Bei Betriebsinventuren sowie der Forsteinrichtung werden in langen zeitlichen Abständen (circa alle 10 Jahre) mittels Schätzungen und Hochrechnungen Daten über die Waldstruktur erhoben. Diese Daten sind eine wichtige Grundlage für die Waldbewirtschaftung, das Monitoring der Waldentwicklung und der Diversität des Waldes. Die bei Betriebsinventuren verwendeten Aufnahmetechniken sind sehr zeitaufwändig und können nur stichprobenartig erhoben werden. Das Pilot-Projekt Five3D zielt darauf ab, mit neuen, innovativen, fernerkundungsbasierten Technologien unter Einsatz von Laser- und Multispektral-Sensoren die Datenaufnahme für das Monitoring der Waldstruktur zu verbessern.

Über mehrere Hektar Waldfläche (Nationalpark Bayerischer Wald, DBU Naturerbe, Stiftung Schönau) wurden hochaufgelöste Laserdaten und multispektrale Daten mittels einer DJI 600 Pro Drohne erfasst. Der Lidarsensor Lightweight Airborne Profiler (LAP) des Fraunhofer IPM integriert einen Laser mit einer Wellenlänge von 905 nm und zwei RGB-Kameras mit einer Auflösung von 4112 x 2176 Pixel. Die Flughöhe betrug bei allen Befliegungen ca. 80 m. Bei einer Querüberlappung von mindestens 50% ergab sich eine durchschnittliche Punktdichte von 150 Punkten/m². Zusätzlich wurden in zwei Teilgebieten eine Multispektralkamera eingesetzt. Diese besteht aus einem RGB-Modul mit 412 x 3008 Pixel und zwei monochromen Modulen mit jeweils 2164 x 2056 Pixel und den Wellenlängen 725 nm und (Red Edge) und 850 nm (NIR).

Zur Segmentierung von Einzelbäumen wurde das neuronale Netzwerk Mask R-CNN verwendet, das mit Multispektralbildern (RGB, CIR) und aus den Laserdaten abgeleiteten Bildern (Durchdringungsmetriken, CHM) trainiert und validiert wurde. Für die Genauigkeitsanalysen standen hochpräzise Baumpositionen und manuell aufbereitete Baumsegmente zur Verfügung. Zum Vergleich wurden vier Baseline Methoden herangezogen.

Die Ergebnisse zeigen deutlich die Überlegenheit des neuen KI-basierten Segmentierungsverfahren im Vergleich zu Standardmethoden. Im Detail erreichten die mit den CHM-basierten Bildern trainierten Modelle bei der Auswertung mit den Referenzdaten der Teilflächen der DBU Naturerbe F1-scores von bis zu 66%. Im Vergleich dazu erreichten die Baseline Methoden ein um ca. 54% F1-score schlechteres Ergebnis. Ähnliche Genauigkeitssteigerungen ließen sich in den Referenzflächen der Stiftung Schönau beobachten. Die trainierten Modelle erreichten F1-scores von bis zu 76% und sind den Baseline Methoden um ca. 47% überlegen. In den Referenzflächen des Nationalparks Bayerischer Wald erreichten die Modelle F1-scores von bis zu 64% und sind den Baseline Methoden um bis zu 39% überlegen.

Zusätzlich wurden mit Hilfe der Baumsegmente aus den Laserpunktwolken Baumparameter (Baumhöhe, Kronengrundfläche, Kronenansatzhöhe, Kronenvolumen) berechnet. Mittels im Feld aufgenommener Referenzdaten (Baumart, BHD, und Kronenansatzhöhe) wurden für repräsentative Probekreise die Biomasse und das Baumvolumen der Referenzbäume bestimmt. Diese, für eine Inventur wichtigen Parameter wurden schließlich - gestützt auf den Baumparametern für alle Bäume der beflogenen Gebiete - mit Hilfe eines Generalized Additive Model (GAM) unter der Annahme einer Gamma-Verteilung modelliert. Die Modelle wurden mit den Referenzdaten der einzelnen Flächen getrennt trainiert und erreichten eine Modellgüte R² von 0,612 bis zu 0,876.

