Few-shot aircraft detection in remote sensing imagery still challenges modern deep learning-based detectors with limited annotated priors. While existing strategies (e.g., data augmentation and transfer/meta learning) could alleviate the serious overfitting to a large extent, the adopted generic inductive bias, rather than aircraft-specific knowledge, does not change the inherently data-hungry property of deep representation learning. In this letter, we propose a grammar-guided detection framework that incorporates high-level structural knowledge with three kinematic grammars of semantic landmarks: symmetry, parallelism, and verticality. These grammatical rules are embedded in a differentiable energy function for geometry-constrained landmark inference. Besides, we incorporate the grammar into the Faster RCNN framework as top-down priors to filter the conventional bottom-up message passing process. Specifically, the top-down landmark-aware attention refines feature responses within ROIs, while the bottom-up scale-aware attention strengthens backbone features by learning multiscale representations. These modulated features are further fed into a contour-aware detection head to predict landmark heatmaps and reconstruct aircraft outlines. The experiments conducted on an enhanced MAR20 benchmark with newly annotated landmarks demonstrate the effectiveness of our method under few-shot settings.

The semantic segmentation of high-resolution remote sensing imagery frequently encounters difficulties stemming from intricate scene layouts and significant variations in object scales. To tackle these issues, we propose a global-local Fourier-aware network (GLF-Net). The core innovation exploits Fourier transforms to decouple features for separate enhancement of global consistency and local details. Specifically, the model consists of three key components: 1) the global Fourier-aware module (GFAM) selectively extracts and fuses low-frequency components to suppress background interference and improve scene-level semantic consistency; 2) the multiscale local fusion (MSLF) module aggregates high-frequency components with multiscale local information to strengthen boundaries and fine-grained structures, thereby alleviating intraclass variations caused by scale differences; and 3) a vHeat backbone is adopted to efficiently capture long-range spatial dependencies among geographic entities. Extensive evaluations conducted on the ISPRS Vaihingen and Potsdam benchmarks indicate that our GLF-Net yields superior results compared with existing methods. These findings confirm the model's robustness in handling precise segmentation tasks, even when faced with intricate multiscale variations arising from object size disparities.

Class-incremental learning (CIL) is of significant value for dynamic Earth observation applications using hyper-spectral images (HSIs). However, existing methods addressing catastrophic forgetting often fail to simultaneously handle limited storage budgets, inter-class semantic ambiguity, and the effective balance between new and old knowledge. To systematically address these challenges, this article proposes a novel class-incremental learning framework based on text guidance and dual distillation (TGDD). First, a distribution-aware sample selection strategy is designed to address the conflict between the inherent intra-class spectral variability of HSI data and limited storage budgets. This strategy adaptively discovers the intrinsic substructures of each class and ensures they are effectively represented in the exemplar set, thereby maximizing the fidelity of historical knowledge under strict storage constraints. Second, a dual distillation mechanism is constructed to achieve a balance between knowledge “retention” and “learning”. Specifically, inter-task knowledge distillation transfers the discriminative ability of the old model to the current one without disrupting consolidated knowledge. Meanwhile, intra-task hierarchical self-distillation promotes collaborative learning between deep and shallow layers, ensuring consistent and highly discriminative feature representations. Furthermore, to tackle misclassification caused by spectrally similar classes across new and old tasks, a multimodal semantic guidance module is introduced. Based on text embeddings of class names, this module dynamically adjusts feature learning and decision boundaries, enhancing the model’s ability to distinguish between spectrally similar but semantically different classes. Extensive experiments on four public HSI datasets demonstrate that TGDD achieves a superior balance between stability (old knowledge) and plasticity (new knowledge). Its classification accuracy and robustness significantly outper-form current state-of-the-art methods. The code is available at https://github.com/Li-ZK/TGDD.

Electronic equipment used for videotaping is often not well imaged at night. Infrared-visible image fusion and image brightness enhancement are two important methods used in nighttime imaging. Existing methods do not unite these two approaches appropriately. In order to cross the gap between them, this paper proposes a multi-task fusion framework (MTF-use), which combines brightness enhancement and fusion tasks into an end-to-end network to solve the environment awareness problem in dark light. Firstly, a fast feature extraction module is constructed based on the reparameterisation theory. The module provides a rich stream of information to the network during training and degenerates losslessly into a convolutional layer during testing. Next, the brightness enhancement module is designed based on Retinex theory as the basic unit of the feature extraction part. Ultimately, the labels are generated using the brightness enhancement algorithm for training. A large number of experiments show that the proposed method has achieved the most advanced performance in vision and quantitative comparisons. In addition, the proposed method also shows great application prospects in the field of multispectral detection. Our code is available at https://github.com/UAVSwarm/MT-fuse-code.

