Terrestrial laser scanning (TLS) enables non-destructive estimation of aboveground biomass (AGB) through direct volumetric modelling and allometric model calibration. However, manual individual tree segmentation of TLS plot point clouds creates a processing bottleneck that limits scalability. We evaluated whether automated tree segmentation using RayExtract can replace manual segmentation to provide plot-level volume estimates aimed at satellite calibration/validation data. We evaluated direct volume estimates and allometric models calibrated using RayExtract tree predictions across seven TLS plots spanning various forest types, ranging from mountainous conifer to tropical rainforests. For direct volume modelling at 0.25 ha resolution, automated segmentation introduced a systematic positive bias of 20.5% (±5.5) relative to manual segmentation. Filtering using outlier detection reduced this to 6.9% (±4.3), while a filtering approach based on visual verification reduced bias to 4.2% (±2.7) on average. After visual verification, which requires only 0.7% of the active processing time compared to manual segmentation, RayExtract produced direct volume estimates with uncertainties comparable to other well-established AGB error sources across all forest types. For allometric model fitting, RayExtract-derived models substantially increased uncertainty and bias. Visual verification generally improved metrics, resulting in uncertainty and bias increases of 9 and 15 percentage points (pp) on V ∼ D models respectively, and 5 and 19 pp on V ∼ HCD models relative to manual segmentation-based models. Bias and uncertainty of automated segmentation-based models were largest in tropical forests, and targeted manual segmentation or correction of a set of representative trees likely remains necessary for local allometric model calibration.
Accurate estimation of maize residue cover (MRC) at field to regional scales is essential for evaluating conservation tillage practices and quantifying soil carbon sequestration potential. However, optical remote sensing-based MRC retrieval is strongly affected by surface moisture variability, which alters spectral responses and exacerbates the spectral confusion between maize residues and bare soil, while single-source data lack sufficient discriminatory power to resolve this complexity. To address these limitations, we propose an integrated MRC estimation framework that combines moisture-stratified correction, multi-source feature fusion, and explainable ensemble learning. First, surface moisture heterogeneity is characterized using fuzzy C-means (FCM) clustering, enabling the partitioning of the landscape into moisture-consistent zones. Within each zone, the Kubelka-Munk (KM) radiative transfer model is employed to explicitly correct moisture-induced absorption effects in Sentinel-2 reflectance, thereby restoring the intrinsic residue spectral properties. Second, Sentinel-2 reflectance, Sentinel-1 SAR backscatter, and texture features are jointly integrated to enhance residue-soil separability. Finally, a Stacking ensemble model is constructed to fuse multi-source features, and SHAP analysis is used to interpret variable contributions and reveal the mechanisms underpinning model performance. Results demonstrate that the proposed FCM + KM framework substantially improves MRC estimation accuracy across heterogeneous moisture conditions, increasing R2 from 0.46 to 0.83 and reducing RMSE from 0.14 to 0.08. The proposed framework provides a physics-informed and interpretable pathway for improving large-scale residue monitoring, particularly in regions where moisture variability constrains optical-based retrievals.
A comprehensive study has been conducted with satellite ocean color, sea surface temperature (SST), and sea surface salinity (SSS) observations, as well as ocean current and ocean nutrient data with the model simulations from 2018 to 2024, to characterize and quantify physical, optical, biological, and biogeochemical ocean properties, assess the ocean processes and understand the mechanism that drives the ocean ecosystem features in the Georges Bank (GB) region. The impact of GB on SST is less significant except in summer and SSS is not influenced by the GB seafloor topography. The enhanced chlorophyll-a (Chl-a) in the GB region is found to be significantly higher than those nearby open ocean regions and the Gulf of Maine with notable spring and autumn phytoplankton blooms. Of the major nutrients, the elevated Chl-a matches well with the higher iron (Fe) level both spatially and temporally, while the other nutrients, i.e., nitrate, phosphate, and silicate are negatively correlated. The linkage between the enhanced Chl-a and increased Fe in the GB region shows that Fe is the key nutrient driver that fuels the high biological activity over the GB region, and depletion of nitrate and phosphate hampers further increase of Chl-a during the spring bloom. Unlike most climate-driven biological activities, e.g., Biological Indian Ocean Dipole (BIOD) driven by nutrients from deeper ocean, the nutrient model shows the enhanced iron at the sea surface is caused by the mixing and re-suspension of iron-rich sediment from the GB sea floor. As a case study, this study sheds the light to address the high productivity and marine biodiversity of the ocean ecosystems in these ocean banks and helps to understand the physical, nutrient, and biological processes that drive these ocean bank ecosystems.
Rapid building damage mapping from very-high-resolution post-event imagery is essential for emergency response, yet supervised models often lose reliability when annotations for a newly affected region are scarce. Existing segmentation approaches also make limited use of explicit information about building locations and their immediate surroundings. This paper presents STRATA, a structure-guided transfer architecture that combines a hierarchical SegFormer-B2 encoder, a Building-Context-Guided Attention module, and a lightweight multiscale decoder. Building-footprint-derived priors represent building regions, their surrounding context, and the remaining background, while joint optimization with labeled source events supports adaptation to label-scarce target events. Experiments were conducted using post-event imagery and harmonized four-class annotations from 11 disasters in the Extensible Building Damage benchmark, with target-label budgets ranging from 10 to 500 samples per class. Under the most restrictive budget of 10 samples per class, STRATA exceeded the strongest baseline by 9.61 percentage points in F1-score and 7.87 points in rare-class mean intersection over union. Ablation analysis showed that the hierarchical transformer encoder provided the largest contribution to damage-severity discrimination, while structural guidance offered complementary improvements, particularly for destroyed buildings. Source-event supervision was most beneficial at the lowest annotation budgets, with its contribution diminishing as the target training set became more representative. Event-level evaluation further showed that transfer effectiveness varied with the correspondence between source and target damage characteristics. At the same time, competitive performance across earthquake, hurricane, and tornado events indicated that the learned representation was not confined to a single hazard type. These findings position STRATA as a practical option for early post-disaster damage screening, when reliable target-event annotations are still limited. The implementation is publicly available at https://github.com/EhAI4EO/STRATA.
