Binocular Stereo Vision in Remote Sensing: A Review

Highlights

What are the main findings?

• This study synthesizes the technical progress and inherent limitations of representative models in the field of remote sensing stereo matching.
• The research provides a critical analysis of the characteristics and domain-specific constraints of major remote sensing stereo datasets.

What are the implications of the main findings?

• This study serves as a practical guide for researchers to select the most suitable models and datasets for specific remote sensing stereo matching tasks.
• This survey identifies the unresolved bottlenecks—large disparity ranges, ill-posed regions, cross-sensor domain shift, and label scarcity—that should guide the design of next-generation remote sensing stereo matching algorithms.

Abstract

Stereo vision leverages binocular imagery to emulate the human visual system in perceiving three-dimensional (3D) structures by estimating disparity from rectified image pairs and converting it to depth via geometric triangulation. In recent years, deep learning-based stereo matching has significantly advanced in accuracy, efficiency, and generalization, surpassing traditional methods and demonstrating great potential in remote sensing applications. However, stereo matching in remote sensing faces unique challenges not commonly seen in terrestrial datasets. These include limited access to satellite imagery, seasonal differences between image pairs, difficulty in identifying small objects, and widespread regions with repetitive textures, such as lakes and forests. Unlike prior surveys that primarily address ground-level scenes, this paper presents a comprehensive review of stereo matching techniques tailored for remote sensing. It synthesizes the progress and limitations of representative models, analyzes the characteristics and domain-specific constraints of remote sensing stereo datasets, and outlines future research directions and application prospects in this field.

1. Introduction

Stereo matching using high-resolution satellite imagery is a fundamental technique in remote sensing and photogrammetry. It supports a variety of applications, including 3D reconstruction, terrain modeling, change detection, disaster assessment, image registration, and environmental monitoring [1,2,3,4]. Given a pair of rectified stereo images, the objective is to identify corresponding pixels—$(x_l,y)$ in the left image and $(x_r,y)$ in the right image—and compute the disparity as $d=x_l-x_r$. This disparity is triangulated to infer depth for subsequent remote sensing analysis.

Traditional stereo matching methods typically follow the classic four-step pipeline established by Scharstein and Szeliski [5]: matching cost computation for pixel-wise similarity measurement, cost aggregation to ensure spatial smoothness, disparity computation for initial estimation, and disparity refinement for final optimization. Representative algorithms within this framework include Graph Cuts [6], Semi-Global Matching [7], and PatchMatch [8]. Although these conventional approaches have achieved significant progress, several critical limitations continue to hinder their effectiveness in complex real-world scenarios:

(1)

Empirically designed cost metrics: Traditional matching costs often rely on heuristic metrics, such as luminance difference and correlation coefficients. These hand-crafted measures are often insufficient for capturing complex radiometric variations in high-resolution images.

(2)

Localized cost aggregation: Cost aggregation is typically performed within a finite neighborhood with a fixed window size. This localized approach lacks the flexibility to adapt to varying terrain scales and often fails in regions with repetitive textures or depth discontinuities.

(3)

Dependency on post-processing: Disparity refinement heavily depends on a sequence of manual techniques, such as median filtering for smoothing and subpixel enhancement. These multi-stage pipelines increase computational complexity and require extensive parameter tuning.

Consequently, these hand-crafted schemes often falter in challenging environments and lack the scalability required to process the massive volumes of intricate data characteristic of modern satellite sensors. To transcend these inherent limitations, deep learning-based stereo matching has emerged as a transformative paradigm. By substituting manual feature engineering with end-to-end learnable architectures, these methods have demonstrated superior robustness and accuracy across diverse and demanding applications.

Since 2016, stereo matching methods based on deep learning [9,10,11] have made substantial advances. These methods overcome many of the limitations inherent in traditional algorithms, which often depend on hand-crafted features and rule-based optimization. Current deep models are typically categorized into three major types: cost volume-based methods, iterative optimization networks, and Transformer-based architectures. GC-Net [10] introduced cost volume construction as a central component and inspired numerous variants [11,12]. RAFT [13], originally proposed for optical flow, was later adapted for stereo with strong performance [14]. More recently, Transformer-based models [15,16] have emerged with promising capabilities in modeling global context and long-range dependencies.

Despite these advances, applying deep stereo models to remote sensing imagery remains highly challenging. Several domain-specific issues must be addressed:

(1)

Computational inefficiency: High-resolution inputs from satellite sensors significantly increase computational demands. For instance, PSMNet [17] requires 576 ms to process a $1024\times 1024$ remote sensing image on an NVIDIA GTX 1080Ti. BGA-Net [12] takes 1572 ms on the same GPU, and MaskCRNet [18] requires 724 ms on an NVIDIA RTX 3090. These requirements hinder real-time use and complicate deployment on edge devices.

(2)

Reduced accuracy in remote sensing scenes: Remote sensing imagery is subject to various challenges, including occlusions in urban environments, repetitive textures such as forests and water surfaces, seasonal appearance changes, illumination variations, human activities, and the presence of small-scale targets. As illustrated in Figure 1, remote sensing stereo pairs often exhibit substantial appearance variations across views. These factors introduce significant ambiguities that degrade the performance of stereo matching models trained primarily on conventional terrestrial datasets. For example, PSMNet [17] achieves a three-pixel error rate of 24.81% on the remote sensing WHU-Stereo dataset [19], but yields a significantly lower error rate of 1.89% on the terrestrial KITTI 2012 dataset [20].

