Atmospheric and Oceanic Sciences

The 15th United Nations Biodiversity Conference (COP15) obligates all countries to reduce pesticide risks by 50% by 2030. In this study, we derived the trends of total applied toxicity (TAT) globally between 2013 and 2019, weighting applied masses by ecotoxicity, of 625 pesticides for eight species groups to assess pathways toward this reduction goal. We found that the TAT of most species groups has increased; that only 20 ± 14 pesticides per group define >90% of the TAT nationally; that fruits, vegetables, maize, soybean, rice, and other cereals contribute 76 to 83% of the global TAT; and that China, Brazil, the United States, and India contribute 53 to 68% of the global TAT. Our target achievement categorization shows that substantial actions, combining shifts to less-toxic pesticides, increased adoption of organic agriculture, and also provision of national pesticide use data, will be required globally to approach the United Nations’ target.

Programmable control of protein or colloidal nanoparticle self-assembly into targeted nanostructures, while maintaining stability across extreme pH conditions, remains a major challenge. We designed coiled-coil bundlemer peptide nanoparticles that form ordered, hierarchical materials across an unusually broad pH range (1, 7, and 14) dependent on patchy surface charge display. Nematic liquid crystal formation was observed at low concentration (~0.5 to 4 weight %) at pH 1 and pH 14, whereas higher concentration at pH 1 yielded hexagonal columnar phases. At neutral pH, the same patchy nanoparticles assembled into ordered lattices through electrostatic complexation. Molecular dynamics simulations revealed end-to-end particle stacking underlying all phases. Coiled coils with identical amino acid composition but lacking designed charge patches displayed no ordered assembly, demonstrating the importance of programmable electrostatic interactions with protein-like specificity of spatial display.

Chemical protein synthesis enables the construction of specific protein architectures but is limited to millimolar reaction concentrations, restricting access to poorly soluble proteins. Potassium acyltrifluoroboronates (KATs) offer a promising alternative through fast and chemoselective amide bond formation, but their application to protein synthesis has been precluded by the lack of a masking strategy. We report chiral, zwitterionic organoboron complexes that mask amino acid–derived KATs. These molecules exhibit unexpected nitrogen-carbon-boron connectivity and are fully compatible with solid-phase peptide synthesis and stereoretentive deprotection. We synthesized C-terminal KAT peptides and demonstrated KAT ligation at micromolar concentrations for the convergent synthesis of the aggregation-prone programmed death ligand 2 (PD-L2) immunoglobulin V domain. This work establishes organoboron chemistry as an enabling strategy for chemical protein synthesis at low concentrations far more suitable for handling large, aggregation-prone biomolecules.

The atmospheric methane (CH 4 ) growth rate surged after 2019, peaking at 16.2 parts per billion per year (ppb year −1 ) in 2020 before declining to 8.6 ppb year −1 in 2023. Using multiple atmospheric inversions constrained by observation- and model-based prescribed hydroxyl radical (OH) fields and CH 4 atmospheric data, we show that a drop of OH radicals in 2020–2021, followed by recovery in 2022–2023, accounted for 83% of year-on-year variations in the CH 4 growth rate, the rest being explained by wetland and inland water emissions, which increased between 2019 and 2020–2022 [+8.6 ± 2.6 teragrams of CH 4 per year (TgCH 4 year −1 )] and then decreased between 2022 and 2023 (−9.9 ± 3.3 TgCH 4 year −1 ). Most emission changes from 2019 to 2023 occurred in northern tropical wetlands in Africa and Asia, whereas South American wetlands emissions declined and Arctic emissions increased after 2019.

Secondary representations enable our minds to depart from the here-and-now and generate imaginary, hypothetical, or alternate possibilities that are decoupled from reality, supporting many of our richest cognitive capacities such as mental-state attribution, simulation of possible futures, and pretense. We present experimental evidence that a nonhuman primate can represent pretend objects. Kanzi, a lexigram-trained bonobo, correctly identified the location of pretend objects (e.g., “juice” poured between empty containers), in response to verbal prompts in scaffolded pretense interactions. Across three experiments, we conceptually replicated this finding and excluded key alternative explanations. Our findings suggest that the capacity to form secondary representations of pretend objects is within the cognitive potential of, at least, an enculturated ape and likely dates back 6 to 9 million years, to our common evolutionary ancestors.

