Search Results (1,296)

To better analyze the health and welfare of individual dairy cows, this study proposes BR-Tracker, a multi-object tracking method designed for dense monitoring environments to address missed detections, tracking failures, and frequent identity switches. In the object detection stage, a Receptive Field Attention Downsampling (RFADown) module is introduced into the neck network of YOLOv10s to effectively process the locally visible regions of occluded cows by dynamically adjusting the receptive field. An improved Partial Bi-Level Routing Attention (PBRA) module is incorporated into the backbone network to simultaneously extract global and local features, while a Spatial Pyramid Pooling with Efficient Layer Aggregation Network (SPPELAN) module is adopted to enhance multi-scale feature aggregation. In the object tracking stage, an MPDIoU-based matching algorithm is designed to improve matching accuracy and the reliability of trajectory association. The dataset contains 15,420 images for object detection and 130 independent videos for multi-object tracking, of which 40 videos are selected for the final tracking evaluation. Experimental results show that BR-YOLOv10s achieves a precision, recall, and mean average precision (mAP) of 95.7%, 90.3%, and 95.2%, respectively. Compared with the YOLOv10s-ByteTrack baseline, the proposed method improves HOTA, MOTA, MOTP, and IDF1 by 4.4, 5.7, 3.0, and 6.1 percentage points, respectively, while reducing identity switches by 23.19%. In addition, the tracker achieves a processing speed of 43.2 FPS. These results demonstrate that the proposed method can effectively perform real-time detection and tracking of densely distributed dairy cows in complex intensive farming environments. Full article

The North China Plain is a core grain-producing region in China. Under the dual pressures of global warming and intensifying human activities, frequent agricultural droughts pose a serious threat to regional food security and sustainable agricultural development. Based on ERA5-Land root-zone soil-moisture data (1982–2022), this study constructed the Standardized Soil Moisture Index (SSMI) and applied an MMK trend test, Partial Wavelet Coherence (PWC), and sub-regional Pearson correlation to investigate drought evolution and driving mechanisms. Groundwater-level and effective-irrigated-area datasets were further incorporated to complement soil-moisture-based analysis and investigate the anthropogenic drivers behind drought variations. The results show that: (1) The regional-average SSMI showed a significant decreasing trend at a rate of −0.013 decade⁻¹ (p < 0.001), indicating the intensification of agricultural drought. Meanwhile, the SPEI presented a non-significant increasing trend of 0.145 × 10⁻² decade⁻¹ (p < 0.001), revealing the existence of hysteresis between the two drought types. Spring suffered the most severe drought, while drought is alleviated by precipitation replenishment in summer, which highlighted the association between seasonal differences and regional climatic characteristics. (2) Groundwater level (GW) is identified as the dominant explanatory factor for agricultural drought (POSC = 70.53%, AWC = 0.93), far exceeding precipitation (16.66%) and evapotranspiration (8.71%), which underscores the pivotal role of groundwater in regulating soil moisture. (3) The driving mechanisms of agricultural drought exhibit remarkable north–south differentiation: southern sub-regions are predominantly driven by meteorological factors, whereas northern sub-regions are closely associated with intensive groundwater irrigation. Anthropogenic groundwater over-extraction shows strong statistical associations with agricultural drought variability. Future work should combine groundwater dynamics, irrigation and crop phenology. These findings support refined drought early-warning and region-tailored mitigation for the North China Plain. Full article

