Conference Agenda

This study presents a deep learning framework integrating Sentinel‑2, Sentinel‑1, and GEDI LiDAR to map Ethiopia’s canopy height at 10‑m resolution from 2019–2024. A shift‑aware loss function was employed to correct geolocation errors inherent in GEDI L2A footprints, and height‑weighted penalties addressed systematic underestimation in tall forests. Results show a national net gain of 23,537 km² in tree cover >8 m, reversing long‑standing deforestation trends. Gains concentrated in low‑to‑mid canopy strata (<20 m), strongly associated with major restoration interventions including the Green Legacy Initiative (GLI), REDD+, and the Sustainable Land Management Program (SLMP). Losses persist in western and southeastern highlands, driven by agricultural expansion, wildfires, infrastructure development, and large‑scale agricultural investments. This work demonstrates the operational value of multi‑sensor deep learning for near‑real‑time monitoring of restoration outcomes in data‑scarce regions.

Accurate tree-level analysis in forests via LiDAR scanning is essential for biomass estimation, canopy structure assessment, and carbon monitoring, yet remains constrained by the scarcity of large-scale annotated LiDAR datasets and the high cost of manual annotation. To address this, we present a novel approach that integrates 3D tree models with UAV-borne LiDAR simulation to generate synthetic forest point clouds with comprehensive annotations. Our approach generates diverse woodland, open, and closed forest structures, producing Synthetic Forest, a benchmark datasets of three 1 ha scenes containing 38–47 million points each, with densities of 3300–3860 points/m² and average spacing of 2 cm. Each scene contains between 70 and 216 individual trees, along with understory vegetation, deadwood, stumps, rocks, and bushes, all automatically annotated with semantic classification IDs, instance IDs, and tree IDs for volume estimation. The proposed pipeline provides automated, error-free ground truth for leaf-wood classification, instance segmentation, and wood volume estimation. We provide a guideline for generating forest plots and utilizing the datasets for diverse forestry tasks. By eliminating the need for costly field data collection, our pipeline offers scalable, customizable synthetic datasets that accelerate forest inventory. The Synthetic Forest dataset is publicly released via Zenodo (DOI: 10.5281/zenodo.17568131), enabling reproducible research and supporting further developments in forest monitoring and management.

Accurate mapping of tree species is essential for forest monitoring, biodiversity assessment, and ecological applications. Very high-resolution UAV imagery provides detailed structural and spectral information, but species-level segmentation remains challenging due to limited annotated data, complex crown geometries, and strong visual similarity among taxa. Recent Remote Sensing Foundation Models (RSFMs) offer new possibilities by providing transferable representations learned from large, multimodal geospatial datasets.

This contribution introduces a two-phase framework that combines foundation model initialization with multi-temporal UAV imagery to enhance fine-grained forest segmentation. In Phase 1, a DeepLabv3+ network is initialized using FoMo-Net, a ViT-based RSFM pre-trained on the multi-scale FoMo-Bench benchmark. This initialization enables strong generalization from heterogeneous global forest datasets to very high-resolution UAV scenes. In Phase 2, phenological information is integrated by fusing May and September UAV acquisitions through temporal difference composites and pseudo-label refinement, allowing the model to resolve species-specific seasonal patterns.

Experiments on the Québec Trees Dataset, covering 14 species at 0.02 m GSD, demonstrate substantial performance gains. Foundation model initialization improves overall accuracy from 52.79% to 71.21%, while incorporating multi-temporal cues further increases accuracy to 78.21%. The results highlight the complementary roles of structural priors learned by RSFMs and phenological information captured by UAV time series for detailed forest species mapping.

Accurate mapping of forest types and vegetation characteristics is essential for monitoring biodiversity and forest dynamics. Traditional Deep Learning (DL) models trained on Sentinel-2 time series achieve high performance, but require extensive preprocessing and sensor-related fine-tuning. In this study, we evaluate the recently introduced AlphaEarth Foundations (AEF) embeddings, which is global, multi-modal feature representation of the earths surface, for forest mapping in Italy. We compare a) a Multi-Layer Perceptron trained on AEF, b) a Time-Series Transformer trained on Sentinel-2 annual time series and CHELSA climate data, and c) a Cross-Attention fusion model combining both feature sets. Using 5-fold cross-validation in a regression and a classifaction task on two datasets (evergreen broad-leaved tree cover ETC, forest vegetation type FVT) we find that the combined model consistently outperforms the single-source approaches (RMSE = 0.161, Accuracy = 0.757). AEF-based models achieve comparable accuracy to the Sentinel-2-based model while reducing extensive time series preprocessing and training time by an order of magnitude. Feature attribution using integrated gradients reveals that AEF provides stable baseline representations, while Sentinel-2 inputs add phenology-related detail. The results show, that integrating generalized embeddings with specialized spectral-temporal features improves predictive performance for forest mapping.

In Japan, many urban park trees are becoming large and aged, increasing the risk of structural failures caused by extreme weather events and biological deterioration. Effective management therefore requires reliable risk assessment, for which accurate tree species identification is one of the fundamental prerequisites. However, species identification still depends heavily on visual assessment by skilled professionals, posing challenges in efficiency and objectivity. This problem is particularly significant for broad-leaved trees, which exhibit high species diversity and morphological variability. In addition, labor shortages have intensified the demand for automated and reliable classification techniques. This study proposes a high-accuracy classification method for broad-leaved tree species using ground-level images captured with a commercially available RGB camera and deep learning. The proposed method extracts small local patches that capture species-specific visual features, such as leaf shape and bark texture, commonly used by professional arborists for species identification. These local features are evaluated individually using deep learning models, allowing fine-scale visual characteristics to be effectively utilized for classification. To address variability in outdoor imaging conditions, including illumination changes, shadows cast by branches and leaves, and moss attachment, multiple patches are classified independently and the results are integrated through majority voting, improving classification robustness. Experiments were conducted on seven tree species commonly found in Japanese urban parks: cherry, ginkgo, zelkova, konara oak, sawtooth oak, plane tree, and flowering dogwood. The results demonstrate that the proposed method achieves a maximum classification accuracy of approximately 95% under real-world conditions, demonstrating its effectiveness for practical urban tree management.