Currently, no effective treatment exists for infertility associated with premature ovarian insufficiency (POI) because affected patients lack hormone-responsive antral follicles. By screening a Food and Drug Administration (FDA)–approved drug library, we identified finerenone, a kidney disease medication, as a promising drug for restoring fertility in POI. Finerenone stimulated follicle development in aged mice and restored antral follicle development in patients with POI following oral administration, resulting in mature oocytes and embryos. Mechanistically, finerenone reduced fibrotic deposition in the ovarian stroma, alleviating collagen-mediated suppression of follicular development. Building on this insight, we identified additional FDA-approved oral antifibrotic drugs as potential treatments for POI-related infertility. Our findings highlight the ovarian stroma—rather than the follicles themselves—as the key therapeutic target and offer potential therapeutic leads for POI-related infertility.

Priming rare subdominant precursor B cells in germinal centers (GCs) is a central goal of vaccination to generate broadly neutralizing antibodies (bnAbs) against HIV. Multivalent immunogen display on protein nanoparticle scaffolds can promote such responses, but it also generates scaffold-specific B cells that could theoretically limit bnAb precursor expansion in GCs. We rationally designed DNA origami–based virus-like particles (DNA-VLPs) displaying a germline-targeting HIV envelope protein immunogen, which elicited no scaffold-specific antibody responses. Compared with a state-of-the-art clinical protein nanoparticle, these DNA-VLPs increased the expansion of epitope-specific GC B cells relative to off-target B cells and enhanced expansion of bnAb-lineage B cells in a humanized mouse model of CD4 binding site priming. Thus, minimizing off-target responses enhances bnAb priming and indicates that DNA-VLPs are a promising vaccine platform.

The sugarcane genus Saccharum is characterized by complex genomes with diverse ploidy levels. We developed a multiscale graph–based pangenome representation, which integrates nine genome assemblies into a unified reference, representing modern cultivars and founding species. Each homo(eo)logous (encompasses both homologous and homeologous relationships) chromosome set retains 47 to 57 haplotypes and ~74,000 to 271,000 gene alleles. This framework enables multiomics exploration, encompassing homo(eo)log systems and epigenomic signatures. The pangenome facilitates population genomics analyses of 417 mixed-ploidy Saccharum accessions, revealing convergent selection and identifying the Andropogoneae TB1 homolog linked to tillering as a promising gene-editing target to boost cane yield. Additionally, the pangenome supports dosage-informed genome-wide association study, improving heritability estimates and identification of sugar or leaf-angle–associated loci, including SaIRX10 and SaBAK5 . Our analytical framework establishes a foundation for graph-based genetic studies in sugarcane and other polyploid genomes.

Brassica rapa ( Br ) encompasses many morphotypes and subspecies, so it is a good model with which to investigate plant diversification and subspeciation. Here, we resequenced the genomes of 1720 Br accessions and de novo assembled 11 representative telomere-to-telomere gapless genomes for seven elite subspecies that underwent intensive morphotypification and developed distinct agronomic traits valued to agriculture. We identified 6992 unknown genes, 110 complete (peri)centromeres, and five new satellites associated with Br morphotypes and subspecies and Brassica species evolution. The pangenome, built on 11 gapless and 20 published genomes, reveals structural variations and gene diversities among Br subspecies. Pangenome-wide association studies uncovered that the gene BrLH1 controls leaf-head formation. We show that structural changes have occurred in satellites, (peri)centromeres, and genes, contributing to fast subspeciation and morphotypification during the short history of Br cultivation, providing invaluable resources for Brassica breeding.

Accurate semantic segmentation of high-resolution remote sensing imagery is crucial for applications such as land cover mapping, urban development monitoring, and disaster response. However, remote sensing data still present inherent challenges, including complex spatial structures, significant intra-class variability, and diverse object scales, which demand models capable of capturing rich contextual information from both local and global regions. To address these issues, we propose ArgusNet, a novel segmentation framework that enhances multi-scale representations through a series of carefully designed fusion mechanisms. At the core of ArgusNet lies the synergistic integration of Adaptive Windowed Additive Attention (AWAA) and 2D Selective Scan (SS2D). Specifically, our AWAA extends additive attention into a window-based structure with a dynamic routing mechanism, enabling multi-perspective local feature interaction via multiple global query vectors. Furthermore, we introduce a decoder optimization strategy incorporating three-stage feature fusion and a Macro Guidance Module (MGM) to improve spatial detail preservation and semantic consistency. Experiments on benchmark remote sensing datasets demonstrate that ArgusNet achieves competitive and improved segmentation performance compared to state-of-the-art methods, particularly in scenarios requiring fine-grained object delineation and robust multi-scale contextual understanding.