This article establishes a strong averaging principle for a class of slow-fast stochastic evolution systems that incorporate both McKean–Vlasov interactions and multivalued operators within a Hilbert-space variational framework. The slow variable $\upsilon$ evolves according to a Caputo fractional differential law of order $\varrho \in (0,1)$ and is driven by fractional Brownian motion (fBm) with Hurst index $H_1\in (1/2,1)$, interpreted as scalar or cylindrical according to the underlying state space, while the fast variable $\nu$ follows a standard Brownian-driven stochastic differential equation. This hybrid formulation preserves the memory effects and long-range dependencies of the slow dynamics while maintaining the Markov property of the fast subsystem, which is essential for the averaging argument. The analysis relies on a weighted Volterra–Young estimate for singular stochastic convolutions. Since $H_1>1/2$, the stochastic integral in the slow equation is interpreted pathwise in the Young sense rather than in the Itô sense. In contrast, the fast equation, driven by standard Brownian motion, is interpreted in the usual Itô sense, preserving its Markovian structure essential for the ergodic averaging argument. Under explicit Hölder, monotonicity, coercivity, compactness, and integrability assumptions, well-posedness is proved via a Hölder-space fixed-point argument combined with variational approximation and maximal-monotone limit identification. A key technical contribution of this work is the rigorous treatment of the multivalued fractional equation: an $L^{q^{\prime }}$-integrable measurable selector $\mathfrak{a}$ for the subdifferential term is constructed by identifying it as the weak limit of the regularized Yosida operators, and its membership in the multivalued graph is established through a Minty argument that requires a limsup inequality rather than merely weak convergence. Subject to the standard ergodicity of the Brownian-driven fast variable, strong convergence of the slow component to the solution of the averaged fractional equation is proved. The averaging principle is established for $\varrho \in (1/2,1)$. This restriction arises from the singular block-integral estimate used in the proof, which requires $\varrho >1/2$; the case $\varrho \le 1/2$ is therefore beyond the scope of the present argument and is left for future work. To demonstrate the broad applicability of the abstract framework, the general results are applied to two important classes of fractional stochastic systems. The first consists of finite-dimensional fractional slow-fast McKean–Vlasov (SF-MV) stochastic variational inequalities (SVIs), including reflected stochastic systems in convex domains as an important special case. The second consists of fractional SF-MV multivalued stochastic partial differential equations (SPDEs). For both classes, explicit averaging rates are obtained under suitable Lipschitz assumptions. Full article

Koinobiont parasitoids have been shown to both reduce and increase the growth rate of caterpillars. However, no comprehensive study has been conducted on several caterpillar hosts of a given plant species. Moreover, most of the existing case studies are on large caterpillars attacking crop plants. This may not be representative of the effect of parasitoids on the growth of caterpillars that feed on trees. We sampled caterpillar communities from the canopy of oak trees (Quercus robur and Q. petraea) in a forest in France and from one in Poland, measured their growth rate, and noted parasitoid exit. The details of the methods varied between study years and locations, and this was partially accounted for statistically. Across six species of Lepidoptera, parasitoids reduced caterpillar growth rate on average. The effect of parasitism appeared to vary widely within and across lepidopteran species, indicating diversity in the outcome of plant–herbivore–parasitoid interactions. Further studies should verify this result and determine if the reduced growth rate is accompanied by lower plant consumption rates. If so, plants would immediately benefit from attracting parasitoids when they are attacked by caterpillars, and forestry practices that increase parasitoid abundance could reduce folivory partly without delay. Full article

This work investigated the effects of annealing temperature and holding time on recrystallization, grain growth, and tensile properties of cold-rolled biomedical Co–20Cr–15W–10Ni (L-605) alloy. The alloy was solution-treated at 1200 °C for 30 min, cold-rolled to 40% area reduction, and annealed at 800–950 °C for 15 min, 1000 °C for 5–60 min, or 1200 °C for 5–60 min. Annealing at 800–900 °C produced partially recrystallized microstructures; 950 °C yielded nearly full recrystallization with an average grain size of ~3.4 μm. At 1000 °C, the fully recrystallized microstructure retained ~5 μm fine grains even after 60 min, likely due to grain boundary migration inhibition by fine secondary-phase particles, whose composition and pinning effect remain unclarified. At 1200 °C, grains grew significantly to ~90 μm, attributed to higher grain boundary mobility and reduced particle pinning at elevated temperature. As annealing temperature rose from 800 °C to 950 °C, yield strength decreased from 1245 MPa to 800 MPa, and elongation increased from 12.6% to 45.9%. At 1200 °C, yield strength fell to 442–455 MPa and elongation reached 73.5–80.8%. The results confirm that the alloy’s strength–ductility performance is directly determined by microstructure evolution from retained cold rolling deformation substructures to fine recrystallized grains, and further to coarse recrystallized grains at higher temperatures. Full article