In recent years, hyperspectral image (HSI) classification has garnered widespread attention in remote sensing applications. However, the scarcity of labeled training samples makes class prototypes susceptible to noise and data distribution shifts, which degrades the model’s generalization performance. Furthermore, the insufficiency of labeled samples results in blurred decision boundaries between different classes, often causing misclassification. To address these issues, this paper proposes a multi-level knowledge distillation framework for cross-domain few-shot classification (MLKD-CFSC). The framework consists of two core modules: the Feature Distillation (FD) module and the Decision Distillation (DD) module. The Feature Distillation module optimizes the student network’s prototype features by leveraging those from a teacher network, which is pre-trained on a large-scale hyperspectral dataset. This process enhances the discriminability and robustness of the features extracted by the student network. Meanwhile, the Decision Distillation module constructs a knowledge distillation method based on the soft labels of the teacher network, guiding the student network to learn more refined classification decision boundaries by uncovering the potential similarities between categories. Experimental results on two public hyperspectral datasets, Houston (HT) and WHU-Hi-HanChuan (HC), demonstrate that the proposed MLKD-CFSC framework significantly outperforms existing cross-domain few-shot methods, improving the Overall Accuracy (OA) over the best-performing baseline by 4.00% and 4.47%, respectively. The code is available at https://github.com/Li-ZK/MLKD-CFSC-2026.

Open-set hyperspectral image classification requires a model to recognize known land-cover categories while rejecting samples from unseen classes. This setting is difficult because hyperspectral images have high spectral dimensionality, limited annotations, large intra-class variation, and strong inter-class similarity. Many existing methods rely on closed-set confidence scores or combine heuristic anomaly cues, which makes the rejection boundary sensitive to dataset-specific score distributions. This paper proposes a Dual-Space Calibrated Prototype Network, called DSCP-Net, for open-set hyperspectral image classification. DSCP-Net learns two complementary feature spaces: one for local spectral-spatial structure and one for high-level semantic information. It uses dual-space multi-prototype representations to describe the multi-mode structure of known classes, and employs a lightweight mixture-of-experts module for sample-adaptive feature transformation. For unknown detection, DSCP-Net adopts a distance-first inference rule in which calibrated prototype distances in the two spaces define the unknown score. A validation-based global calibration strategy determines the rejection threshold from known-class validation statistics. Experiments on four public hyperspectral datasets show that DSCP-Net improves the balance between known-class classification and unknown-class rejection over the compared closed-set and open-set methods. Further analysis shows that calibrated prototype-distance inference provides the clearest and most consistent contribution to the open-set gain, whereas the expert transformation and differentiable auxiliary objectives have smaller, dataset-dependent effects. The code is available at https://github.com/AIYAU/DSCP-Net.

Addressing the critical challenge of sample scarcity in hyperspectral image classification, this paper presents a semantic-anchor-enhanced prototype learning (SAEPL) framework that is specifically tailored for few-shot, cross-domain expert systems. Unlike conventional prototype methods that rely solely on the mean of a handful of support samples, SAEPL introduces trainable, orthogonally-initialized semantic anchors to explicitly capture inter-class relationships and guide prototype generation. These anchors are jointly optimized with a lightweight contrastive-learning objective that aligns class prototypes across source and target domains, thereby dramatically improving generalization under heterogeneous distributions. In addition, a Dynamic Adaptive Spatial Occlusion data-augmentation module is embedded to simulate real-world spatial degradations, further boosting robustness without extra annotation cost. Extensive experiments on four public hyperspectral datasets (Chikusei -> Indian Pines, Houston, Salinas) demonstrate that SAEPL attains overall accuracies of 80.49%, 81.35%, and 93.22%, outperforming eight state-of-the-art algorithms by significant margins while reducing parameter count by up to 90%. Ablation studies and t-SNE visualizations confirm that both semantic anchors and Dynamic Adaptive Spatial Occlusion are indispensable for achieving compact, discriminative feature clusters. The source code is available at: https://github.com/AIYAU/SAEPL.

