Cross-domain few-shot image semantic segmentation is widely applied in fields such as medical image analysis and remote sensing image processing. This paper focuses on the field of cross-domain few-shot image semantic segmentation, presenting the first systematic review in this direction. Firstly, the development from image semantic segmentation and few-shot image semantic segmentation to cross-domain few-shot image semantic segmentation is outlined, identifying the core challenges as domain shift, scarcity of annotated data, and insufficient model generalization capability. Subsequently, nine commonly used datasets and five key evaluation metrics are summarized. Existing cross-domain few-shot image semantic segmentation methods are categorized into ten subclasses from three dimensions: feature alignment strategies, model architecture design, and data utilization approaches; and their key strategies are analyzed. Finally, the limitations of current methods and potential future research directions are discussed, aiming to provide researchers with a comprehensive overview of the current state and emerging trends in this field.

In the field of energy management for fuel cell vehicles, energy management strategies based on deep reinforcement learning have become a research hotspot for hybrid powertrains. However, insufficient learning capability, low learning efficiency, and difficulties in model convergence have still been observed under various complex driving conditions. To address these problems, an energy management strategy based on an improved deep reinforcement learning algorithm is proposed. The traditional TD3 algorithm is modified, and prioritized experience replay together with a reward-oriented adaptive noise scale is adopted to obtain an improved AE-TD3 algorithm. The performance of the AE-TD3 algorithm under standard driving conditions is evaluated on the CLTC, UDDS, FTP75 and RCDC cycles, and its performance under real urban road driving conditions is also tested. Simulation studies are carried out, and the AE-TD3 algorithm is found to converge faster and to possess stronger exploration capability than the TD3 algorithm. Regarding the output performance of the energy storage unit, the fluctuation range of the battery SOC relative to its target value obtained by the AE-TD3-based energy management strategy is significantly smaller than that obtained by the TD3-based strategy under all four training cycles. In addition, when the vehicle operats at high speed, the fuel cell output power under the AE-TD3-based strategy is significantly higher than that under the TD3 algorithm. In terms of fuel cell hydrogen consumption, the hydrogen usage is reduced by 2.7%, 1.1%, 5.7% and 7.3%, respectively, compared with the TD3 algorithm. By applying the energy management strategy trained under standard driving cycles to real-world road conditions, it is shown that the AE-TD3-based energy management strategy has good online applicability and strong adaptability to different driving conditions.

With the large-scale development of Electric Vehicle (EV), their regulatory potential as "mobile energy storage units" cannot be overlooked, which profoundly influence the operational paradigm of power systems. In the context of EV grid integration, fully considering the dual characteristics of EV as controllable loads and mobile energy storage, a comprehensive dynamic charging/discharging scheduling model for EV is constructed, incorporating multiple key factors such as EV charging demand, dynamic electricity prices, time-coupling constraints of energy storage, and battery degradation. To address the randomness of EV charging start times and initial states, as well as the curse of dimensionality and convergence difficulties of traditional reinforcement learning methods in scenarios with continuous decision variables, an intelligent charging/discharging control and optimal scheduling algorithm based on Twin Delayed Deep Deterministic Policy Gradient (TD3) is proposed. Through continuous interaction between the agent and the environment and the design of a reward feedback mechanism, this algorithm can make optimal charging/discharging decisions based on electricity price fluctuations, ensuring that the expected charging capacity is achieved after the charging process, thereby realizing intelligent control and optimal scheduling of EV charging/discharging behavior to minimize charging costs. Simulations based on real-world scenario data demonstrate that the proposed algorithm effectively adapts to dynamic electricity price changes in smart grids and significantly reduces charging costs for EV users. Compared with a series of mainstream algorithms (such as DDPG, DQN, PSO, etc.), the proposed algorithm reduces charging costs by 4.41% to 24.23%, fully validating its performance and economic advantages.

