Accurate and dynamic monitoring of wheat phenotypes is essential for breeding decision-making and crop management. However, RGB image-based phenotyping still suffers from expensive pixel-level annotation, unstable organ-level segmentation across growth stages, and limited multi-trait extraction under complex field conditions. To address these issues, a high-throughput phenotyping framework (WheatScoper) was proposed, enabling organ-level segmentation and plot-level multi-trait extraction. To reduce annotation cost, a structure-aware geometry-assisted annotation (SAGA) algorithm was developed, yielding an approximately 7.6-fold improvement in annotation efficiency over fully manual annotation. To enable efficient organ-level segmentation, a lightweight semantic segmentation network (WheatScopeNet) was developed by integrating parallel hybrid spatial modeling with cross-scale feature fusion. On the held-out test set from the same site and growing season, WheatScopeNet achieved an mIoU of 0.869 and an mDice of 0.930. Leveraging the segmentation results, an automated system was established to extract 41 multi-view image-derived traits (I-traits) across four core phenotypic dimensions. The extracted I-traits supported the estimation of eight manually measured agronomic traits, with R2 values ranging from 0.477 to 0.697. Notably, the correlations between stay-green-related-traits and yield varied distinctly with viewing-position. Only upper side-view indicators remained significantly correlated with yield, with Side-up final GPAR showing the strongest association, whereas top-view GPAR-derived indicators showed weak associations. Finally, a web-based platform integrating cascaded inference, segmentation visualization, and automatic I-trait extraction was developed. The platform provides an end-to-end solution for field wheat phenotyping and supports breeding decision-making and crop management.
Wildfire serves a critical function in shaping the structure, composition and processes of Earth’s ecosystems. However, anthropogenic climate change over the past century has catalyzed a shift in wildfire regimes, particularly in California, as evidenced by increases in the extent, frequency and severity of fires. The broad social-economic consequences of extreme wildfire events necessitate systematic frameworks that incorporate multi-source observations to support risk assessments and pre-fire interventions. This study presents an ensemble Fire Potential Index (eFPI) with multi-decadal Moderate Resolution Imaging Spectroradiometer (MODIS) and climatological data, along with a comprehensive validation of its accuracy in capturing wildfire susceptibility across Southern California between 2002 and 2024. The eFPI captures the fire susceptibility of both live and dead fuels by accounting for variability in ignitability and fuel moisture, while weighting fuel types according to their fractional coverage at the sub-pixel scale. Spectral mixture analysis was first applied on daily MODIS imagery at 500-m spatial resolution to quantify sub-pixel fractions of green vegetation (GV), non-photosynthetic vegetation (NPV) and substrate. Second, we adopted spatially-contiguous climate datasets and in-situ measurements to estimate the moisture content of live and dead fuels, which were subsequently synthesized with extinction coefficients and associated fractional coverage to produce daily estimates of fire potential. Our results indicate the eFPI, through spatially-explicit integration of sub-pixel fuel characteristics, demonstrated a promising capacity to discriminate between burned and unburned regions across Southern California (χ2 = 3247; AUC = 0.74). Substantial differences in antecedent FPI estimates were also found between fire and no-fire pixels, which facilitate short-term projections of fire risk in the absence of real-time remote sensing and climate data. This paper provides an integrative framework for wildfire risk assessments, and offers a new approach for land management and fire protection agencies for efficient pre-fire planning practices.
Turbidity, a fundamental water quality parameter, profoundly influences aquatic ecosystems by regulating light penetration, primary productivity, and the fate of nutrients and pollutants. Despite prior large-scale optical water-quality studies of U.S. rivers, systematic, multi-decadal analyses integrating Landsat surface reflectance, machine-learning retrieval, and ecoregion-scale driver attribution have remained limited. This study developed and validated a long-term Landsat-based turbidity reconstruction framework for major rivers of the conterminous United States from 1984 to 2025. By integrating harmonized Landsat 5/7/8/9 surface reflectance (Collection 2 Tier 1) with river geometry from the Surface Water and Ocean Topography Mission River Database (SWORD), named after the SWOT satellite mission, and in-situ turbidity measurements from the USGS, we established a robust Random Forest (RF) regression model (validation R2 = 0.95, RMSE = 21.25 NTU) for turbidity estimation, outperforming three other machine-learning algorithms. The RF model was further evaluated using strict spatiotemporal transferability tests, including chronological hold-out, shared-site, new-site, and leave-one-region-out (LORO) spatial cross-validation. Model interpretability was rigorously validated using SHAP (SHapley Additive exPlanations) analysis, confirming that the model learned physically meaningful relationships. Our study revealed distinct spatial patterns in river turbidity, with generally higher values in the Southern Plains (SPL) and Temperate Plains (TPL) ecoregions and lower values in the Western Mountains (WMT) and Xeric (XER) regions. Moreover, we detected a heterogeneous but overall declining trend in turbidity over time, particularly in historically high-turbidity areas. Model diagnostics indicate that predictive performance is higher in mid- to high-turbidity waters, and the reconstructed dataset should be interpreted primarily at regional and long-term scales. To explore the drivers of these changes, we integrated climatic, land cover, and dam data into the analysis. The results show that anthropogenic factors, including dams (∼33% contribution), forest cover (∼27%, reflecting the absence of disturbance), and urban/built-up areas (∼17%), are the primary contributors to turbidity variance at the national scale, far outweighing the influence of climatic variables (each generally < 5%). Further investigation of ten representative dammed rivers indicated that dam construction is associated with significant downstream reductions in turbidity, with the effect typically stronger for taller dams but modulated by local topography. As an exploratory application, framed as an illustrative case study rather than a primary finding, we further developed a turbidity-linked hydro-morphodynamic sensitivity indicator that combines turbidity decline and velocity-related channel proxies. This indicator is intended to provide a preliminary reference for identifying reaches where sediment reduction may coincide with enhanced hydraulic forcing; independent geomorphic validation is required before any hazard-related interpretations can be drawn.