(3)

Dependence on predefined priors: Many stereo matching methods rely on a fixed disparity range (e.g., 0–192), typically determined by the camera baseline and image resolution. This assumption limits model adaptability to the varying imaging geometries present in satellite-based systems. Moreover, the disparity distributions in remote sensing datasets, such as US3D [21] and WHU-Stereo [19], differ markedly from those in terrestrial datasets like KITTI [20] and ETH3D [22]. As illustrated in Figure 2a, remote sensing disparities often span $(-64,64)$ or $(-112,64)$, whereas terrestrial datasets commonly cover ranges such as $(0,192)$ or $(0,256)$. This disparity mismatch not only impedes model generalization but also complicates joint training across domains, thereby limiting the transferability of terrestrial datasets and pretrained models to remote sensing scenarios.

(3)

Limited data: The scarcity of labeled stereo image pairs in remote sensing significantly constrains the generalization ability of deep learning models, particularly in supervised settings. As shown in Figure 2b, the WHU-Stereo dataset provides only 1757 annotated samples, whereas the widely used terrestrial SceneFlow dataset includes 39,049 samples. This severe data imbalance poses a major obstacle to training robust and transferable stereo matching models in remote sensing domains.

Stereo matching in remote sensing (RS) has evolved largely by inheriting and adapting techniques from the computer vision (CV) community. Many representative RS algorithms are extensions of classical CV models, such as SGM, GC-Net, PSMNet, and RAFT-Stereo, because these frameworks provide strong geometric priors and scalable architectures. However, despite their successful transferability, CV models usually experience a noticeable performance drop when applied directly to RS imagery, mainly due to unique characteristics such as larger imaging baselines, radiometric inconsistencies, seasonal variations, and the presence of small repetitive patterns. As a result, most RS-oriented methods take CV architectures as baselines and further introduce domain-specific adaptations, including radiometric normalization, adaptive disparity range modeling, and integration of auxiliary cues like DSM (Digital Surface Model) or LiDAR priors. Therefore, the development of RS stereo matching can be viewed as a progressive evolution from CV methods toward domain-adaptive designs that address the challenges of remote sensing data.

This survey presents a comprehensive review of binocular stereo vision techniques in remote sensing, including representative stereo matching methods, as shown in

Table 1. Summary of representative stereo matching methods.

Type	Category	Characteristic	Representative Works (Common Data)	Representative Works (Remote Sensing)
Traditional	Classical	Interpretable, efficient, strong generalization	Scharstein and Szeliski [5], SGM [7], Hirschmuller et al. [23], Kordelas et al. [24], MGM [25], tMGM [26]	Lee et al. [27], Ghuffar [28], Qin et al. [29], SGM-ForestM [30], Wang et al. [31], Tatar et al. [32], DPSM [33], LGSM [34], Zhao et al. [35], L2GSM [36], Yue et al. [37]
Hybrid	Combine deep learning with traditional	Combines geometric priors with learning flexibility	MC-CNN [38], GA-Net [39], Deeppruner [40]	S2P-GANet [41], Gómez et al. [42], Albanwan and Qin [43], Sat-MVSF [44], Zheng et al. [45]
Supervised	2D Convolution-based	Fast inference	DispNet [9], AANet [46], Bi3D [47], HITNet [48], SMD-Nets [49]	Ji et al. [50], Wang et al. [51]
	3D Convolution-based	High accuracy, high computing cost	Bosch et al. [21], GC-Net [10], CFNet [52], PCW-Net [53], Abc-Net [54], SDA [55], ADL [56], UPFNet [57]	Tao et al. [11], HMSM-Net [58], Jiang et al. [59], Tao et al. [60], PSMNet-FusionX3 [61], SRCV-Net [62], DBMSMNet [63]
	Iterative Optimization	High accuracy, iteratively updating	RAFT-Stereo [14], IGEV [64], IGEV++ [65], MoCha-Stereo [66], Selective-Stereo [67], TC-Stereo [68], MC-Stereo [69]	MaskCRNet [18], Patil and Guo [70]
	Transformer-based	Global context modeling	STTR [15], SSTTStereo [71], ELFNet [72], S2M2 [73]	Wei et al. [74], MaskCRNet [18], A-SATMVSNet [16]
	Multi-task Learning	Cross-task supervision	RTS2Net [75], SGNet [76], NNNet [77], DWARF [78], Rao et al. [79]	BGA-Net [12], S2Net [80], Chen et al. [81], Qin et al. [82], SDBF-Net [83], SemStereo [84]
	Vision Foundation Model Integration	Using priors from VFM, strong generalization	FoundationStereo [85], ViTAStereo [86], Monster [87], DEFOM-Stereo [88]	HDSM-Net [89]
	Other DL Models	Innovative, flexible	LEAStereo [90], NMRF-Stereo [91], CroCo v2 [92], DiffuVolume [93], DMIO [94]	DeepSim-Nets [95], MEMF-Net [96]
Alternative Supervised	Self-supervised	No GT, geometry-driven	SsSMnet [97], ActiveStereoNet [98], Flow2Stereo [99], ES3Net [100], Yuan et al. [101], DualNet [102]	Knöbelreiter et al. [103], Igeta and Iwasaki [104], Chen et al. [105]
	Semi-supervised	Limited GT	Smolyanskiy et al. [106], SemiDepth [107], SSMF [108], Semi-Stereo [109]	HDADE [110]
	Weakly-supervised	Indirect cues	Tulyakov et al. [111], PA-Net [112], SUW-Stereo [113]	Albanwan and Qin [114]

. We begin with traditional methods in Section 2, followed by deep learning approaches in Section 3. Acceleration strategies are discussed in Section 4, and publicly available datasets are summarized in Section 5. Finally, Section 6 outlines key findings and future research directions.