To improve the accuracy of wind power forecasting, a combined ultra-short-term wind power prediction model integrating decomposition optimization and a multi-objective loss function is proposed. Firstly, the optimal number of modal components for variational modal decom position is dynamically searched based on an improved Gray Wolf Optimization Algorithm to achieve efficient decomposition of wind power series. The hyper-parameters of the prediction model are adaptively optimized to enhance the model’s generalization ability. Secondly, the improved gray wolf algorithm is introduced for adaptive hyper-parameter optimization, further improving generalization. A multi-objective loss function integrating prediction accuracy, stability, and grid-connection eligibility is designed. The prediction results of modal components are co-trained, and the final wind power prediction is reconstructed through weighted superposition of each component’s results. Validation experiments are conducted using actual wind power data from different seasons. The results show that the model’s optimal values of standardized mean absolute error, standardized root mean square error, and coefficient of determination reached 1.68%, 0.01%, and 99.45% respectively, significantly outperforming other models. Experimental results show that the proposed model has significant advantages in prediction accuracy and dynamic adaptability.

To solve the chaotic synchronization problem with unknown system parameters model uncertainty, and external disturbances,a new robust adaptive synchronization control strategy is proposed, which consists of two parts. Thefirst part is a nonlinear robust controller that ensures the stability of the closed-loop system and exhibits excellent robustness and fast convergence characteristics.The second part is a nonlinear adaptive law, which utilizes the estimated model uncertainty and bounds to effectively compensate for external disturbances and model uncertainty. The strategy combines adaptive control with parameter identification to enable the controller to achieve efficient synchronization under system uncertainties and perturbations by estimating the unknown parameters of the system in real time. The global asymptotic stability of the control system is proved by applying the Lyapunov stability theory, which shows that the strategy achieves the smooth convergence of the synchronization system error to zero within 0.1 s and improves the robustness of the system. Finally, the correctness and validity of the theoretical analysis are verified by simulation of the PMSM system.

To address the issues of low recognition accuracy for tiny targets and strong background interference in medical glove surface defect detection, this paper proposes a lightweight detection model based on the Transformer architecture, named PAM-DETR. First, the PRT-Block module is constructed based on the RT-DETR model, introducing a re-parameterization structure and an attention mechanism. This significantly enhances the representation capability of small target features while eliminating the computational redundancy of standard convolutions. Second, an Adaptive Sparse Encoding Module (ASEM) is designed to replace the original AIFI structure, thereby optimizing the interaction efficiency across different feature scales. Third, by integrating the ASEM with the newly designed CSPOmniKernel structure, a Multi-Scale Enhanced Feature Pyramid (MSEFP) is constructed to achieve efficient multi-scale fusion of tiny defect features. Finally, a dedicated medical glove defect dataset is built based on industrial on-site collection for validation. Experimental results demonstrate that the improved PAM-DETR algorithm achieves an increase of 4.12 and 4.84 percentage points in the mAP@50 metric compared with the strong baseline models YOLO11n and RT-DETR-R18, respectively. Furthermore, compared with RT-DETR-R18, the number of parameters in PAM-DETR is reduced by 14.5%, and the computational cost (FLOPs) is decreased by 7.9%, effectively meeting the production line’s dual requirements for high-precision and lightweight defect detection of medical gloves.

Graph Neural Network (GNN) have been widely applied in numerous real-world scenarios due to their powerful modeling capabilities for graph-structured data. Recent studies indicate that GNN may inherit and amplify biases inherent in training data, leading to unfair treatment of specific groups defined by sensitive attributes. To mitigate bias in graph data, this paper proposes FairDC, a fair graph learning framework based on view decoupling and counterfactual augmentation, which addresses feature bias and structural bias separately. The framework first decouples raw graph data into feature views and structural views, then introduces counterfactual views. Finally, a multi-view fusion strategy is employed to learn fair node representations. Experimental results on multiple benchmark datasets demonstrate that FairDC maintains stable prediction performance while reducing fairness metrics DP and EO by 32% and 36%, respectively, compared with the state-of-the-art baseline DAB. This validates the proposed method’s effective trade-off between utility and fairness.