Alheira is a traditionally smoked, fermented, non-ready-to-eat meat sausage produced in the Trás-os-Montes region of Portugal, where lactic acid bacteria (LAB) are major determinants of product safety and quality. This study used whole-genome sequencing to characterize 59 LAB isolates collected from artisanal alheira produced in six municipalities. Eight species belonging to six genera were identified, and average nucleotide identity analysis resolved 24 non-redundant strain groups. Functional annotation revealed broad repertoires of carbohydrate-active enzymes, proteases, lipases, transport systems, and genes associated with tolerance to acid, oxidative, bile, heat, and salt stress. Genotype–phenotype inference was performed with a subset of 22 strain groups represented by unflagged genome assemblies. After false-discovery-rate correction, carbon-metabolism, transport, energy-metabolism, total functional-gene, acid-stress, bile-stress, and total-stress counts were positively associated with total pH decline. A mixed-effects model showed a positive association between standardized total functional-gene count and acidification ($b=0.137\pm 0.059$, $p=0.049$; marginal $R^2=0.266$, conditional $R^2=0.582$), but within–between decomposition indicated that the association was detectable between species rather than within species. Regularized regression, Random Forest, AICc model comparison, and partial least-squares regression converged on carbon metabolism, transport, energy metabolism, and acid- or bile-stress functions as the principal exploratory signals; however, sparse LASSO regression for variable selection was unstable, and prediction of unobserved species was limited. None of the five targeted pathways: acetaldehyde, citrate, diacetyl, exopolysaccharide, and lactose metabolism, remained significant after multiple-testing correction. These findings provide a genomic and phenotypic basis for selecting native starter-culture candidates while emphasizing the need for independent functional and safety validation. Full article

Cyclic heat stress compromises egg production and antioxidant defenses in laying hens. The present study employed a short-term, 14-day cyclic heat-stress model to evaluate whether dietary supplementation with a chlorogenic acid-rich green coffee bean extract could mitigate these effects. Before initiation of the heat-stress experiment, all hens were maintained under thermoneutral conditions and received the basal diet without the extract. Following 1 week of adaptation and 4 weeks of thermoneutral feeding, 240 approximately 50-week-old Jingfen No. 8 hens were allocated to a thermoneutral control, a heat-stress control, or heat-stressed groups receiving 100 or 200 mg/kg extract containing 70% chlorogenic acid, with four independent replicates of 15 hens each (60 hens per treatment). Heat-stressed hens were subjected to a 14-day cyclic heat-stress challenge at 32–35 °C for 9 h/day. Heat stress decreased average egg weight and laying rate, increased serum malondialdehyde, and reduced total antioxidant capacity (p < 0.05). Both supplemented groups improved average egg weight and laying rate, reduced malondialdehyde and total cholesterol, and increased total antioxidant capacity relative to heat-stressed controls (p < 0.05). Supplementation altered hepatic expression of genes associated with proliferation, apoptosis, and autophagy. Exploratory metabolomics detected changes in the 100 mg/kg group but not the 200 mg/kg group relative to heat-stressed controls. The extract partially mitigated production and antioxidant impairments during cyclic heat stress. Full article

This study investigated the effects of graded dietary barley replacement for corn on growth performance, rumen fermentation, carcass traits, meat quality, fatty acid profiles, and the rumen microbiome in Arbas White Cashmere goats, and further applied untargeted muscle metabolomics to elucidate the metabolic mechanisms underlying meat quality alterations. Growth performance was monitored for all goats (n = 10 per group); rumen fermentation, carcass traits, serum biochemistry and meat quality were determined in six randomly selected goats per group (n = 6); and rumen microbiota profiling and muscle metabolomics were performed on three selected goats per group (n = 3). Partial barley substitution (33% and 67% of starch from barley) significantly improved final body weight and average daily gain (p < 0.05), whereas total replacement (100% barley starch) did not confer additional growth advantages. Moderate barley inclusion increased ruminal propionate concentration (p = 0.001) and enriched the fiber-degrading bacterium Prevotella, while total barley replacement markedly reduced rumen microbial diversity (Sobs, Chao1, ACE, and Shannon indices; p < 0.01) and depleted multiple fibrolytic and hydrogenotrophic genera, including Christensenellaceae_R-7_group, NK4A214_group, and Methanobrevibacter. The 33% barley group exhibited elevated serum total bile acids, FGF-19, and CYP27A1 levels (p < 0.05), suggestive of modulated bile acid metabolism, along with decreased glucose-6-phosphatase (p < 0.05), pointing to suppressed hepatic gluconeogenesis. Muscle metabolomics revealed that moderate barley substitution predominantly affected glycerophospholipid metabolism, with upregulation of PE(P-18:1/20:4) and PC(16:0/16:0), whereas total replacement induced extensive metabolic reprogramming characterized by downregulation of glycine and glutathione-related metabolites, indicative of compromised antioxidant defense. The 33% barley group also exhibited the highest muscular C18:3n-3 content (p < 0.01), while the 67% and total replacement groups showed increased C22:6n-3 deposition (p < 0.05). Correlation analysis confirmed that PE(P-18:1/20:4) and PC(16:0/16:0) were significantly positively correlated with ruminal propionate and ADG (r = 0.841–0.986; p < 0.05), and Prevotella abundance showed a strong negative trend with G-6-Pase (r = −0.771). Collectively, these findings demonstrate that replacing one-third of dietary corn starch with barley starch optimally balances productivity, rumen health, and meat quality through enrichment of Prevotella, enhanced propionate production, and modulation of bile acid metabolism, while excessive barley inclusion destabilizes the rumen ecosystem and triggers muscle oxidative stress, providing a theoretical basis for precision grain formulation in intensive cashmere goat production. Full article