2. Traditional Stereo Matching Methods

Stereo matching has long relied on a classical framework formalized by Scharstein and Szeliski [5] which decomposes the process into four key stages: matching cost computation, cost aggregation, disparity computation, and disparity refinement. This formulation has guided the development of traditional stereo matching algorithms for over two decades. In the first stage, similarity between pixel candidates in the left and right images is estimated, typically using photometric or gradient-based metrics. Lower costs signify higher correspondence likelihood. Cost aggregation then reinforces spatial consistency by integrating matching evidence from neighboring pixels. Disparity computation selects the disparity with minimum aggregated cost per pixel, and a final optimization stage refines results by correcting mismatches and enforcing global constraints.

Traditional methods are broadly categorized into local and global approaches. Local methods restrict computation to finite support windows around each pixel, making implicit smoothness assumptions via cost aggregation. These methods are computationally efficient and well-suited to real-time applications, but their performance often degrades in low-texture or occluded regions due to their limited contextual modeling. Global methods, by contrast, formulate stereo matching as an energy minimization problem, typically using Markov Random Fields (MRFs), where the cost function incorporates both data fidelity and smoothness terms. These models achieve higher accuracy in challenging regions but incur higher computational costs.

To balance these trade-offs, Hirschmüller proposed Semi-Global Matching (SGM) [7], which performs multiple 1D global optimizations along multiple paths to approximate global regularization while maintaining efficiency. Due to its effectiveness, SGM quickly became the method of choice for many real-world applications. In parallel, Hirschmüller et al. [23] conducted a comprehensive study of matching cost functions, offering valuable guidance for algorithm design. Since its introduction, SGM has been extensively extended and adapted for diverse datasets and domains [24,25,26].

Building on the SGM framework, a series of methods were specifically developed for remote sensing stereo applications. Lee et al. [27] pioneered the use of stereo matching in satellite 3D reconstruction, incorporating a conjugate search strategy and correlation-based patch design to improve runtime efficiency and matching quality. Ghuffar et al. [28] applied SGM to WorldView-3 imagery and qualitatively evaluated its reconstruction performance.

Machine learning has also been introduced to augment traditional pipelines. Qin et al. [29] trained a Support Vector Machine (SVM) [115] to predict the potential stereo matching quality. SGM-ForestM [30] leveraged random forests to estimate optimal disparities per scanline, significantly enhancing SGM performance. Wang et al. [31] developed a feature-based SGM acceleration method that reduced noise and improved robustness in textureless or discontinuous regions. Similarly, Tatar et al. [32] improved aggregation by introducing homogeneity weights and edge-guided filtering, which were particularly effective for urban scenes.

Recent research has increasingly focused on integrating geometric and semantic priors. LGSM [34] (LiDAR-Guided Semi-Global Matching) constrained disparity search ranges using LiDAR data, thereby reducing mismatches in difficult matching areas and refining the boundaries of objects. DPSM [33] (Dual Propagation Stereo Matching) introduced a novel bidirectional propagation strategy that exploited building regularities and shape constraints, optimizing disparity via an energy minimization framework. Zhao et al. [35] incorporated building edge cues into cost aggregation using a multi-order Census transform to mitigate the effects of noise. L2GSM [36] fused LiDAR data and depth discontinuity lines with SGM, yielding improved results in low-textured and depth discontinuity regions. Most recently, Yue et al. [37] proposed a hierarchical aggregation strategy guided by semantic and geometric edge features. Their method adaptively constrained local disparity candidates, improving accuracy in complex urban environments.

In summary, traditional stereo matching methods, especially those based on the SGM framework, have laid a solid foundation for stereo vision in remote sensing. By integrating geometric priors, semantic cues, and learning-based enhancements, these methods have continued to evolve. However, their reliance on hand-crafted features and limited accuracy in complex scenes underscores the need for more intelligent, data-driven approaches.

3. Deep Learning-Based Stereo Matching Methods

Deep stereo matching has progressed rapidly with the advancement of deep learning. In this section, we present a comprehensive taxonomy of deep stereo methods, grouped into three major categories: (1) hybrid approaches that combine deep learning with traditional stereo algorithms, (2) fully supervised end-to-end methods, and (3) alternative supervision strategies, including self-supervised, semi-supervised, and weakly supervised learning.

The fully supervised category is further subdivided based on architectural design, encompassing 2D convolutional models, 3D convolutional networks, iterative refinement frameworks, Transformer-based architectures, multi-task learning strategies, and methods integrating vision foundation models.