The Lunar Reconnaissance Orbiter Camera (LROC) Narrow Angle Camera (NAC) and Korea Pathfinder Lunar Orbiter (KPLO) ShadowCam provide high-resolution (0.5–2 m per pixel) images of the Moon. These high-resolution images facilitate the creation of highly detailed controlled mosaics, which can be used for applications such as regional geomorphic maps, crater size-frequency distribution analysis, and mission planning. We establish the methodology used to produce most of our LROC NAC and ShadowCam regional controlled mosaics, conduct an analysis of the accuracy of our controlled mosaics, and discuss the utility of these products. This accuracy analysis includes a comprehensive analysis of the bundle adjustment results for both our LROC NAC and ShadowCam controlled mosaics and a comprehensive analysis of the positional accuracy of our LROC NAC controlled mosaics, with median positional offsets of our NAC controlled mosaics being <12 m in latitude and <5 m in longitude.

Video synthetic aperture radar could provide more valuable information than static images. However, it suffers from several difficulties, such as strong clutter, low signal-to-noise ratio, and variable target scale. The task of moving target detection is therefore difficult to achieve. To solve these problems, this paper proposes a model and data co-driven learning method called look once on principal components (PC-YOLO). Unlike preceding works, we regarded the imaging scenario as a combination of low-rank and sparse scenes in theory. The former models the global, slowly varying background information, while the latter expresses the localized anomalies. These were then separated using the principal component decomposition technique to reduce the clutter while simultaneously enhancing the moving targets. The resulting principal components were then handled by an improved version of the look once framework. Since the moving targets featured various scales and weak scattering coefficients, the hierarchical attention mechanism and the cross-scale feature fusion strategy were introduced to further improve the detection performance. Finally, multiple rounds of experiments were performed to verify the proposed method, with the results proving that it could achieve more than 30% improvement in mAP compared to classical methods.

Transportation infrastructure is vulnerable to heterogeneous deformation, yet such deformation remains insufficiently monitored and characterized in metropolitan regions due to the lack of high-resolution deformation gradient products and comparison with industrial standards. Here, we generated a 45 m resolution interferometric synthetic aperture radar (InSAR) surface displacement time series across the Beijing Plain using Sentinel-1 SAR imagery acquired between 2014 and 2024, and calculated deformation gradients along all ring roads, major expressways, and airport runways. These deformation gradients are compared with national standards to evaluate their structural risks. Our analysis shows that (1) subsidence in the Beijing Plain is concentrated in the northern, eastern, and southern regions, where the northeastern region has been uplifting since 2018 due to the groundwater recovery in Beijing; (2) all ring roads, expressways, and airport runways are relatively stable during our observation period of 2015–2021, except for the central runway of Beijing Capital International Airport, which has accumulated a deformation gradient of 1.9‰ during 2015–2021, exceeding the safety limit of 1.5‰, indicating structural risks. These results demonstrate the effectiveness of high-resolution InSAR time series for monitoring deformation and pinpointing potential structural risks.

Multi-label remote sensing scene classification (MLRSSC) requires autonomous discovery of all relevant land-cover categories without human guidance. Conventional expert classifiers return only label vectors without spatial evidence, while foundation segmenters (e.g., SAM, RemoteSAM) remain passively dependent on external prompts—misaligned with autonomous interpretation. We introduce SAFI-XRS, a parameter-efficient self-prompted framework that transforms passive prompting into active scene parsing. By training only <2% of a 332M-parameter segmenter (∼2.4M parameters), SAFI-XRS generates class-aligned queries from images via a Semantic Query Generator (SQR), replacing external prompts with self-generated conditioning. A Mask-Guided Classifier (MGC) aggregates spatial evidence into label confidences, enabling mask-based explainability. Experiments on UCM-ML, DFC15-ML, and AID-ML show SAFI-XRS surpasses text-prompted foundation segmenters (+3.9/+3.8 mAP on balanced datasets) while achieving 6.8× parameter efficiency compared to expert models, validating a practical path toward autonomous, explainable RS scene understanding.