Many current studies have purely regarded green finance as green credit, ignoring the background of carbon neutrality and energy transition. This paper investigates the relationships among green finance, green technological innovation, and urban ecological welfare performance. Using 279 Chinese cities as examples, this paper reveals the following: there is (1) a direct positive effect, where a 0.1-unit absolute increase in the green finance index (GFI) is associated with an average 0.0605-unit rise in ecological welfare performance (EWP); (2) a partial mediation mechanism through green technological innovation, establishing a “finance → technology → ecology” pathway; (3) temporal persistence of positive effects across pre-2012 and post-2012 policy periods; and (4) regional disparities with significant impacts in Eastern/Western/Northeastern China but insignificant effects in Central China due to lower green credit allocation and weaker R&D intensity. Robustness checks, including Winsorization and subgroup analyses, validate these results. This study advances the integration of environmental finance with sustainable development theory, offering actionable insights for achieving ecological welfare goals. Full article

Background/Objectives: Sleep duration and physical activity are important lifestyle behaviors that may influence adolescents’ academic and Physical Education (PE) performance. However, evidence regarding their associations with Physical Education (PE) performance remains limited, particularly in North African populations. This study aimed to examine the relationships between sleep duration, physical activity, and academic and PE performance among Moroccan adolescents. Methods: A cross-sectional study was conducted among 439 Moroccan adolescents aged 14–20 years. Data were collected using the Arab Teens Lifestyle Study (ATLS) questionnaire. Sleep duration, physical-activity level (metabolic equivalent minutes per week [MET-min/week]), overall academic average, average grade in science subjects, and final PE grade were assessed. Pearson correlation analyses and multivariate analysis of variance (MANOVA) with follow-up univariate analyses were performed, with Games–Howell post hoc comparisons used for PE performance. Results: Sleep duration was positively associated with overall academic average (r = 0.147, p = 0.002) and final PE grade (r = 0.167, p = 0.001), but not with average grade in science subjects (r = 0.080, p = 0.099). Physical activity was positively associated with overall academic average (r = 0.169, p < 0.001) and final PE grade (r = 0.166, p = 0.001), but not with average grade in science subjects (r = −0.009, p = 0.851). Overall academic average was strongly correlated with science average (r = 0.809, p < 0.001) and moderately correlated with final PE grade (r = 0.474, p < 0.001). multivariate analysis of variance (MANOVA)revealed a significant but small overall difference in combined academic and PE performance across the four sleep-duration and physical-activity groups (Pillai’s Trace = 0.034, F(6, 870) = 2.522, p = 0.020, partial η² = 0.017). Follow-up analyses showed no significant group differences in overall academic average (F(3, 435) = 1.460, p = 0.225, partial η² = 0.010), whereas PE performance differed significantly across groups (F(3, 435) = 4.040, p = 0.007, partial η² = 0.027). Conclusions: Sleep duration and physical activity showed weak but significant positive associations with overall academic average and PE performance among Moroccan adolescents, whereas neither was significantly associated with average grades in science subjects. Combined sleep-duration and physical-activity groups differed significantly in PE performance, but not in overall academic average, with small effect sizes. These findings suggest that sleep duration and physical activity are associated with adolescent performance, although the magnitude of these associations is modest and should not be interpreted as causal. Further longitudinal research using objective measures of sleep and physical activity is warranted to clarify these relationships. Full article