This survey is organized following the evolution from general computer-vision stereo algorithms toward domain-adaptive designs for remote sensing. Each methodological category is discussed first in the CV context and then extended to RS scenarios. A detailed review of each category, along with representative methods, is provided in the following subsections.

3.1. Combining Deep Learning and Traditional Algorithms

Deep learning techniques have demonstrated remarkable capabilities in tasks such as object recognition, classification, and semantic segmentation since 2012. The first application of deep learning to stereo matching can be traced back to 2015. From 2015 to 2020, researchers began integrating deep learning modules into traditional stereo pipelines, leading to what are commonly called hybrid methods [38,39,40]. These approaches aim to retain the robustness and interpretability of classical algorithms while leveraging the learning power of deep networks.

In the field of remote sensing, several notable works have adopted this hybrid strategy. To address the integration gap between deep learning models and satellite stereo pipelines, as well as the lack of systematic performance evaluation in complex multi-temporal scenarios, S2P-GANet [41] introduced a stereo processing pipeline tailored to satellite imagery, in which Rational Polynomial Coefficients (RPCs) were utilized to model satellite viewing geometry. Within this framework, GA-Net [39] was employed to regress bidirectional disparity maps. Furthermore, by integrating this pipeline with multi-view stereo techniques such as COLMAP [116], the method achieved enhanced 3D reconstruction performance. To address the sub-optimal performance caused by heuristic pair selection and noisy data integration in traditional satellite multi-view stereo pipelines, Gómez et al. [42] proposed an iterative refinement approach incorporating bilateral filtering and multi-view fusion, further enhancing the geometric accuracy of remote sensing reconstructions. In a broader evaluation, Albanwan and Qin [43] systematically compared deep stereo networks with classical algorithms such as SGM. Their analysis revealed two key observations: (1) deep learning models typically outperform classical methods under in-distribution scenarios, but often suffer from poor generalization to unseen domains; and (2) traditional approaches exhibit more consistent performance across domains. These findings underscore a fundamental trade-off between the adaptability of learned models and the robustness of handcrafted priors.

Recent studies have investigated integrating deep learning with traditional geometric constraints to address the challenges of geometric adaptation in satellite multi-view stereo matching, the scarcity of training samples, and the limited generalization across multi-temporal imagery. Specifically, Sat-MVSF [44] introduced a self-optimization framework that leverages pseudo-labels guided by structural consistency, demonstrating strong robustness across diverse and seasonal datasets. Additionally, Zheng et al. [45] proposed a two-stage stereo pipeline, wherein a deep network produces an initial disparity estimate, which is subsequently refined via handcrafted post-processing.

Overall, these approaches integrate the robustness of handcrafted geometric priors with the representational flexibility of deep neural networks, thereby improving generalization across domains. Nevertheless, their multi-stage pipeline is relatively complex, and the overall performance gain remains limited when compared with fully learning frameworks.

3.2. End-to-End Supervised Algorithms

End-to-end supervised stereo matching methods take a pair of stereo images as input and directly predict the corresponding disparity map. We categorize existing end-to-end stereo matching methods into seven architectural groups: (1) 2D convolution-based networks, (2) 3D convolution-based networks, (3) iterative refinement frameworks, (4) Transformer-based architectures, (5) multi-task learning models, (6) methods integrating vision foundation models, and (7) other emerging designs.

3.2.1. 2D Convolution-Based Models

One of the earliest milestones in end-to-end stereo matching was DispNet [9], which introduced a fully convolutional architecture integrating a 3D cost volume within a 2D convolutional framework. The model comprises four key components: feature extraction using 2D convolutions, construction of a 3D cost volume from stereo image features, cost volume regularization via an encoder-decoder structure, and disparity regression. Subsequent methods, including AANet [46], Bi3D [47], HITNet [48], and SMD-Nets [49], followed this design paradigm by leveraging 2D convolutions for efficient representation learning while maintaining the expressiveness of volumetric cost aggregation.

In the remote sensing domain, Ji et al. [50] extended two-view deep stereo networks (GC-Net [10], MC-CNN [38]) into a dense multi-view matching framework with multi-view geometry constraints, tailored for strip-overlap acquisition of aerial photogrammetry. Their evaluation demonstrated a consistent superiority of deep models over traditional methods. Wang et al. [51] surveyed dense matching cost computation for satellite stereo and, through comparative analysis including DispNet-C and SGM, reinforced the benefits of learning-based approaches in remote sensing scenarios.

2D convolution-based stereo matching models were the first to enable fully end-to-end disparity estimation, outperforming traditional approaches in terms of accuracy. Owing to their exclusive use of 2D convolution operations, these models offer notable advantages in inference efficiency, making them suitable for real-time and edge computing applications. However, they still face limitations in accuracy and robustness, particularly in geometrically complex or textureless regions. As high-precision 3D convolutional architectures emerged, research interest in 2D convolution-based methods has gradually declined.

3.2.2. 3D Convolution-Based Models

3D convolution-based stereo models closely follow the classical four-step stereo matching pipeline and aim to enhance spatial regularization through 4D cost volume construction, as shown in Figure 3. A seminal contribution in this category is GC-Net [10], which first extracts features from left and right images using 2D convolutional layers. These features are concatenated across the disparity dimension to construct a 4D cost volume, which is then regularized using 3D convolutions, enabling the aggregation of both spatial and disparity cues. The final disparity map is obtained via a differentiable Soft ArgMin operation. GC-Net established a foundational design that inspired numerous successors [21,52,53,54,55,56,57], improving accuracy across a variety of scenes.

