Ponderosa pine (Pinus ponderosa) is a foundational species in dry-forest landscapes across western North America. Ongoing warming and aridification intensify climatic stress across its native range. Prior research on ponderosa pine health is regional in scope and focuses on single disturbances, leaving the combined influence of climate, stand structure, and biotic agents on dieback and mortality poorly understood. To address these knowledge gaps, we surveyed sites across California, Colorado, Idaho, Montana, and Wyoming during 2024 and 2025. At each site, we characterized stand conditions, assessed tree health, and collected tissue samples for pathogen identification. Long-term changes in climate were also recorded at each site to test associations between tree health and biotic and abiotic factors. Recent climate trends included reduced precipitation, elevated mean temperatures, and changes in vapor pressure deficit, indicating the potential for climatic stress on trees. Of the 324 trees surveyed, 25.8% had detections of pathogens and symptoms consistent with pathogen infection. Pathogens were composed of two functional groups: (1) needle and branch pathogens, including the widely distributed endophyte, Sydowia polyspora, primarily associated with trees of moderately poor health status, and (2) vascular stain fungi that were strongly linked to recent mortality. Only S. polyspora showed significant relationships with abiotic conditions in generalized linear models, while non-metric multidimensional scaling revealed clear geographic structuring of pathogen assemblages. In summary, our study suggests that climatic stress is associated with increases in host vulnerability to biotic agents, while fungal pathogens exhibit distinct associations with moderate decline and recent mortality.
Ensuring consistent quality in friction stir welding (FSW) remains challenging because defect formation and mechanical performance arise from complex thermo-mechanical interactions. This study presents an interpretable convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) framework for patch-level surface-defect detection and weld-level tensile-strength classification in ultrasonic vibration-assisted FSW AA2060-T8E30 joints. Surface images from 54 welds were divided into six ordered patches to preserve contextual information along the welding direction. DenseNet121 and VGG16 were evaluated as standalone CNNs and hybrid CNN-BiLSTM models. DenseNet121-BiLSTM achieved the best defect-detection performance, with 98.15% fixed-test accuracy and 95.54% mean leave-one-weld-out accuracy, compared with 92.59% and 90.23% for VGG16-BiLSTM. The DenseNet121-BiLSTM model was then applied to three-class tensile-strength classification relative to base-material ultimate tensile strength, achieving 88.89% fixed-test accuracy and 77.78% +/- 7.41% under repeated stratified grouped cross-validation. Gradient-weighted Class Activation Mapping (Grad-CAM) visualisations identified surface regions associated with the predictions, while a graphical user interface integrated image input, classification, confidence reporting, and interpretation. The results demonstrate the feasibility of using standard weld-surface images for both visible defect detection and preliminary tensile-strength categorisation to support post-weld decision-making.
Timber harvesting operations are restricted on steep slopes to limit soil damage, erosion and water quality perturbation. Tethered logging, an approach that uses winch-assisted equipment to operate on steep slopes may provide an option for expanding the pace and scale of fuel reduction and salvage operations. This is the first study conducted in the United States to measure hillslope erosion following tethered logging. We used sediment fences to quantify hillslope-scale erosion during three post-harvest years in beetle-killed stands where slopes generally exceeded 40%. Tethered machinery is intended to travel upon logging debris, but foliage and branches were scarce on the long-dead trees, and bare soil patches were widespread along the tether roads. Thus, we compared soil erosion on slash-covered and bare tether roads and on harvested hillslopes between the roads. Erosion on slash-covered tether roads that comprised 73% of tether road surfaces was ≤0.6 Mg/ha/yr. In contrast, erosion exceeded 10 Mg/ha after high intensity rainstorms on some bare tether roads and annual erosion reached 12 Mg/ha/yr. Hillslope scale erosion measured using sediment fences increased with rainfall intensity, as it did during plot-scale simulated rainstorms. The highest erosion we observed on bare tether roads was less than levels typical of unmaintained forest roads; owing to their limited extent, erosion from the bare patches was mitigated by logging debris on the tether roads and harvested hillslopes. Though additional research is needed, this case study suggests that tethered logging has potential to jointly address fuel reduction and watershed protection objectives.
Rapid and accurate building damage assessment after tornadoes is critical for emergency response and recovery, yet automated methods struggle with the visual complexity of tornado-induced wreckage, primarily due to severe domain shift and extreme class imbalance. We introduce ASTRA, a large-scale framework evaluating 79 open-source models (67 CNNs and 12 Transformers-based architectures) across 2,420 experiments on an expertly curated Quad-State Tornado Damage (QSTD) benchmark of 5,517 street-view images annotated under the IN-CORE engineering taxonomy. We find that achieving strong performance hinges on the interaction among architecture, optimization, and pre-training source, not on architecture alone. Attention-based models were acutely optimization-sensitive: although bottom-tier under the default configuration, they recovered by + 25 to + 38 Macro-F1-points once optimization was corrected, either by switching from Adam to SGD or, equivalently, lowering the learning rate. A controlled ablation attributes this swing primarily to the learning rate, with a smaller optimizer effect specific to Vision Transformers; CNNs showed no comparable reversal. Furthermore, we uncover a counterintuitive behavior in which scene-centric Places365 pre-training did not surpass object-centric ImageNet after optimization, despite its closer semantic proximity to building and streetscape scenes and its competitive zero-shot performance. In a strict zero-shot test on the held-out Tuscaloosa-Moore Tornado Damage dataset of 2,393 images, naive baselines failed to transfer, while QSTD-optimized models achieved substantially better performance. The champion model, ConvNeXt-Base, achieved a + 34.6 Macro-F1-point gain over its baseline, despite temporal, sensor, and geographic shifts. Consequently, ASTRA offers practical optimization guidelines and a rigorously validated, open-source foundation for automated tornado-damage assessment. Project Website: https://crumeike.github.io/astra-page.