Vision-language multimodal learning has exhibited remarkable advantages in few-shot hyperspectral image classification, where prompt learning effectively enhances feature-extraction accuracy and representation quality by guiding the model to focus on critical information. However, static prompts lack flexibility while dynamic prompts suffer from unstable generation. To address these issues, this paper proposes a Dual Prompt-Driven Cross-Modal Fusion Learning method (DPCFL). Specifically, we utilize static prompt templates to provide stable prior guidance for the model.Simultaneously, a unique learnable prompt vector is designed for each category, which operates independently of the pretrained model's input to effectively mitigate potential influences from prior semantics. To alleviate semantic bias, we further design a multi-loss joint optimization strategy incorporating parameter space, feature space, and cross-modal space consistency constraints, thereby improving the robustness of class-level prototype features. Additionally, to tackle data scarcity, a cross-domain collaborative training mechanism is introduced to facilitate knowledge transfer. The experimental results on multiple standard hyperspectral image datasets confirm its superior classification performance under both few-shot and cross-domain scenarios. The related code will be made publicly available at the following URL: https://github.com/AIYAU/DPCFL.

Hyperspectral imaging technology provides an effective means for non-destructive detection of soluble solids content (SSC) in blueberries. However, multiple interference factors, including the diversity of fruit biological characteristics and differences during the data acquisition process, significantly affect the predictive performance of models. To address this challenge, this study proposes a differential enhancement and multi-scale feature fusion framework. This framework extracts gradient dynamic features and curvature characteristics of spectral curves through a multi-order differential feature enhancement module, achieving hierarchical fusion of static spectral and local dynamic features, thereby improving the model’s robustness against multiple interference factors. Simultaneously, by combining a multi-scale feature extraction network, it promotes deep fusion of information at different scales and enhances feature representation capabilities. Experimental results on 13 sets of blueberry datasets demonstrate that the proposed method significantly outperforms existing models in SSC prediction performance, validating its effectiveness and superiority.

To meet the demands for fusion efficiency and perceptual performance of infrared and visible images on resource-constrained unmanned aerial vehicle (UAV) platforms, this paper proposes a text-guided lightweight multimodal image fusion network with heterogeneous encoders. The network employs a lightweight dual-branch heterogeneous encoding architecture designed to complementarily represent infrared and visible image information. Specifically, the infrared-encoding branch emphasizes thermal targets and edge responses, while the visible-encoding branch focuses on modeling texture and detail information. This design effectively avoids feature redundancy and performance bottlenecks commonly associated with homogeneous encoders. To enhance the collaborative representation capability of multimodal features, a lightweight cross-modal attention fusion module is introduced. This module jointly models attention relationships across channel and spatial dimensions, thereby strengthening complementary information interactions between modalities. Furthermore, leveraging semantic features extracted from the pre-trained vision–language model CLIP, the fusion process incorporates explicit semantic prior guidance. Through hierarchical feature-level modulation, the weights of infrared and visible features are dynamically adjusted, improving the semantic consistency and environmental adaptability of the fused images. Systematic comparative experiments and comprehensive evaluations were conducted on three publicly available multimodal image datasets: TNO, LLVIP, and M3FD. The proposed method was compared with nine representative image fusion algorithms. Results demonstrate that the proposed network achieves state-of-the-art performance on multiple mainstream evaluation metrics, including mutual information and structural similarity. The fused images surpass existing methods in detail clarity, edge structure consistency, and target discernibility. Ablation studies further show that the inference time of the proposed model is reduced by approximately 50% compared with baseline methods, achieving higher efficiency without significant performance degradation. In addition to quantitative evaluations, qualitative experiments guided by textual instructions were performed, demonstrating the model’s strong semantic responsiveness and content adaptability. For example, in low-light enhancement tasks, the model significantly improves brightness and visibility of fused images, highlighting thermal sources present in infrared images. However, in daytime scenarios, this instruction may cause excessive brightness in background areas, reflecting the model’s selective sensitivity to semantic inputs. In overexposure correction tasks, the model preserves thermal contrast features in infrared images, reducing interference from overexposed regions in visible images and producing fusion outcomes dominated by infrared characteristics. Under conditions of low contrast in infrared images, the model enhances texture and detail information from visible images, generating fusion results closer to the visual appearance of the visible modality. In noise-robustness tasks, when visible images are corrupted by noise, the model preferentially leverages the stable structural information from infrared images to reconstruct fused outputs, effectively mitigating noise-induced degradation and demonstrating strong anti-interference capability. In summary, the proposed model integrates heterogeneous dual-branch encoding with cross-modal attention and semantic guidance to achieve improved fusion quality and adaptability on resource-limited UAV platforms. Experimental results confirm that it can dynamically adjust fusion strategies based on different semantic inputs, enhancing the consistency and relevance of fused images for various tasks. Moreover, the model achieves a favorable balance between computational efficiency and fusion performance, making it suitable for practical deployment in complex environments.