In remote sensing domain, Ji et al. [117] demonstrated that GC-Net [10] and MC-CNN [38] significantly outperform traditional algorithms when applied to satellite imagery. To address the specific challenge of occlusions in satellite scenes, Tao et al. [11] integrated an unsupervised bidirectional loss into PSMNet [17] via bidirectional pyramid network, achieving notable improvements in heavily occluded regions. Further extending this direction, HMSM-Net [58] introduced a hierarchical multi-scale stereo matching framework designed to handle intractable regions caused by repetitive patterns, texture-less areas, and disparity discontinuities. By constructing multi-scale cost volumes, the model effectively enhances disparity estimation in these challenging scenarios. Moreover, recognizing that the scarcity of training data is a major bottleneck hindering the deployment of CNN-based techniques, the authors released the Gaofen-7 dataset, establishing a new benchmark for evaluating remote sensing stereo algorithms.

Subsequently, to address the high cost of acquiring LiDAR ground truth and the limited generalization of stereo networks across different sensors and scenarios, Jiang et al. [59] introduced a systematic training strategy aimed at enhancing the generalization ability of remote sensing stereo matching networks. Their approach was validated on several state-of-the-art architectures, including CFNet [52], HMSM-Net [58], and PASMNet [118], and demonstrated consistent performance improvements across diverse satellite scenes. To address the challenges of matching multi-scale objects in large scenes and the ambiguity of multi-modal probability distributions in occluded or textureless regions, Tao et al. [60] proposed a confidence-aware cascade refinement strategy, which progressively refines disparity maps from coarse to fine resolution. To address the limitations of noisy photogrammetric surfaces and the excessive sparsity of LiDAR point clouds in aerial 3D reconstruction, PSMNet-FusionX3 [61] introduced a fusion framework combining stereo image features and LiDAR point clouds via triangular irregular interpolation, improving performance in sparse regions. To address the problem of disparity shifts caused by repetitive textures and textureless regions in satellite imagery, SRCV-Net [62] introduced a Cost Volume Refinement Strategy (CVRS) that uses a “left-left” cost volume as a reference to suppress false matches in the standard “left-right” cost volume. Integrated into their SRCV-Net, this method significantly improves disparity consistency and accuracy for satellite stereo matching, especially in repetitive and textureless areas. To address the challenges of disparity estimation in complex intractable regions, such as textureless, repeated texture, and occlusion areas, DBMSMNet [63] proposed a lightweight stereo matching network featuring a dual-branch module for multiscale feature extraction. It uses a coarse-to-fine cost aggregation strategy with disparity-channel attention for enhanced fusion, and a final refinement step guided by image intensity and gradients to produce accurate disparity maps for satellite images.

3D convolution-based stereo matching networks closely adhere to the classical pipeline, offering strong interpretability. Compared to 2D convolution-based network, these models achieve higher disparity estimation accuracy by constructing and regularizing a 4D cost volume. However, the computational overhead associated with 3D convolutions imposes significant constraints on inference speed, limiting their applicability in real-time or large-scale remote sensing scenarios.

3.2.3. Iterative Optimization-Based Models

Iterative optimization-based models draw inspiration from classical optimization-based methods by refining disparity estimates through multiple recurrent updates. A seminal contribution in this direction is RAFT [13], originally developed for optical flow and awarded Best Paper at ECCV 2020. RAFT extracts 2D features at one-eighth the input resolution, computes pairwise similarity to build a correlation volume, and iteratively refines the flow or disparity using a Gated Recurrent Unit (GRU). The final disparity map is generated through upsampling to the full resolution.

RAFT-Stereo [14] adapted this framework for stereo matching by tailoring the similarity computation to disparity-specific cues, achieving competitive results on multiple benchmarks. Its success has inspired a series of follow-up models, such as IGEV [64], IGEV++ [65], MoCha-Stereo [66], Selective-Stereo [67], TC-Stereo [68], and MC-Stereo [69], which further improve matching precision, runtime, or domain robustness through architectural refinements.

In the remote sensing domain, Patil and Guo [70] introduced the Stellar dataset to address the challenges of significant visual appearance variations caused by non-simultaneous imaging and the scarcity of large-scale annotated data. By evaluating RAFT-Stereo and 3D convolutional networks on this benchmark—which provides rectified stereo pairs with true disparity and semantic labels—they demonstrated that both iterative refinement frameworks and 3D architectures are capable of producing dense and accurate disparity maps even under complex radiometric configurations. Building upon this direction, to address the challenges of imperfect epipolar rectification, missing data, and significant domain differences in high-resolution satellite imagery, MaskCRNet [18] proposed a cascaded recurrent stereo architecture specifically designed for satellite images, as shown in Figure 4. The model integrates a Transformer-based encoder, a CNN-based encoder, multi-scale cascaded recurrent refinement modules, and a self-supervised image reconstruction branch, achieving SOTA accuracy on high-resolution remote sensing benchmarks.