The multi-clause deduction algorithm is the reasoning core of automated theorem provers based on the contradiction separation rule. It is characterized by multi-clause and dynamic deduction features that differ from binary deduction methods. Currently, clause selection strategy methods are a hotspot in research on multi-clause deduction, effectively optimizing multi-clause deduction paths. However, there is a lack of comprehensive evaluation aimed specifically at the deduction paths themselves. The standard contradiction separation clause evaluation method is a novel multi-clause deduction path evaluation mechanism that can effectively guide the search for multi-clause deduction paths. The multi-criteria decision making method is applied to the evaluation of standard contradiction separation clause. Firstly, the attribute of the contradiction separation clause is measured, objectively weighted using the entropy weight method, and evaluated through a combination of multi-criteria optimization and compromise solutions. Secondly, based on this evaluation method, a multi-clause deduction algorithm is proposed, which can evaluate the standard. Finally, this multi-clause deduction algorithm is applied to the international advanced first-order logic contradiction separation clause while dynamically updating its evaluation criteria. It can avoid searching for invalid paths through a backtracking mechanism, thereby effectively improving the inference ability of multi-clause deduction. The proposed algorithm is implemented in the automated theorem prover Eprover 3.2 and tested on the problems from the last three years of international automated theorem provers competition and TPTP (Thousands of Problems for Theorem Provers) problem library with a rating of 1. Eprover3.2 with the proposed algorithm solves 14, 14 and 20 additional theorems compared with the original Eprover3.2 respectively, and it also solves 9 theorems with a rating of 1. The experimental results show that the proposed multi-clause deduction method can be effectively applied to the first-order logic automated theorem proving.

The problem of weakly persistent centers for a ten-parameter family of complex planar cubic polynomial differential systems at the origin is studied. Firstly, the necessary conditions for the origin of the system to be a weakly persistent center are obtained by calculating and decomposing the algebraic variety of the ideal generated by the first seven focal quantities. Then, it is proved that these conditions are also sufficient by either using Darboux integrable theory to construct the Darboux first integral or Darboux integrating factor, or verifying the time reversibility of the system, or applying the induction to demonstrate the existence of polynomial first integral. Finally, the complete classifications for a weakly persistent center in a family of real planar cubic polynomial differential systems is derived, which is obtained by setting the variables and the coefficients of the complex systems are conjugation.

This paper studies a class of two-disease epidemic models that incorporate both SIS and SIRS transmission mechanisms. First,the local asymptotic stability of the four local equilibrium points in the deterministic model is rigorously established,and the global stability of the disease-free equilibrium is proved under the condition that the basic reproduction numbers satisfy R₁<1 and R₂<1. Subsequently, the corresponding stochastic model is formulated,for which the existence and uniqueness of a global positive solution are proved. Furthermore,sufficient conditions for disease extinction and persistence in the mean are derived. Specifically,when the stochastic reproduction number, the disease is shown to become extinct,whereas if, the disease is persistent in the mean. Finally, theoretical results are validated through numerical simulations.

In practice, HMVCM faces two main challenges: the presence of strong influence points and the absence of data. Strong influence points may seriously distort the model parameter estimation and lead to poor prediction performance. Missing data may lead to information loss, affecting the accuracy and reliability of the model. Therefore, effectively addressing these two types of problems is critical to improve model performance. In order to solve these problems, this paper uses an outlier detection scheme based on Bayesian method, introduces indicative variables to deal with the missing data mechanism, and uses MH algorithm and Gibbs sampling to obtain Bayesian estimation of missing data. The results show that HMVCM combined with Bayesian inference can adapt to datasets with complex dynamic characteristics, and shows good performance in the face of strong influence points and missing data. The effectiveness of the proposed method is verified in simulation experiments, which shows that the model can maintain high accuracy and robustness in the dataset with complex structures.