Abstract Climate-change-driven drought intensification increasingly threatens forest ecosystems, highlighting an urgent need for accurate monitoring of forest water stress. Leaf water potential (Ψleaf) is a key integrative indicator, yet conventional measurements are destructive and unsuitable for large-scale or high-frequency monitoring. Hyperspectral remote sensing offers a promising alternative, but robust canopy-level Ψleaf estimation remains constrained by limited labeled data and heterogeneous environmental conditions. Here, we develop a cross-scale framework integrating supervised contrastive learning with deep transfer learning to translate robust leaf-scale pretraining into canopy-scale Ψleaf estimation from hyperspectral data in a Populus × euramericana ‘I-214’ plantation. Hyperspectral imagery was captured at the leaf scale under controlled laboratory conditions (n = 229) and at the canopy scale using a UAV-based platform (n = 200), together with paired Ψleaf measurements. Reflectance consistently increased with declining Ψleaf at both scales, supporting the feasibility of cross-scale modeling. At the leaf scale, physics-consistent spectral augmentation coupled with contrastive learning enhanced feature discrimination and predictive stability under small-sample conditions (R2 = 0.8030). Transfer learning via progressive fine-tuning enabled efficient scaling of the leaf-level pretrained model to canopy-level prediction despite structural and environmental complexity and restricted field data ranges, achieving R2 = 0.7605 and RMSE = 0.1056 MPa. Coupling with individual-tree crown segmentation further enabled spatially explicit mapping of canopy Ψleaf and plot-level forest water stress dynamics. These results demonstrate that combining contrastive representation learning with cross-scale transfer provides a practical pathway for physiological monitoring and scalable, climate-smart forest phenotyping in data-constrained forested environments.

Unsupervised domain adaptation (UDA) has become an important approach to address spectral drift in cross-scene hyperspectral image (HSI) classification. However, existing methods suffer from two key defects in learning target domain features. First, without label supervision in the target domain, the model struggles to learn discriminative features and tends to overfit to the specific statistical properties of the source domain, resulting in insufficient generalization to the target domain. Second, while existing domain alignment methods can narrow the distribution gap between the source and target domains, they lack explicit constraints on the inter-class structure of the target domain, leading to blurred boundaries between different classes in the feature space and insufficient inter-class separability. To address these issues, this letter proposes a UDA framework based on masked robustness learning and class structure regularization (MRSR-UDA). This framework uses a masked robustness learning (MRL) module to randomly mask pixels to create difficult samples with incomplete information. By using a combination of strong and weak constraints to augment view prediction consistency, MRSR-UDA forces the model to learn robust discriminative features from the context of spectral-spatial neighborhoods, thereby improving generalization to the target domain. Furthermore, a class structure regularization (CSR) module is introduced to constrain the inter-class correlation matrix to remain consistent before and after data augmentation from a global class relationship perspective, explicitly optimizing inter-class separability in the target domain. The two modules are collaboratively optimized at the sample-level and the class-level to improve feature discriminability and structural rationality of the feature space, respectively. Experiments on two cross-domain HSI datasets verify the effectiveness of the proposed method. Our source code is available at https://github.com/Li-ZK/MRSR-UDA-2025

Hyperspectral imaging technology enables rapid and non-destructive assessment of fruits. However, acquiring a large-scale and well-balanced labeled dataset remains a significant challenge, primarily due to destructive reference measurements and difficulty sourcing samples that cover a wide and balanced distribution of quality attributes, which exacerbates overfitting risk when sample diversity is limited. Additionally, most deep learning methods optimize using mean squared error and regularization losses without explicitly learning feature representations relevant to regression tasks, making it difficult to learn robust, continuous feature representations. To address these challenges, we propose a spectral quantitative regression framework with supervised contrastive learning (SQR-SCL). A C-Mixup method selects sample pairs based on target-value similarity and interpolates their features and target values, enhancing sample diversity. A differential preservation and continuity feature learning (DPCFL) module learns continuous feature representations while preserving differences between samples. The DPCFL module draws samples with similar target values closer together and pushes those with differing values further apart, sorting features in the embedding space by target value. Such ordering prevents the model from making significantly different predictions for similar inputs, thereby improving prediction stability. A projection head in the DPCFL module ensures that extracted features preserve inter-sample differences and enhances their discriminative power. In this study, reflectance hyperspectral images and soluble solids content (SSC, degrees Bx) were acquired for 1010 blueberries in five sample groups. On this dataset, SQR-SCL achieved a test R2 of 0.884 and an RMSEP of 0.641 (degrees Bx), outperforming compared deep learning and traditional methods.