Remote sensing solutions for avalanche segmentation and mapping are key to supporting risk forecasting and mitigation in mountain regions. Synthetic Aperture Radar (SAR) imagery from Sentinel-1 can be effectively used for this task, but training an effective detection model requires gathering a large dataset with high-quality annotations from domain experts, which is prohibitively time-consuming. In this work, we aim to facilitate and accelerate the annotation of SAR images for avalanche mapping. We build on the Segment Anything Model (SAM), a segmentation foundation model trained on natural images, and tailor it to Sentinel-1 SAR data. Adapting SAM to our use case requires addressing several domain-specific challenges: (1) domain mismatch, since SAM was not trained on satellite or SAR imagery; (2) input adaptation, because SAR products typically provide more than three channels while the SAM is constrained to RGB images; (3) robustness to imprecise prompts that can affect target identification and degrade the segmentation quality, an issue exacerbated in small, low-contrast avalanches; and (4) training efficiency, since standard fine-tuning is computationally demanding for the SAM. We tackle these challenges through a combination of adapters to mitigate the domain gap, multiple encoders to handle multi-channel SAR inputs, prompt-engineering strategies to improve avalanche localization accuracy, and a training algorithm that limits the training time of the encoder, which is recognized as the major bottleneck. We integrate the resulting model into a segmentation tool and show experimentally that it speeds up the annotation of SAR images.

This study, conducted in the framework of the LIAISE field campaign in NE Spain (May–September 2021), investigates how near-surface relative humidity influences early-stage rainfall characteristics when precipitation is most affected by temperature and relative humidity before rainfall onset. Two instrumented sites were examined, using disdrometers, Micro Rain Radar (MRR), C-band weather radar data, and automatic weather stations. Rainfall events were first classified as stratiform or convective using weather radar data based on a texture analysis of the reflectivity field. Then, only stratiform events were selected and further classified into dry and moist categories according to the upper and lower terciles of near-surface (2 m) relative humidity at the rainfall onset (dry < 54%; moist > 72%). Results show that during dry events, the time delay between the detection of precipitation at ~750 m above ground level (AGL) (by MRR or C-band radar) and its arrival at the surface (measured by the disdrometer) is consistently longer than during moist events, indicating possible evaporation of raindrops during their descent. Surface drop size distributions also differ: dry cases have generally fewer small drops (with diameters < 0.8 mm) but relatively more large drops, leading to higher radar reflectivity values despite similar surface rainfall amounts. However, reflectivity observed aloft by C-band radar and MRR does not present the dependence on relative humidity found at ground level. Findings reported here increase our understanding of the impact of low-level conditions on precipitation characteristics and microphysical associated processes and may contribute to improve correction schemes in operational weather radar quantitative precipitation estimates.

Transformer-based deep learning techniques have recently shown outstanding potential in remote sensing scene classification (RSSC), benefiting from their ability to capture global semantic relationships and contextual dependencies. However, effectively utilizing the raw image and global semantic information while simultaneously taking into account detailed features and multi-scale spatial relationships remains a major challenge. Therefore, this paper proposes a novel FG-Swin KANsformer model that integrates frequency domain and gradient prior information from raw images with the Kolmogorov–Arnold Network (KAN) to enhance nonlinear feature modeling. The FG-Swin KANsformer consists of three key components: the Discrete Cosine Transform (DCT) module, the gradient-spatial feature extraction (GSFE) module, and the Swin Transformer module integrated with KAN. In the feature embedding phase, the DCT module extracts frequency domain features, while the GSFE module uses multi-scale convolutions and Sobel operators to extract spatial structures and gradient information at different scales, thereby enhancing the utilization of the original image’s frequency domain and gradient prior information. In the Swin Transformer feature modeling phase, the conventional multilayer perceptron (MLP) in Swin Transformer Blocks is replaced by KAN, which decomposes complex multivariate functions into univariate compositions, thereby improving nonlinear representation capacity and enhancing feature discrimination. The thorough experiments on three distinct public remote sensing (RS) datasets demonstrate that FG-Swin KANsformer exhibits outstanding performance.