Cadmium (Cd) contamination in agricultural soils has become increasingly prominent, exerting a serious impact on the growth and quality of rice and other major crops, thereby threatening the stability of agricultural ecosystems and human health. To investigate the application potential of industrial by-products in the remediation of Cd-contaminated soil, this study focused on two typical industrial by-products from Guangxi, steel slag (SS) and lime slag (LM), and examined their effects on Cd bioavailability in the soil-rice system through soil incubation and rice pot experiments. The soil incubation experiment was conducted with different application rates of SS and LM (1, 2, 5, 7, and 10 g·kg^-1, denoted as SS1, SS2, SS5, SS7, SS10 and LM1, LM2, LM5, LM7, LM10, respectively). The results showed that soil pH and available silicon content increased with increasing application rates of SS and LM, while soil redox potential (E_h) decreased. Compared with the control (CK), SS2 and LM10 had the best effect on reducing soildiethylenetriaminepentaacetic acid extractable Cd(DTPA-Cd) content, decreasing 49.32% and 42.43%, respectively. The rice pot experiment was conducted with different application rates of steel slag (SS) (1, 2, 3, and 5 g·kg^-1, denoted as SS-R1, SS-R2, SS-R3, SS-R5) and lime slag (LM) (1, 3, and 5 g·kg^-1, denoted as LM-R1, LM-R3, LM-R5). The results showed that the application of steel slag and lime slag affected Cd uptake by rice. Specifically, SS-R3 and LM-R5 significantly reduced soil DTPA-Cd content and Cd concentrations in rice roots, stems, leaves, and brown rice (P < 0.05). Compared with CK, the reductions were 40.00%, 80.59%, 86.83%, 78.13%, and 78.57% for SS-R3, and 40.91%, 24.77%, 60.41%, 65.53%, and 69.05% for LM-R5, respectively. The brown rice Cd contents across treatments followed the order: CK (0.38 mg·kg^-1) > LM-R1 (0.25 mg·kg^-1) > SS-R1 (0.22 mg·kg^-1) > LM-R3 (0.19 mg·kg^-1) > SS-R5 (0.17 mg·kg^-1) > LM-R5 (0.12 mg·kg^-1) > SS-R2 (0.09 mg·kg^-1) > SS-R3 (0.07 mg·kg^-1). Notably, the brown rice Cd contents in theSS-R2, SS-R3, SS-R5, LM-R3, and LM-R5 treatments met the national food safety standard (< 0.2 mg·kg^-1, GB 2762-2022). In conclusion, the industrial by-products steel slag and lime slag significantly reduced soil Cd bioavailability and inhibited Cd uptake by rice, with steel slag demonstrating superior efficacy in reducing Cd content in various rice tissues compared to lime slag.

Land surface temperature (LST), as an important indicator of the earth’s surface energy exchange process, plays a crucial role in the stability of ecosystems. Based on the 2001-2021 MOD11A1 V6 time-series LST data, combined with land cover types, NDVI(normalized difference vegetation index), and elevation data, this study employed Theil-Sen trend analysis, Mann-Kendall nonparametric tests, correlation analysis, centroid migration model, and an improved radius method to conduct a multi-scale, multi-dimensional comprehensive analysis of the spatiotemporal distribution and evolution characteristics of LST in Gansu Province at different temporal scales (seasons and years) and for different land cover types. The results showed that: 1) Between 2001 and 2021, the annual mean LST in Gansu Province was maintained at 21.25 ℃ with the area of cooling regions larger than that of warming regions, and an overall decreasing trend of 0.12 ℃ per decade.LST rises significantly in Spring while LST cools down significantly in Summer.LST in the Hexi region was generally higher than in other areas of Gansu Province, forming a north-high, south-low geographic distribution pattern. The centroid of high-temperature regions continuously shifted northwest, whereas the centroid of low-temperature regions moved southeast. 2) Significant differences in LST were observed among different land cover types. Bare land had the highest annual mean LST at 26.30 ℃, while glaciers had the lowest at -1.77 ℃. Vegetation cover was significantly negatively correlated with LST. 3) Among the six selected representative local cities, most exhibited varying degrees of heat island effects. Jiuquan City showed the most significant heat island effect, followed by Lanzhou and Qingyang; Longnan and Gannan cities had relatively weak heat island effects, while Wuwei City displayed a cold island effect.

To reveal the distribution characteristics of glomalin-related soil protein (GRSP) in the heterogeneous habitats (inside, fringe, and outside) of the Guangxi Dashiwei Tiankengs and its relationship with soil physicochemical properties, soils from these three microhabitats of the tiankeng were selected as the research objects. The contents of easily extractable glomalin-related soil protein (EE-GRSP) and total glomalin-related soil protein (T-GRSP) were determined. Combined with soil physicochemical indicators, methods such as Spearman correlation analysis and redundancy analysis were employed to systematically analyze the habitat differentiation characteristics of GRSP and explore its correlation with soil physicochemical factors. The results showed that among the three microhabitats, EE-GRSP was significantly higher outside the tiankeng (1.55 mg/g) than inside (1.35 mg/g), while no significant habitat difference was observed at the Fringe (1.52 mg/g). The T-GRSP content ranged from 3.17 to 3.51 mg/g, with no significant differences among habitats. The proportion of EE-GRSP to organic carbon (EE-GRSP/SOC) ranged from 3.51% to 4.72%, with the Fringe being significantly lower than outside and inside. The proportion of T-GRSP to organic carbon (T-GRSP/SOC) showed no significant differences among habitats. Based on the relationship between soil physicochemical factors and GRSP, it was indicated that organic carbon (SOC) inside the tiankeng, total nitrogen (TN) and total carbon (TC) at the Fringe, and total carbon (TC) outside the tiankeng were the core soil factors affecting GRSP content. Moreover, the accumulation of GRSP increases with the increase in soil carbon and nitrogen storage, while the proportion of GRSP to SOC decreased with the increase in soil carbon and nitrogen storage due to its accumulation rate lagging behind that of soil carbon and nitrogen. Additionally, the effects of soil moisture, pH, and TP on GRSP were habitat-specific. The study confirmed that the accumulation of GRSP in the tiankeng was influenced by habitat heterogeneity and soil carbon-nitrogen factors, and its carbon and nitrogen sequestration function played a key role in maintaining soil quality in the karst tiankeng ecosystem.