Prototype-driven few-shot learning (FSL) holds promise for cross-domain hyperspectral image (HSI) classification but encounters two key limitations: 1) over-reliance on visual features overlooks deep semantic information, hindering the differentiation of visually similar categories and 2) inade quate semantic connections between local and holistic features restrict the utilization of spatial-spectral characteristics. To tackle these issues, we propose a multimodal prototype correction and multidimensional knowledge distillation framework for cross-domain few-shot hyperspectral image classification MCMD-CFSC. This framework implements a multimodal prototype correction strategy that leverages pre-trained language models to extract spatial and spectral text features from spatial and spectral text descriptions, thereby aligning image spatial prototypes with spatial text features and spectral prototypes with spectral text features to enhance the recognition of diverse ground objects. In addition, a multidimensional knowledge distillation mechanism is developed to ensure semantic consistency between local and holistic features across spatial-spectral dimensions, bolstering the model's discriminative power and generalization ability. Experimental results demonstrate that MCMD-CFSC significantly surpasses existing deep learning and FSL techniques on four hyperspectral datasets. The code is available at https://github.com/Li-ZK/MCMD-CFSC-2025.

Multimodal learning has demonstrated outstanding performance in various fields, including visual tasks, due to its ability to integrate information from different data sources. In the field of hyperspectral image (HSI) classification, although existing research can handle complex image data, the utilization of semantic information is insufficient. Over-reliance on a single image modality overlooks the synergistic effects of cross-modal information, thereby limiting the model's performance in category center representation and discrimination, especially in scenarios with scarce samples. To address these challenges, this article proposes a dual-prototype learning with multisemantic fusion (DPL-MSF) method. Specifically, within the DPL-MSF framework, precise text prototypes are constructed by leveraging category textual information and prompt learning strategies to capture the first layer of semantic information. In addition, soft labels (SLs) are introduced into the image prototypes, and a feature fusion module is designed to deeply integrate SL information, thereby obtaining another layer of semantic information. The primary role of the text prototype is to assist in the generation of image prototypes before the introduction of SLs, through contrastive learning methods. Under the synergistic effect of multiple semantic information, the final class-level prototypes generated are more representative and discriminative, significantly enhancing the performance of few-shot learning. Experiments conducted on multiple standard HSI datasets have shown that DPL-MSF has significant advantages in scenarios with scarce samples and cross-domain settings.

The information obtained from a single sensor is often insufficient for accurate environmental perception. In contrast, multi-sensor fusion significantly enhances perceptual accuracy, with the integration of infrared and visible light sensors being one of the most common approaches in multi-sensor fusion. Infrared and visible image fusion aims to combine the texture details of visible images with the thermal radiation information of infrared images to produce fused images with enhanced visual quality, which find broad applications in night vision, surveillance, and target detection. Existing fusion methods primarily focus on improving the quality of fused images, often at the expense of computational efficiency due to the reliance on complex, multi-layered architectures. This limitation makes them unsuitable for deployment on low-power terminal devices. To address these challenges, this paper proposes a novel image fusion framework, termed TSFusion, which leverages teacher-student learning to achieve both high-quality fusion and high computational efficiency. A teacher network, equipped with sequence-model-based feature extraction blocks and progressive integration modules, generates high-quality fusion labels by overcoming the limited receptive field of conventional convolutional networks. The lightweight student network adopts a reparameterisation-based Fast Inception Module, ensuring robust fitting capabilities while maintaining high-speed operation. Additionally, a novel multi-scale loss function is introduced to guide the student network in learning from the teacher at both semantic and pixel levels. Experimental results on three benchmark datasets demonstrate that the proposed TSFusion framework achieves superior visual and quantitative performance compared to state-of-the-art methods. Furthermore, TSFusion exhibits exceptional inference speed, outperforming all existing fusion algorithms, making it highly suitable for real-time applications. The implementation of TSFusion is publicly available at https://github.com/UAVSwarm/TSFusion.