Deep learning–based pansharpening methods for remote sensing have advanced rapidly in recent years. However, current methods still face three limitations that directly affect reconstruction quality. Content adaptivity is often implemented as an isolated step, which prevents effective interaction across scales and feature domains. Dynamic multi-scale mechanisms also remain constrained, since their scale selection is usually guided by global statistics and ignores regional heterogeneity. Moreover, frequency and spatial cues are commonly fused in a static manner, leading to an imbalance between global structural enhancement and local texture preservation. To address these issues, we design three complementary modules. We utilize the Adaptive Convolution Unit (ACU) to generate content-aware kernels through local feature clustering, thereby achieving fine-grained adaptation to diverse ground structures. We also develop the Multi-Scale Receptive Field Selection Unit (MSRFU), a module providing flexible scale modeling by selecting informative branches at varying receptive fields. Meanwhile, we incorporate the Frequency–Spatial Attention Unit (FSAU), designed to dynamically fuse spatial representations with frequency information. This effectively strengthens detail reconstruction while minimizing spectral distortion. Specifically, we propose the Hybrid Attention Fusion Network (HAFNet), which employs the Hybrid Attention-Driven Residual Block (HARB) as the fundamental utility to dynamically integrate the above three specialized components. This design enables dynamic content adaptivity, multi-scale responsiveness, and cross-domain feature fusion within a unified framework. Experiments on public benchmarks confirm the effectiveness of each component and demonstrate HAFNet’s state-of-the-art performance.

The concentrations of carbon, nitrogen, and phosphorus in water bodies significantly influence aquatic ecological conditions. By collecting multitemporal hyperspectral data and water quality parameter data from water bodies and through systematic preprocessing of hyperspectral data combined with multimethod sensitive band selection, an optimal spectral feature subset was determined. Within a machine learning framework, multiple combined remote sensing inversion models were constructed to identify the optimal inversion model for each water quality parameter, along with corresponding preprocessing methods and sensitive bands. The results indicate that differential processing of remote sensing reflectance enhances model accuracy. Sensitive band selection effectively eliminates redundant bands, significantly improving the computational efficiency of inversion models. XGBoost demonstrated superior accuracy in constructing 240 water quality parameter inversion models because of its unique algorithmic design. However, model accuracy is not solely determined by algorithmic complexity or predictive capability but rather by the combined effect of algorithm performance and input feature quality. Verification of the inversion model’s generalization ability via an independent dataset demonstrated its capacity for generalization. These findings provide valuable insights for the reliable application of hyperspectral data in aquatic environmental remote sensing and offer support for regional water quality conservation efforts.

Landslide susceptibility maps (LSMs) are crucial for risk mitigation, but integrating Multi-temporal Interferometric Synthetic Aperture Radar (MT-InSAR) data is often hampered by a lack of physical interpretation. To address this issue, this study proposes an enhanced modeling framework that integrates multi-source monitoring data by coupling dynamic deformation features. Ground deformation velocity is obtained using MT-InSAR and embedded as dynamic physical constraints into the loss function of a Multi-Layer Perceptron (MLP) model. This approach enables the joint optimization of static geological factors and dynamic deformation characteristics in landslide susceptibility prediction. The proposed framework was applied to Zunyi City, Guizhou Province, China, utilizing an inventory of landslide hazard sites and a dataset of 16 susceptibility factors for model training and evaluation. The results demonstrated that the dynamically constrained model significantly improved predictive performance (AUC = 0.976, an increase of 0.032 compared to the baseline model), and enhanced spatial consistency, reflected by an average increase of 0.0184 in predicted susceptibility for inventoried landslide hazard sites. The framework also outperformed other conventional machine learning models across multiple evaluation metrics. Furthermore, SHAP (SHapley Additive exPlanations) analysis revealed that slope (18.68%), DEM (13.26%), rainfall (11.57%), and mining activities (8.79%) were the primary contributing factors in high-susceptibility areas. This study offers a physically interpretable and robust methodology that advances landslide risk assessment and contributes to disaster prevention strategies.

The lack of high-precision imaging data for lunar volcanic regions currently hinders the detailed characterization of lava tube systems and their associated fine-scale geomorphology. To address this information deficit, this study establishes the Jingpo Lake Volcanic Field (JLVF) in Northeast China as a primary terrestrial analog for the lunar Marius Hills complex. We systematically characterize the basaltic morphometric continuum, tracing the geological evolution from proximal scoria cones through medial lava tube skylights to distal lava plateaus. Focusing on the subsurface transport system, we identify a linear chain of discontinuous skylights that structurally mirrors the “proto-rille” stage of lunar sinuous rilles. Quantitative morphometry reveals that these terrestrial vents reproduce the geometric duality of lunar pits, ranging from stable “deep shafts” to degraded “funnel pits,” effectively validating the mechanical diversity of the lunar inventory. Critically, the “U-to-V” cross-sectional transition observed in JLVF collapse trenches serves as diagnostic ground-truth evidence, confirming that lunar rilles originate from the catastrophic roof failure of subsurface tubes rather than purely thermal erosion. Regarding the lava plateau, our field investigation resolves sub-meter micro-textures—including laminar pahoehoe ropes and inflation fissures—that are typically obscured by the resolution limits of current lunar orbiters. These findings suggest that the seemingly “smooth” lunar maria likely host complex, rugged micro-terrains. Therefore, comparing lunar volcanic regions with simulated volcanic fields from Earth is crucial. Analyzing potential volcanic products from angles undetectable by some lunar satellites can offer vital insights for future lunar exploration.