To elucidate the characteristics of microbial diversity in different forest soil types within the subtropical monsoon region, this study investigated O/A (0~10 cm) and AB (10~30 cm) horizons of limestone soils (LS) and red soils (RS) under forest vegetation in Northern Guangxi. Combining 16S rRNA amplicon sequencing with soil physicochemical property analysis differences in microbial community structure, functional potential, and co-occurrence networks between soil types and horizons are explored. The results showed that: 1) Both soil types were dominated by Proteobacteria, Actinobacteria, and Acidobacteria. Compared with limestone soils, RS exhibited higher relative abundance of Acidobacteria. Archaeal communities were predominantly Thaumarchaeota (>76%) in both soils. 2) Bacterial and archaeal richness indices were higher in limestone soils and red soils, respectively, though overall alpha diversity showed no significant difference. The O/A horizon displayed higher bacterial richness and diversity than the AB horizon, whereas archaeal richness was greater in the AB horizon. 3) Limestone soils bacterial and archaeal networks exhibited higher modularity and information transfer efficiency, while red soils networks demonstrated enhanced node complexity and robustness, reflecting adaptive mechanisms to acidic environments. 4) Soil pH, organic carbon (SOC), total nitrogen (TN), and available phosphorus (AP) indirectly regulated microbial functional divergence by modulating resource availability. 5) Limestone soils bacteria showed stronger carbon-nitrogen metabolic functions (e.g., nitrification, aromatic compound degradation), whereas red soils bacteria exhibited enhanced nitrogen fixation capacity. Archaeal functional prediction revealed significantly higher aerobic ammonia oxidation activity in limestone soils and greater methanogenic potential in AB horizons of both soils.

Farmland soils developed in the karst regions of distinctive aboveground-underground coupled structure, generally exhibit serious soil erosion and nutrient deficiency, which constrains sustainable agricultural development of the regions. Soil organic matter, particularly the soluble low-molecular-weight fraction therein, is a key driver of soil fertility conservation. Thus, understanding of their composition and distribution is of great significance to improve soil nutrient availability and crop productivity. In this study, organic carbon abundance, chemical structure of dissolved organic matter (DOM), and species and contents of low-molecular-weight organic acids (LMWOAs) in typical farmland soils of the Guilin karst area, Guangxi, were investigated to examine their relationships with soil fertility. The results showed that the compositional structure of DOM were similar among the soils, but significant differences were observed in their contents. The composition of DOM is dominated by aliphatic compounds with a relatively low abundance of aromatic compounds while that of LMWOAs was comprised mainly of oxalic and formic acids. These components were enriched in the high-fertility soils but depleted in the low-fertility ones. Significant positive correlations were observed between SOM content and the contents of aliphatic compounds, aromatic compounds, and LMWOAs, suggesting that aliphatic and aromatic compounds served as effective components for enhancing soil fertility and soil structural stability while LMWOAs for facilitating soil nutrient cycling and dynamic balance. This has important implications for optimizing soil fertility management and guiding fertilization practices in karst farmlands.