Polarimetric synthetic aperture radar (PolSAR) provides full polarimetric measurements, from which physically meaningful scattering parameters can be derived through target decomposition. Traditional model-driven methods often suffer from rigid power allocation, sensitivity to noise, and limited capability in representing complex scattering scenes. In contrast, deep learning-based approaches can learn intricate scattering patterns directly from data with greater flexibility. Nevertheless, these purely data-driven methods usually model scattering mechanisms as black boxes, resulting in limited physical interpretability and reduced generalization to complex or unseen scenarios. Building on these observations, we investigate the integration of physical priors into deep neural networks and propose PELNet, a physics-embedded end-to-end framework for polarimetric target decomposition. The core innovation of PELNet lies in structurally embedding a differentiable scattering model within the network decoder, establishing a strict correspondence between the latent space and interpretable physical parameters. Specifically, a multi-scale SE-attention encoder is used to extract context-aware polarimetric features, while a lightweight Mixture-of-Experts module enables adaptive modeling of different volume scattering components. Furthermore, a differentiable scattering reconstruction layer embeds the physical scattering model directly into the network decoder, establishing an interpretable mapping between latent representations and physical parameters. A physically guided loss function is further introduced to enhance reconstruction fidelity and spatial consistency. Experiments on multi-band PolSAR datasets across diverse terrain types demonstrate that PELNet achieves superior accuracy and interpretability compared to traditional methods, providing a new perspective for PolSAR image interpretation and physical parameter inversion. The source code and pretrained models are publicly available at https://github.com/shurongz/PhyDecomp.
Accurate and uncertainty-aware retrieval of vegetation biophysical variables is essential for reliable Earth observation-based monitoring. We developed Gaussian Process Regression (GPR) models trained on globally distributed Copernicus Ground-Based Observations for Validation (GBOV) reference data to retrieve leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), and fractional vegetation cover (FVC) from Sentinel-3 SYNERGY (SY_2_SYN) surface reflectance. The models are designed for general applicability, and are illustrated here through large-scale application and evaluation over Europe. Over 3650 GBOV samples upscaled to satellite pixel per fundamental vegetation trait (FVT) supported training of the GPR–SYN-GPR models (2020–2021), whilst over 1700 samples were reserved for validation (2022–2024). The models achieved robust performance for most vegetation types, with R2 and NRMSE over 0.8 and below 13%, respectively, for croplands, while revealing persistent limitations in dense evergreen canopies. Comparison with Copernicus Global Land Service (CGLS) products reveals overall agreement, with biome-dependent differences that guide future refinement. A key feature of the GBOV–SYN–GPR framework is the explicit decomposition of predictive uncertainty into epistemic (model-driven) and aleatoric (observation-driven) components. European-scale mapping for 14–23 July 2024 reveals coherent biogeographical gradients and epistemic uncertainty patterns consistent with expected vegetation dynamics. Overall, epistemic uncertainty dominates and exhibits systematic spatial structures linked to gaps in training representativeness, while aleatoric uncertainty remains generally smaller and is mainly enhanced in mountainous regions and under challenging observation conditions. To contextualize the magnitude of the total uncertainty, integrating both epistemic and aleatoric components, we compared it with GBOV-provided uncertainties and found that GBOV–SYN–GPR total uncertainties were of similar magnitudes to those of the upscaled GBOV land products at three GBOV sites. The proposed framework provides a scalable and traceable approach for generating uncertainty-quantified vegetation trait products from Sentinel-3 and offers a pathway for future optical missions.
