Accurate monitoring of Spartina alterniflora eradication is essential for evaluating coastal wetland restoration and supporting invasive-species management. However, vegetation removal and natural surface variation can produce similar spectral changes in heterogeneous tidal flats, complicating image-based detection. To address this issue, we propose COAST ( Coastal Online Adaptive Spatiotemporal Transformer ), an online ecological-prior-guided bitemporal Transformer for mapping potential S. alterniflora eradication areas in the Jiangsu coastal tidal flats, China. The framework uses geographically weighted discriminant analysis (GWDA) posterior probabilities to represent spatially heterogeneous relationships between S. alterniflora pre-eradication spectral features and removal-related change. These probabilities provide soft supervision for a lightweight online prior encoder, whose output modulates bitemporal features through spatial prior gating. Training uses spatial-block out-of-fold GWDA predictions, and inference uses paired images with the prior generated from the pre-eradication image. Under a spatially separated evaluation design, COAST achieved the highest mean F1-score (95.00 %) and IoU (90.48 %) among the evaluated baselines across three random-seed runs. It also achieved a boundary F1-score of 81.25 %, a 95th-percentile Hausdorff distance (HD95) of 256.04 m, and a changed-area fraction mean absolute error of 0.82 percentage points. Ablation experiments supported the contribution of prior-guided learning relative to the BiT backbone. Evaluation on manually annotated cross-year patches yielded F1-scores of 88.27 % for 2020–2023 and 93.13 % for 2020–2025, supporting temporal transfer within the sampled coastal settings. Local errors included agricultural-background false positives and incomplete target recovery. COAST provides an image-based workflow for repeated regional screening of eradication-related change and the prioritisation of follow-up verification.
Dothistroma needle blight (DNB) is a major foliar disease of Pinus radiata , an important plantation species in several Southern Hemisphere countries. DNB management and resistance breeding depend on subjective, labour-intensive visual severity scoring that is difficult to scale. Prior UAV studies have used broad severity categories and crown-mean reflectance, discarding both the ordinal structure of the field grade and the within-crown spatial pattern of infection. This study evaluated whether UAV multispectral imagery can quantify tree-scale DNB severity while preserving the ordinal 5 % field grading, treating adjacent-class errors as less severe than distant ones. From imagery of two trials in the North Island, New Zealand, 8,136 manually delineated crowns were characterised by 552 candidate features (82 vegetation indices and 10 raw bands, each summarised by six within-crown distributional statistics). After collinearity filtering, 40 features were used to fit Ordinal Forest (OF) and Frequency-Adjusted Borders Ordinal Forest (fabOF) models at three granularities (5, 10, and 19 severity classes). At 19-class resolution, predictions fell within ± 10 points of the true severity for approximately 80 % of crowns. Quadratic weighted kappa rose with granularity, consistent with finer models erring predominantly by small, adjacent-class amounts. fabOF improved ordinal ranking, with markedly less underestimation between 50 and 70 % severity, whereas OF was sharper in the well-sampled moderate range, a calibration-balance trade-off across applications. Within-crown variability and tail-percentile statistics, not crown-mean reflectance, carried the strongest signal (SHAP analysis), consistent with DNB’s patchy progression. UAV multispectral imagery can therefore support fine-grained DNB severity assessment, an automated complement to visual scoring for management and resistance phenotyping, demonstrated here across two juvenile trials.
Dothistroma needle blight (DNB) is a major foliar disease of Pinus radiata, an important plantation species in several Southern Hemisphere countries. DNB management and resistance breeding depend on subjective, labour-intensive visual severity scoring that is difficult to scale. Prior UAV studies have used broad severity categories and crown-mean reflectance, discarding both the ordinal structure of the field grade and the within-crown spatial pattern of infection. This study evaluated whether UAV multispectral imagery can quantify tree-scale DNB severity while preserving the ordinal 5 % field grading, treating adjacent-class errors as less severe than distant ones. From imagery of two trials in the North Island, New Zealand, 8,136 manually delineated crowns were characterised by 552 candidate features (82 vegetation indices and 10 raw bands, each summarised by six within-crown distributional statistics). After collinearity filtering, 40 features were used to fit Ordinal Forest (OF) and Frequency-Adjusted Borders Ordinal Forest (fabOF) models at three granularities (5, 10, and 19 severity classes). At 19-class resolution, predictions fell within ± 10 points of the true severity for approximately 80 % of crowns. Quadratic weighted kappa rose with granularity, consistent with finer models erring predominantly by small, adjacent-class amounts. fabOF improved ordinal ranking, with markedly less underestimation between 50 and 70 % severity, whereas OF was sharper in the well-sampled moderate range, a calibration-balance trade-off across applications. Within-crown variability and tail-percentile statistics, not crown-mean reflectance, carried the strongest signal (SHAP analysis), consistent with DNB’s patchy progression. UAV multispectral imagery can therefore support fine-grained DNB severity assessment, an automated complement to visual scoring for management and resistance phenotyping, demonstrated here across two juvenile trials.