Evapotranspiration (ET) links the water cycle with the energy balance and serves as a key driving process for ecosystem functioning and water resource management. Canopy conductance (Gc) plays a central role in regulating transpiration, but many models inadequately represent its regulatory mechanisms and show varying applicability across different land cover types. This study develops a remote-sensing ET estimation approach suitable for large scales and diverse land cover types and proposes an improved canopy conductance model for daily latent heat flux (LE) estimation. By integrating the canopy radiation transfer concept from the K95 model into the multiplicative Jarvis framework, an improved canopy conductance model is developed that includes limiting effects from photosynthetically active radiation (PAR), vapor pressure deficit (VPD), air temperature (T), and soil moisture (θ). Eighteen combinations of limiting functions are designed to evaluate structural performance differences. Using observations from 79 global flux sites during 2015–2023 and integrating multi-source datasets, including ERA5, MODIS, and SMAP, a two-stage parameter optimization was applied to determine the optimal limiting function combination for each land cover type. And nine sites from nine different land cover types were selected for independent spatial validation. Temporal validation within the optimization sites shows that, at the daily scale, the model achieves a Kling–Gupta efficiency (KGE) of 0.82, a correlation coefficient (R) of 0.82, and a Root Mean Square Error (RMSE) of 27.83 W/m2, demonstrating strong temporal stability. Spatial validation over independent holdout sites achieved KGE = 0.84, R = 0.84, and RMSE = 22.53 W/m2. At the 8-day scale, when evaluated over the holdout sites, the model achieves KGE = 0.87, R = 0.88, and RMSE = 18.74 W/m2. Compared with the K95 and Jarvis models, KGE increases by about 34% and 15%, while RMSE decreases by about 38% and 12%, respectively. Relative to the MOD16 and PML-V2 products, KGE increases by about 32% and 16%, while RMSE decreases by about 33% and 17%, respectively. Comprehensive comparisons show that explicitly coupling canopy structure with multiple environmental constraints within the Jarvis framework, together with structure optimization across land cover types, can markedly improve large-scale remote-sensing ET retrieval accuracy while maintaining physical consistency and physiological rationality. This provides an effective pathway and parameterization scheme for producing ET products applicable across ecosystems.

Accurate fine-scale forest mapping is fundamental for ecological monitoring and resource management. While deep learning semantic segmentation methods have advanced the interpretation of high-resolution UAV imagery, their generalization across diverse forest regions remains challenging due to high spatial heterogeneity. To address this, we propose two enhanced versions based on the PP-LiteSeg architecture for robust cross-regional forest segmentation. Version 01 (V01) integrates a multi-branch attention fusion module composed of parallel channel, spatial, and pixel attention branches. This design enables fine-grained feature enhancement and precise boundary delineation in structurally regular artificial forests, such as the Huayuan Forest Farm. As a result, V01 achieves a mIoU of 92.64% and an F1-score of 96.10%, representing an approximately 18 percentage-point mIoU improvement over PSPNet and DeepLabv3+. Building on this, Version 02 (V02) introduces a lightweight residual connection that directly shortcuts the fused features, thereby improving feature stability and robustness under complex textures and illumination, and demonstrates stronger performance in naturally heterogeneous forests (Longhai Township), attaining an mIoU of 91.87% and an F1-score of 95.77% (5.72 percentage-point mIoU gain over DeepLabv3+). We further conduct comprehensive comparisons against conventional CNN baselines as well as representative lightweight and transformer-based models (BiSeNetV2 and SegFormer-B0). In bidirectional cross-region transfer (train on one region and directly test on the other), V02 exhibits the most stable performance with minimal degradation, highlighting its robustness under domain shift. On a combined cross-regional dataset, V02 achieves a leading mIoU of 91.50%, outperforming U-Net, DeepLabv3+, and PSPNet. In summary, V01 excels in boundary delineation for regular plantation forests, whereas V02 shows more stable generalization across highly varied natural forest landscapes, providing practical solutions for region-adaptive UAV forest segmentation.