Closed set domain adaptation (CSDA) assumes identical class sets between source and target domains and is an important solution for reducing domain bias. Compared with CSDA, open set domain adaptation (OSDA) is closer to real-world applications by allowing unknown class samples in the target domain. In addition, previous OSDA methods mainly rely on similarity detection between the target and source domains to identify unknown classes, which does not fully capture the characteristics of the target domain. To address these limitations, this letter proposes an OSDA method integrating entropy-guided weighted adversarial (EGWA) networks and contrastive self-supervised learning for hyperspectral image (HSI) classification. The approach introduces an EGWA network to distinguish between known and unknown classes in the target domain, while weighing their importance for aligning the feature distributions. Contrastive self-supervised learning is introduced to learn the intrinsic structure and discriminative features of the target domain from unlabeled target domain data. Experimental validation on two HSI cross-domain datasets demonstrates significant performance improvements over existing methods. The code is available at http://github.com/Li-ZK/EGWA-OSDA-2025

Meta-learning has demonstrated significant potential in addressing the limited annotation data challenge in hyperspectral target detection (HTD). However, existing meta-learning-based methods face two major challenges: 1) weak intertask correlation leading to unstable optimization directions and 2) meta-knowledge adaptation (MKA) based on single prior spectral information fails to effectively characterize spectral variation properties of targets in new scenarios. To overcome these challenges, this letter proposes a cooperative meta-learning (CML) framework with spectral diversity adaptation for HTD. The framework introduces a co-learner through cooperative learning to dynamically capture cross-task knowledge and stabilize optimization directions, while designing a spectral diversity-based MKA strategy to enhance the model's ability to understand the spectral variation characteristics of targets in new scenarios and precisely distinguish the spectral features between targets and backgrounds. Experimental results on two public datasets demonstrate that the proposed method outperforms state-of-the-art HTD algorithms. The code is available at https://github.com/Li-ZK/CMLSDA

Few-shot open-set recognition (FSOSR) for hyperspectral images (HSIs) is a challenging task of great practical value. It aims to classify a sample to one of the known classes by few examples while being able to reject the sample from unknown classes. Prior research mainly focuses on high-level semantic features and ignores mid-level and low-level detailed features. Furthermore, recent endeavors in generative few-shot open-set recognition (FSOSR) have the limitation that the generated samples do not accurately reflect the original data distribution and introduce additional noise. To solve these problems, a novel FSOSR method for HSI based on multilevel feature score learning (MFSL) is proposed, which considers multilevel features and combines semantic information with detailed information. By designing the multilayer score fusion (MSF) module, the similarity of multilevel features and prototypes is converted into scores, so that the score of known samples is lower and the score of unknown samples is higher. The intra-inter similarity contrast enhancement (ISCE) module is also introduced to enhance the discrimination of features, increasing the intraclass similarity while decreasing the interclass similarity by adjusting the similarity of known samples and prototypes. Experiments on two HSI datasets verify the effectiveness of MFSL. The code is available at https://github.com/Li-ZK/MFSL-2025

Enhancing target domain discriminability is a key focus in Unsupervised Domain Adaptation (UDA) for HyperSpectral Image (HSI) classification. However, existing methods overlook bringing similar cross-domain samples closer together in the feature space to achieve the indirect transfer of source domain classification knowledge. To overcome this issue, we propose a Multi-Task Learning-based Domain Adaptation (MTLDA) method. MTLDA incorporates an inductive transfer mechanism into adversarial training, transferring the source classification knowledge to the target representation learning during the process of domain alignment. To enhance the target feature discriminability, we propose utilizing dual-domain contrastive learning to construct related tasks. A shared mapping network is employed to simultaneously perform Source domain supervised Contrastive Learning (SCL) and Target domain unsupervised Contrastive Learning (TCL), ensuring that similar samples across domains are positioned closely in the feature space, thereby improving the cross-scene HSI classification accuracy. Furthermore, we design a feature-level data augmentation method based on feature masking to assist contrastive learning tasks and generate more varied training data. Experimental results obtained from testing on three prominent HSI datasets demonstrate the MTLDA method’s superior efficacy in the realm of cross-scene HSI classification.