Multilingual online reviews combine ordinal ratings, asymmetric positive and negative evidence, sparse attribute occurrence, and substantial cross-country imbalance. This study presents an integrated inferential framework. Signed topic scores are converted to within-country rank intensities, and country-specific cumulative-logit models distinguish positive and negative occurrence baselines from their corresponding intensity contrasts. The two intensity contrasts are then synthesized jointly through a bivariate Bayesian normal-normal random-effects model that retains their within-country covariance. The primary analysis uses all 30,042 reviews observed in the common 2019–2024 window; the complete 85,373-review corpus and repeated country-capped samples are sensitivity analyses. Separating the two occurrence baselines improves summed AIC from 32,347.7 to 32,167.8, while a transformation-by-function comparison shows that natural splines improve AIC and quadratic-weighted agreement. The linear-rank model is retained to provide comparable scalar intensity contrasts. Targeted partial proportional-odds fits substantially improve in-sample AIC in China and Japan but leave repeated-validation performance and all nine average effect directions essentially unchanged. A secondary semantic mapping audit agrees with 30 of 31 topic assignments. Simulation results show generally adequate interval coverage but reduced Kano-state accuracy under small K, sparse occurrence, and proportional-odds violations. The findings therefore support tiered, uncertainty-aware prioritization rather than a deterministic global ranking. Full article

The cooperative tracking of multiple ground targets by multiple UAVs remains challenging under partial observability, limited communication, and obstacle constraints, owing to complex target association, difficult task handover, strong coupling among low-level continuous control decisions, and unstable critic value estimation. To address these issues, this paper proposes a hierarchical-guidance and Gaussian-process-corrected multi-agent proximal policy optimization method, termed HGP-MAPPO. Built upon the centralized-training and decentralized-execution paradigm, HGP-MAPPO introduces low-frequency task-guidance signals derived from target-association information, task handover and recovery cues, task priorities, and desired observation geometry. These guidance signals are incorporated as conditional inputs into the low-level actor–critic framework, thereby reducing the policy learning difficulty in jointly handling target tracking, occlusion recovery, obstacle avoidance, and smooth control. Moreover, to alleviate local estimation bias in the neural-network critic under complex partially observable conditions, a Gaussian-process-based residual correction mechanism is designed. Specifically, the posterior mean is used to compensate for value residuals, while the posterior uncertainty adaptively regulates the correction intensity, improving the stability of value evaluation and policy optimization. A sparse inducing-point approximation is adopted to control the training-stage computational cost, while the Gaussian-process module is removed during decentralized execution and, therefore, introduces no additional online inference overhead. Experiments are conducted in standard-obstacle and densely obstructed multi-UAV multi-target tracking scenarios, with DDPG-MHSA, MAPPO, MADDPG, and MATD3 adopted as baselines. The experimental results demonstrate that HGP-MAPPO achieves faster training convergence, higher average episode rewards, and improved target retention rates. It also effectively reduces UAV–target distance fluctuations and the mean absolute temporal-difference (TD) error. Ablation studies further confirm the contributions of task-guidance signals, Gaussian-process residual correction, and uncertainty-aware weighting to cooperative tracking performance and training stability. Full article

The increasing demand for sustainable agricultural practices has driven growing interest in organic fertilizers derived from waste streams as alternatives to mineral fertilizers. This study evaluated, in a pilot trial, the effectiveness of using raw digestate, composted digestate and composted sewage sludge in a crop rotation over two consecutive years of maize and Italian ryegrass. Crop yield, forage quality, and soil physicochemical properties were assessed to determine the potential of these recycled organic nitrogen sources. The results differed between the two species of forage. In maize, organic fertilizers maintained yield and forage quality comparable to chemical fertilization, while composted sewage sludge significantly increased average plant height. These findings indicate that recycled organic amendments can partially or totally replace mineral nitrogen fertilizers, contributing to nutrient recycling and waste valorization within a circular economy framework. In contrast, Italian ryegrass showed lower yield under organic fertilization than under chemical fertilization, suggesting that additional management strategies may be required to optimize nutrient availability for this crop. Furthermore, raw digestate application tended to increase soil pH, which may provide benefit subsequent crop. Overall, the results suggest that digestate and sludge-based fertilizers are promising alternatives for maize performance, although their effectiveness in Italian ryegrass requires additional strategies. Full article

