI) for fall for realized distribution and the diffuse irradiation and precipitation of the driest quarter (BIO17) being the most frequent and important for potential distribution. On average, fine-resolution models outperformed coarse resolution models (250 m) for realized distribution (TSS = +6.5%, R2 logloss = +7.5%). The framework shows how combining continuous and consistent Earth Observation time series data with state of the art machine learning can be used to derive dynamic distribution maps. The produced predictions can be used to quantify temporal trends of potential forest degradation and species composition change.