Soil organic carbon (SOC) is the core material basis for maintaining the soil quality of plantation, ensuring the long-term forest productivity and exerting their carbon sequestration potential. To investigate the short-term effects of the "forest-fungus" model on soil organic carbon storage and stability in Chinese fir (Cunninghamia lanceolata) forests and promote the carbon sink function of "forest-fungus" model, the content of soil active carbon components and its contribution to total soil organic carbon in Chinese fir forest after 3 years of cultivation of Ganoderma lucidum with bag were analyzed taking Chinese fir pure forest as the control. The results showed that after planting G. lucidum, the soil moisture content, clay ratio, total nutrient and available nutrient content of nitrogen and phosphorus, and carbon-transforming enzyme activities in the 0-20 cm and 20-40 cm layers of Chinese fir pure forest were significantly increased (P<0.05), but the gravel ratio showed the opposite (P<0.05).In the short term, G. lucidum planting significantly increased the soil organic carbon content in the 0-20 cm and 20-40 cm layers of Chinese fir pure forest, by 117.72% and 42.73%, respectively (P<0.05). Although the contents of Easily oxidizable organic carbon (EOC), Particulate organic carbon (POC), Dissolved organic carbon (DOC) and Microbial biomass carbon (MBC) in the 0-20 cm layer of Chinese fir-G. lucidum forest were significantly higher than that in Chinese fir pure forest. However, EOC/SOC, DOC/SOC and MBC/SOC decreased significantly (P<0.05). In the 20-40 cm layer, only the DOC content of Chinese fir-G. lucidum forest was significantly higherthan in Chinese fir pure forest (P<0.05). The biological factors explained 48.1% of soil carbon components variations, and the activities of carbon conversion-related enzymes were significantly positively correlated with the content of each carbon component. Soil water content was the most critical positive factor affecting DOC change. Structural equation modeling showed that G. lucidum planting can indirectly increase biological activity by increasing soil water content and nutrient content, thereby promoting the increase of soil carbon components and soil organic carbon content. Based on this, the bag cultivation of G. lucidum can effectively improve soil carbon storage and stability, achieving the coordinated development of economic and ecological benefits.

The increasing dimensionality of feature variables in remote sensing-based estimation of forest aboveground biomass (AGB) necessitates effective feature selection to enhance model accuracy. This study focuses on Nanning City as the research area, utilizing Sentinel-2 data as the remote sensing source. Spectral information from various bands, texture features, and additional factors such as elevation, slope, and aspect were extracted. Three feature selection methods including stepwise regression, bivariate correlation, and random forest were employed to identify modeling variables. Biomass estimation models were established based on CatBoost and random forest (RF) machine learning algorithms. Five-fold cross-validation was applied to evaluate model performance, and the best model was used to complete biomass mapping. The results indicated that among the three feature selection methods, the bivariate correlation method performed the best across three tree types: pine, eucalyptus, and broadleaf species. For Chinese fir forests, the random forest method showed superior performance. Specifically: For Chinese fir forests, the combination of the random forest feature selection method and the RF algorithm was optimal (R²=0.58, RMSE=8.53 Mg·hm^-2). For Masson pine forests, the bivariate correlation method combined with the RF algorithm was the best choice (R²=0.51, RMSE=11.10 Mg·hm^-2). For eucalyptus forests, the bivariate correlation method combined with the RF algorithm yielded the best results (R²=0.56, RMSE=14.91 Mg·hm^-2). For broadleaf forests, the bivariate correlation method combined with the RF algorithm also proved optimal (R²=0.35, RMSE=40.55 Mg·hm^-2). Feature selection methods have a significant impact on the predictive performance of models. Combining feature selection methods with ensemble machine learning algorithms is conducive to improving the estimation accuracy of AGB.

This study aimed toprobe into the physiological responses of one-year-old seedlings of Acacia mangium, Acacia auriculiformis, and their hybrid clones (No. 8, No. 9, and No. 19) under low-temperature stress. A randomized block design was adopted, with five temperature treatments (25 ℃ as the control, 15, 10, 5, and 0 ℃); samples were taken and physiological indiceswere measured every 3 days. The results showed that under low-temperature stress the chlorophyll content of all Acacia materials declined significantly, and the longer the stress lasted, the lower the chlorophyll level became; malondialdehyde content rose overall, whereas peroxidase and catalase activities first increased and then decreased, and soluble-protein content likewise followed a trend of initial increase followed by decline. Comprehensive evaluation revealed that clone No. 8 of the Acacia mangium × A. auriculiformis hybrid possesses strong low-temperature adaptability, whereas clone No. 19 is markedly sensitive to low-temperature stress. This study provides a scientific basis for elucidating the low-temperature adaptation mechanisms of Acacia mangium × A. auriculiformis and is of great significance for breeding cold-tolerant varieties and optimizing cultivation management practices.