Background/Objectives: Artificial intelligence (AI)-based chatbots are increasingly used as sources of health information, yet the quality and semantic stability of their responses may vary across systems, languages, repeated queries, and prompt formulations. This study compared expert-rated general information quality and embedding-based semantic stability of responses generated by seven AI chatbot systems to expert-derived, patient-oriented complete denture questions in Turkish and English. Methods: Twenty-two complete denture-related questions were submitted to seven chatbot systems on three study days across scheduled morning, afternoon, and evening sessions. Two additional semantically equivalent Turkish variants were generated for each original question. Five prosthodontists assessed the original Turkish and English responses using the 5-point Global Quality Score (GQS). Semantic stability was quantified using the multilingual SentenceTransformer checkpoint paraphrase-multilingual-MiniLM-L12-v2 and cosine similarity. To incorporate complete responses, responses were divided into non-overlapping token-based chunks, chunk embeddings were combined into one normalized full-response representation, and the 36 pairwise similarities from the nine repeated responses were averaged to one stability estimate per question–model–language/variant condition. Linear mixed-effects models with question-level clustering were used, with Bonferroni adjustment and partial eta squared effect sizes with 95% confidence intervals. Results: AI model significantly affected embedding-based semantic stability (p < 0.001; ηp² = 0.820, 95% CI 0.795–0.837) and GQS (p < 0.001; ηp² = 0.249, 95% CI 0.152–0.316). Grok had the numerically highest overall semantic stability mean (0.952 ± 0.015), whereas GPT-4o had the numerically highest overall GQS (4.48 ± 0.85). English responses showed higher overall GQS than the original Turkish responses (3.96 ± 0.94 vs. 3.66 ± 1.23; p = 0.004) and higher embedding-based similarity than the Turkish conditions overall (p < 0.001). No significant overall effect of scheduled study day was observed (p = 0.504; ηp² = 0.001), whereas semantic stability differed across scheduled query sessions (p < 0.001; ηp² = 0.055), with means of 0.863 ± 0.106, 0.874 ± 0.092, and 0.841 ± 0.140 for morning, afternoon, and evening sessions, respectively. Inter-rater reliability was moderate for English GQS ratings (ICC = 0.675, 95% CI 0.581–0.752) and good for Turkish ratings (ICC = 0.771, 95% CI 0.687–0.833). Conclusions: The evaluated chatbot systems differed in expert-rated information quality and full-response embedding-based semantic stability. English responses showed higher overall values than Turkish responses, although cross-language measurement effects of the embedding model cannot be excluded. Differences across scheduled query sessions should be interpreted as run-to-run output variability rather than intrinsic temporal behavior. These findings characterize comparative chatbot performance under the tested conditions but do not establish clinical accuracy, safety, or patient education effectiveness. Full article

Foundation models for Earth observation are commonly transferred to downstream tasks without explicit attention to the preprocessing mismatch between pretraining and target distributions. This study isolates preprocessing mismatch as a controlled experimental factor in transferring the Hydro multispectral foundation model, a Swin Transformer V2 Base encoder pretrained with SimMIM on Sentinel-2 Level-2A water-body imagery, to the Marine Debris and Oil Spill (MADOS) marine pollution benchmark dataset, processed through ACOLITE Rayleigh reflectance and providing 11 of the 12 spectral bands used during pretraining. The two datasets are therefore produced by different atmospheric correction algorithms under different reflectance conventions, and the resulting per-band statistical discrepancy is quantified as the starting point of the analysis. Three preprocessing dimensions are then systematically varied while all other settings are held constant: input normalisation, spectral band adaptation for the missing B09, and encoder transfer mode. From this, four findings emerge. Normalisation mismatch between training and inference is the single largest source of performance degradation, reducing the mean Intersection over Union (mIoU) by 0.458, more than seven times the largest radiometric perturbation tested. A zero-parameter Frobenius-matched column crop of the patch embedding adapts the 12-band pretrained encoder to the 11-band target, at least as effectively as any learnt linear or nonlinear adapter, at a lower cross-seed variance. Under limited target supervision (1433 training patches against an 87.9 million-parameter encoder), freezing the encoder outperforms both fine-tuning in full and random initialisation training from scratch. The gains of partial unfreezing are attributable to augmented training (very simple copy–paste (VSCP) augmentation, exponential moving average (EMA), and test-time augmentation (TTA)) rather than to encoder adaptation. With matched preprocessing, the frozen encoder reaches 0.600 mIoU and matches the published MariNeXt baseline within seed variability. Mechanistic analysis via band-occlusion attribution and feature-space separability shows that input normalisation determines which spectral bands the encoder relies upon, with the magnitude of the shift correlated to the per-band gap between the source and target distributions. Operationally, preprocessing alignment, rather than architectural modification, carries most of the practical effort in transferring a multispectral foundation model to marine surface segmentation. These results are established for a single encoder–benchmark pair under limited target supervision. The mechanism they identify is more portable than the magnitude reported. A frozen encoder’s representations remain bound to the normalisation statistics of its pretraining dataset, so any transfer that departs from these statistics at inference is predicted to degrade sharply in proportion to the per-band distance between the two distributions. Full article