Iterative optimization-based models provide a compelling balance between computational efficiency and refinement capacity. Their recurrent architecture enables accurate and memory-efficient updates, supporting robust performance across diverse domains. However, due to their inherently sequential inference process, these models may encounter latency bottlenecks in high-resolution or large-scale deployments, posing a critical challenge for real-time remote sensing applications.

3.2.4. Transformer-Based Models

The Transformer architecture, originally introduced by Vaswani et al. [119] for sequence modeling in natural language processing, was later adapted to vision tasks via the Vision Transformer (ViT) [120]. ViT segments an image into fixed-size patches, encodes each patch into a feature token, and leverages multi-head attention to capture long-range dependencies across the spatial domain. Motivated by its success in image classification, the Transformer paradigm has since been extended to stereo matching.

As illustrated in Figure 5, we present four types of Transformer-based stereo matching models: (a) direct attention-based disparity estimation, exemplified by STTR [15]; (b) Transformer-based feature extraction, as in A-SATMVSNet [16]; (c) hybrid models combining Transformer and CNN backbones, as in MaskCRNet [18]; and (d) Transformer-based cost volume regularization, such as SSTTStereo [71].

STTR [15] represents one of the earliest stereo matching frameworks based on attention mechanisms. It employs a convolutional stem to extract features from input images, followed by a Transformer module that performs both self-attention and cross-attention to estimate initial disparity values, which are subsequently refined through a context adjustment stage. SSTTStereo [71] proposes a Sliding Space-Disparity Transformer that regularizes the 4D cost volume using localized attention, balancing accuracy and computational efficiency. ELFNet [72] introduces a deep evidence learning framework that fuses Transformer-based context encoding with cost-volume representations, enabling two-stage fusion and generating both aleatoric and epistemic uncertainty maps alongside the predicted disparity. S²M² [73] is a scalable global stereo matching model that leverages a Multi-Resolution Transformer (MRT) to efficiently handle high-resolution inputs. It combines optimal transport for robust correspondence with a novel Probabilistic Mode Concentration (PMC) loss, achieving state-of-the-art accuracy and reliable depth estimation.

In the field of remote sensing, A-SATMVSNet [16] incorporated attention mechanisms into the feature extraction stage to alleviate the problem of incomplete surface detail recovery in satellite images. Similarly, MaskCRNet [18] combined a ViT-based encoder with masked representation learning in a cascaded recurrent framework to handle missing data, domain shift, and rectification errors in high-resolution satellite stereo. Wei et al. [74] further improved robustness by constructing multi-scale cost volumes and applying feature-level attention to suppress image noise and enhance disparity map quality for satellite stereo.

Transformers offer strong capabilities in hierarchical feature extraction and global context modeling. On terrestrial datasets, Transformer-based stereo matching models have progressively demonstrated SOTA performance. However, their application in remote sensing stereo tasks remains limited. Given the complex textures and the presence of small-scale targets in satellite imagery, the ability of Transformers to capture fine-grained and long-range dependencies makes them particularly well-suited for this domain. As such, developing Transformer-based stereo models represents a promising direction for advancing stereo matching in remote sensing.

3.2.5. Multi-Task Learning-Based Models

Multi-task learning (MTL) seeks to jointly optimize multiple related tasks, enabling shared feature representations and cross-task supervision. In the context of stereo matching, MTL is commonly combined with auxiliary tasks such as semantic segmentation, surface normal estimation, or optical flow prediction. This integration facilitates a more comprehensive understanding of scene geometry and semantics, thereby improving disparity estimation performance.

Several notable methods have demonstrated the efficacy of MTL. RTS²Net [75] proposed a lightweight, real-time architecture that performs joint semantic segmentation and disparity estimation using a shared encoder to balance accuracy and computational efficiency. SGNet [76] introduced a multi-task framework incorporating three modules: a confidence module assessing consistency between semantic and disparity features, a residual initial disparity refinement module guided by semantic categories, and a fusion module for final disparity adjustment. NNNet [77] added a surface normal estimation branch to improve disparity accuracy via a normal consistency loss and refinement scheme. DWARF [78] presented a coarse-to-fine architecture that simultaneously estimates disparity, optical flow, and disparity change using compact correlation volumes and hierarchical feature warping. More recently, Rao et al. [79] integrated masked self-supervised learning into a pseudo-multi-task framework to enhance model robustness and cross-domain generalization.

In the domain of remote sensing, a key milestone was the 2019 IEEE GRSS Data Fusion Contest (DFC) [121], which, for the first time, provided paired disparity and semantic segmentation annotations for satellite imagery via the US3D dataset. The competition aimed to assess the potential of joint learning strategies for stereo matching in satellite scenarios, where models must contend with drastic object-scale variation (from vehicles to large buildings), weak-texture regions (flat rooftops, farmland), heavy occlusions and shadows, and the imperfect rectification typical of incidental satellite imagery. Chen et al. [81], the winning team, proposed a multi-receptive-field contextual fusion module to aggregate context across heterogeneous object scales, solving the scale-variance problem inherent to high-resolution satellite scenes. Qin et al. [82] introduced a dynamic loss-weighting scheme and an ensemble voting strategy to jointly train U-Net and PSMNet [17], balancing the convergence rates of the semantic and disparity branches and stabilizing predictions in low-confidence regions caused by satellite-specific noise, securing second place. SDBF-Net [83] combined DeepLab-v3 and GC-Net [10] as backbones and introduced a semantic–disparity bidirectional fusion mechanism, leveraging semantic priors to disambiguate disparity in weak-texture and occluded regions on incidental satellite images.