Accurate monitoring of Spartina alterniflora eradication is essential for evaluating coastal wetland restoration and supporting invasive-species management. However, vegetation removal and natural surface variation can produce similar spectral changes in heterogeneous tidal flats, complicating image-based detection. To address this issue, we propose COAST (Coastal Online Adaptive Spatiotemporal Transformer), an online ecological-prior-guided bitemporal Transformer for mapping potential S. alterniflora eradication areas in the Jiangsu coastal tidal flats, China. The framework uses geographically weighted discriminant analysis (GWDA) posterior probabilities to represent spatially heterogeneous relationships between S. alterniflora pre-eradication spectral features and removal-related change. These probabilities provide soft supervision for a lightweight online prior encoder, whose output modulates bitemporal features through spatial prior gating. Training uses spatial-block out-of-fold GWDA predictions, and inference uses paired images with the prior generated from the pre-eradication image. Under a spatially separated evaluation design, COAST achieved the highest mean F1-score (95.00 %) and IoU (90.48 %) among the evaluated baselines across three random-seed runs. It also achieved a boundary F1-score of 81.25 %, a 95th-percentile Hausdorff distance (HD95) of 256.04 m, and a changed-area fraction mean absolute error of 0.82 percentage points. Ablation experiments supported the contribution of prior-guided learning relative to the BiT backbone. Evaluation on manually annotated cross-year patches yielded F1-scores of 88.27 % for 2020–2023 and 93.13 % for 2020–2025, supporting temporal transfer within the sampled coastal settings. Local errors included agricultural-background false positives and incomplete target recovery. COAST provides an image-based workflow for repeated regional screening of eradication-related change and the prioritisation of follow-up verification.
Vegetation Optical Depth (VOD) is a key microwave-derived remote sensing parameter reflecting vegetation water content and supporting ecological and carbon-cycle studies. However, existing VOD products have coarse spatial resolution (∼9–36 km), limiting fine-scale applications. Statistical and machine-learning downscaling methods remain immature, lacking reliability and physical consistency, especially over large heterogeneous regions. To address this research gap, this study proposes the Physics-informed Differencing Downscaling Framework (PhyDiF). PhyDiF integrates a differencing paradigm with radiative transfer model (RTM)-based physical constraints. Rather than directly predicting absolute VOD values, the differencing paradigm predicts the differences in VOD (ΔVOD) induced by spatial resolution differences. The predicted ΔVOD is then combined with the coarse-resolution baseline VOD to reconstruct high-resolution absolute VOD. In addition, an RTM-based physical constraint is incorporated into a joint loss function to improve VOD prediction accuracy while promoting consistency between the predicted VOD and the corresponding brightness temperature (TB) simulated using the RTM. At the 9 km scale, direct validation of the differencing paradigm within PhyDiF and model selection among various baseline models show that U‑Net performs best in predicting ΔVOD, achieving a correlation coefficient with the true values of 0.854, with RMSE and MAE of only 0.016 and 0.011, respectively. Based on this, PhyDiF was constructed by incorporating physical constraints, generating a 3‑day 1 km PhyDiF VOD dataset over China from 2015 to 2021. Comprehensive evaluation shows that the PhyDiF VOD product outperforms the baseline product (MT‑DCA) and traditional interpolation results (Cubic) in spatial details, clearly depicting texture information and VOD spatial heterogeneity that more closely resemble actual land surface features. Correlations with above‑ground biomass (AGB), canopy height (CH), and leaf area index (LAI) are significantly improved compared with existing methods, both across China and in typical local areas. For example, in the low‑VOD area, compared with the Cubic method, the differencing‑paradigm‑based U‑Net and PhyDiF achieve improvements in correlation with AGB, CH, and LAI of 19.13 %/5.83 %/9.48 % and 32.17 %/10.42 %/17.43 %, respectively, while PhyDiF, incorporating physical constraints, outperforms U‑Net by 10.95 %, 4.33 %, and 7.26 %, respectively. In temporal validation, the product accurately exhibited the seasonal cycles and interannual dynamic trends of vegetation phenology, maintaining high consistency with MT‑DCA. Further scale consistency testing confirmed a correlation of 0.9997 with MT‑DCA and a standard deviation of only 0.007. Overall, PhyDiF provides a generalizable and reliable framework for generating high‑resolution VOD, advancing vegetation microwave remote sensing technology.