We present an evaluation of different deep learning and machine learning approaches for tree health classification in the Black Forest, the Harz Mountains and the Göttinger Forest on a unique, highly accurate tree-level dataset. The multispectral UAV data were collected from eight forest plots with diverse tree species, mostly conifers. As ground truth data (GTD), nearly 1,500 tree polygons with related attribute information on the health status of the trees was used. This data was collected during extensive fieldwork using a mobile application and subsequent single tree segmentation. Extensive pre-processing included normalization, NDVI calculations, data augmentation to deal with the underrepresented classes and splitting the data into training, validation and test sets. We conducted several experiments using classical machine learning (Random Forest) as well as different Convolutional Neural Networks (CNNs): ResNet50, ResNet101, VGG16 and Inception-v3 on different datasets and classes to evaluate the potential of these algorithms for tree health classification. Our first experiment was a binary classifier of healthy and damaged trees, not considering the degree of damage or tree species. Best results of 0.99 test accuracy and a F1 score of 0.99 were obtained with ResNet50 on RGBI images while VGG16 performed worst with an F1 score of only 0.78. In a second experiment, we also distinguish between coniferous and deciduous trees. F1 scores ranged from 0.62 to 0.99 and with highest results obtained using ResNet101 on NDVIre images. Finally, in a third experiment, we aimed at evaluating the degree of damage: healthy, slightly damaged, and medium or heavily damaged trees. Again, ResNet101 performed best, this time on RGBI images with a test accuracy of 0.98 and an average F1 score of 0.97. These results highlight the potential of CNNs on high-resolution multispectral UAV data for early detection of damaged trees when good training data is available.

Large-scale area-wide mapping of standing deadwood can contribute widely to informed forest management and habitat modeling of various forest species. Numerous automated remote sensing based methods have been developed for this purpose. In this study, we evaluate the effect of increased spatial resolution and the influence of alternated resampling of the input data (orthophotos and Vegetation Height Models (VHMs)) on the accuracy of a two-stage deadwood mapping approach based on a random forest (RF) classification combined with an uncertainty model (UNC) (Zielewska-Buettner et. al. 2020).

The baseline algorithm (Zielewska-Buettner et. al, 2020) and reference in this study was based on publicly available data: orthophotos of 0.5m spatial resolution and VHM of 1m spatial resolution and highest point value selection per pixel. In 2016, the input data was derived from aerial photographs (AP) of 0.2m resolution and 60/30% end/side overlap regardless the camera type. Between 2016 and 2019, the AP data acquisition parameters changed from 60/30% to 80/30% end/side overlap. Subsequently, the spatial resolution of orthophotos changed from 0.5m to 0.2m. Based on data from 2019, we tested two different settings of orthophotos (with spatial resolution of 0.2m and resampled to 0.5m) combined with four settings of VHMs. We calculated VHMs with spatial resolutions of 1m and 0.5m using the highest (VHM-H) and the lowest (VHM-L) point per pixel, hypothesizing that the second resampling method would provide a better bare ground detection and thus less misclassifications of bare ground as deadwood. All derived deadwood maps (RF and UNC) were validated based on visual assessment using the same stratified random sample of pixels of all model classes. Model performance was assessed using confusion matrix and associated measures: i.e. user’s accuracy (UA) and producer’s accuracy (PA) for deadwood and overall accuracy (OA) and Kappa for all model classes.

Input data with higher resolution (2019) delivered in all cases better results than the lower resolution baseline data (2016). In addition, VHM-L always performed better than the standard VHM-H. The best RF results were obtained with the highest tested resolution of orthophoto (0.2m) and VHM-L (0.5m), delivering a deadwood UA=0.75 and PA=0.87 and a mapping OA=0.8 and Kappa=0.73. The best UNC model included a combination of orthophoto and VHM-L, both with a resolution of 0.5m: deadwood UA=0.82, PA=0.89 and mapping OA=0.82 and Kappa=0.75. In comparison, the results in 2016 were for RF: UA=0.60, PA=0.82, OA=0.70 and Kappa=0.60 and for the UNC: UA=0.74, PA=0.8, OA=0.74 and Kappa = 0.65. Two-stage deadwood mapping with UNC outperformed stand-alone RF models in all cases, with the most significant increase in mapping accuracy observed for the datasets including an orthophoto of 0.5m resolution, probably due to the exact matching in resolution with the settings of the fixed UNC model. The results highlight the importance of a thorough selection of spectral and structural input data settings in line with the mapping goal.

In light of the increasing risk of global flooding and subsequent effects on both infrastructure and emergency response, this study analyses the crucial role of remotely sensed imagery in assisting organisations such as UNICEF during critical situations. Employing unmanned aerial vehicles (UAVs) to frequently capture these images, the research emphasises the importance of timely assessments in enabling lifesaving decision making. To optimise resource allocation and streamline the assessment process for on-site specialists, the implementation of computer vision assistance becomes a crucial enhancement. This allows for more efficient evaluation of visuals, contributing to improved response strategies.
This study investigates refinement of water detection methods within high-resolution airborne imagery, addressing the escalating challenges associated with flooding events. Using a semi-supervised approach that combines human expertise with machine learning, this study highlights the importance of integrating the Near-Infrared (NIR) channel for accurate labelling. Focused on the Blessem-Erftstadt flood event, this iterative methodology aims to enhance both the accuracy and complexity of water delineation significantly. Deep learning is used for training segmentation models and establishing an algorithmic baseline, which complements this process. The study's findings show potential improvements to disaster response strategies and increased situational awareness in similar flood scenarios. This can strengthen future response efforts.