Subsequent research further advanced these ideas. BGA-Net [12] introduced top-down and bottom-up bidirectional guided attention to share information between semantic segmentation and disparity estimation, alleviating the seasonal-appearance variation that hampers cross-temporal 3D detection in remote sensing imagery. S²Net [80] built a dual-task framework in which semantic features supervise cost-volume construction to stabilize disparity under data disturbance, while disparity-derived RGB-D features in turn disambiguate foreground–background confusion in satellite RGB imagery. Figure 6 illustrates the architectures of BGA-Net and S²Net, along with an example of downstream 3D reconstruction enabled by their outputs. SemStereo [84] argued that earlier multi-task designs—being either loosely parallel or only implicitly interacting—fail to capture the inherent coupling between the two heterogeneous tasks in remote sensing. It therefore tightly couples them through a Semantic-Guided Cascade (SGC) that propagates deep semantic features into disparity estimation, a Semantic Selective Refinement (SSR) branch for explicit disparity refinement, and a Left–Right Semantic Consistency (LRSC) loss enforcing cross-view consistency, demonstrating the effectiveness of tightly coupling semantic and stereo tasks.

MTL offers a promising strategy for enhancing stereo matching through the integration of complementary tasks. Its application in remote sensing has demonstrated notable improvements in both accuracy and generalization. However, future work must address challenges in task balancing, annotation cost, and computational overhead to fully realize its potential in operational settings.

3.2.6. Integrating Vision Foundation Models

Vision Foundation Models (VFMs), such as CLIP [122], DINO [123,124], and SAM [125], have recently garnered significant attention due to their impressive generalization capabilities across diverse tasks and domains. These models are pretrained on large-scale vision-language datasets or on unlabeled image data using self-supervised learning, serving as powerful task-agnostic feature extractors and promptable learners. While VFMs have been extensively applied to classification, segmentation, object detection, and text-to-image, their integration into stereo matching remains a nascent yet promising research direction.

Current approaches typically adapt VFMs in two main ways: (1) as high-capacity feature extractors for stereo matching pipelines; and (2) as providers of monocular depth priors that guide stereo estimation. ViTAStereo [86] represents one of the earliest explorations in this domain, leveraging features from the vision foundation model DINOv2 [124] to enhance the generalization of stereo models. MonSter [87] introduced a dual-branch architecture consisting of a monocular depth estimation module and a stereo matching module, coupled via a mutual refinement mechanism. The monocular branch utilizes DepthAnythingV2 [126], while the stereo branch adopts the IGEV framework [64]; both outputs are iteratively refined through cross-task feedback. DEFOM-Stereo [88] integrated monocular depth predictions into traditional iterative optimization-based stereo networks, using pretrained depth priors to initialize disparity estimation in a recurrent matching process. Later, FoundationStereo [85] proposed a comprehensive framework comprising a large-scale synthetic dataset (one million stereo pairs), monocular priors from DepthAnythingV2, and an attentive hybrid filter that adaptively fuses spatial and disparity cues. The architectures of these four approaches are depicted in Figure 7.

In the field of remote sensing, HDSM-Net [89] addressed the constraints of fixed disparity ranges and the scarcity of high-quality ground-truth labels by proposing a hierarchical vision transformer (ViT) framework for satellite stereo matching. This approach integrates self-supervised DINO for robust feature learning and a Context-Enhanced Path (CEP) to fuse global contextual information. Furthermore, by employing a pixel-by-pixel matching strategy and improved position encoding, the model enhances spatial accuracy while effectively mitigating the geometric limitations inherent in traditional satellite imagery.

Collectively, these pioneering studies underscore the potential of VFMs in advancing stereo matching, particularly in enhancing cross-domain generalization and fusing multi-modal representations. As VFMs continue to evolve with richer geometric understanding and multi-scale semantic encoding, their application in remote sensing stereo tasks represents a fertile ground for future research, especially for achieving robust and semantically-aware 3D reconstruction in open-world environments.

3.2.7. Other Emerging Models

In addition to convolutional, recurrent, and Transformer-based models, recent research has explored a variety of alternative strategies for stereo matching, drawing inspiration from architecture search, probabilistic modeling, universal visual pretraining, generative learning paradigms, and contrastive learning.

LEAStereo [90] leveraged a hierarchical Neural Architecture Search (NAS) framework [127], which incorporated stereo-specific priors into the search process to discover architectures optimized for disparity estimation. This approach yielded improved task-adaptive performance and enhanced model accuracy. NMRF-Stereo [91] introduced a Neural Markov Random Field (NMRF) formulation that learns pixel-level dependencies in a fully data-driven manner, mitigating the inefficiency of manual design and enabling more flexible cost aggregation. CroCo v2 [92] extends the original CroCo framework [128] by pretraining a unified architecture that consists of a monocular encoder and a binocular decoder. The model is trained using a cross-view completion task, enabling it to learn dense correspondence. Notably, this approach eliminates the need for conventional components such as cost volumes, feature warping, or multi-scale refinement.

Generative learning has also made inroads into stereo vision. DiffuVolume [93] was the first to apply diffusion models to stereo matching, modeling the disparity estimation process as an iterative denoising procedure over a latent cost volume. Building on this, DMIO [94] combined diffusion-based reasoning with an iterative optimization framework, introducing a time-aware gated recurrent unit (T-GRU) to simultaneously capture disparity and temporal dynamics.

In the remote sensing domain, to address the over-reliance on specific scene geometries and the limited generalization of end-to-end models in unseen scenarios, DeepSim-Nets [95] proposed three multi-scale architectures that learn pixel-level similarity using contrastive loss. By remaining explicitly geometry-agnostic, these models decouple similarity learning from the underlying scene geometry, ensuring robust and flexible matching performance across diverse aerial and satellite imagery datasets.

Furthermore, to address the loss of global contextual information caused by patch-wise processing in high-complexity Transformers and the difficulty of matching multiscale objects in ill-posed regions, MEMF-Net [96] introduced a Mamba-based stereo matching framework for high-resolution satellite imagery. By leveraging a Mamba-based feature extractor for efficient global context modeling and a channel-spatial attention-enhanced multifrequency fusion module, the network enhances disparity estimation in challenging areas. Moreover, it incorporates a gradient-based convex upsampling module to refine output details.

Collectively, these emerging approaches reflect a growing interest in rethinking stereo matching beyond traditional pipelines. By incorporating principles from architecture search, probabilistic reasoning, foundational visual pretraining, generative modeling, and contrastive learning, they provide new pathways for building more interpretable, generalizable, and scalable stereo vision systems.

3.3. Weakly-, Semi-, and Self-Supervised Algorithms

Most deep stereo matching models rely heavily on large-scale, densely annotated disparity maps for supervised training. However, acquiring such annotations is particularly challenging in real-world scenarios, especially in remote sensing, due to the significant labor, time, and cost involved in generating high-quality ground truth data. To mitigate this dependency, recent research has focused on weakly-, semi-, and self-supervised approaches that aim to reduce the reliance on labeled data while maintaining competitive model performance.

3.3.1. Self-Supervised Models

Self-supervised learning refers to a class of approaches that generate supervisory signals directly from the input data, without requiring explicit ground truth. In stereo matching, most self-supervised methods are grounded in the principle of geometric consistency: given a left view and a predicted disparity map, the right view is reconstructed via image warping. The photometric difference between the reconstructed and original right views is then used as a loss function to guide network optimization [97,98,99,100,101]. Recently, DualNet [102] proposed a self-supervised stereo framework that combines robust self-supervised teacher learning and pseudo-label supervised student training.

In the remote sensing domain, Knöbelreiter et al. [103] generated training data directly by applying their own 3D reconstruction method to the target dataset, and introduced an outlier filtering step to ensure data quality. To circumvent the high cost and difficulty of acquiring ground-truth labels for high-resolution satellite imagery, Igeta et al. [104] proposed an unsupervised stereo matching network. Their approach addresses the inherent model selection challenge in unsupervised settings by introducing a novel criterion for identifying the optimal training epoch without the need for validation labels, ensuring robust performance even in the absence of reference data. To enhance robustness in texture-less and discontinuous regions, Chen et al. [105] developed a self-supervised method that incorporates a superpixel random walk pre-matching (SRWP) strategy alongside a parallax-channel attention mechanism (PCAM), as shown in Figure 8.

Despite recent advancements, self-supervised stereo matching methods still face significant challenges in achieving consistent and accurate results, especially in remote sensing. Unlike terrestrial datasets, remote sensing imagery often exhibits: (1) significant seasonal or temporal differences between left and right views; and (2) large homogeneous or repetitive textures (e.g., forests, water bodies, and building facades). These conditions undermine the reliability of photometric and geometric consistency assumptions, thereby limiting the effectiveness of traditional self-supervised loss formulations. Developing self-supervised algorithms that can robustly handle such domain-specific challenges remains an important and open research problem in remote sensing stereo vision.

3.3.2. Semi-Supervised Models

Semi-supervised learning serves as a promising compromise between fully supervised and self-supervised paradigms, aiming to leverage both labeled and unlabeled data to improve model performance. Smolyanskiy et al. [106] employed a semi-supervised loss function that combined LiDAR supervision with photometric consistency to train a deep stereo neural network. SemiDepth [107] leveraged left-right consistency in stereo reconstruction and integrated ground truth depth derived from LiDAR to guide network training. Xu et al. [108] proposed a unified framework that incorporates consistency regularization and entropy minimization to effectively utilize large-scale unlabeled data. Semi-Stereo [109] introduced a semi-supervised stereo matching framework based on the teacher–student paradigm, wherein both networks are co-trained in a mutually beneficial manner. They further proposed a consistency-based pseudo-labeling regularization strategy with weak-strong data augmentation to better exploit information from noisy and incomplete data.

In the remote sensing domain, HDADE [110] tackles the two key challenges of domain shifts and the high cost of annotations by implementing a semi-supervised hierarchical domain adaptation framework. This multi-stage approach systematically facilitates feature alignment across disparate satellite sensors, proving that robust disparity estimation is achievable in cross-domain environments with minimal supervision, as shown in Figure 9.

Semi-supervised approaches show great potential in mitigating data scarcity issues and improving cross-domain generalization. However, further research is needed to address challenges such as pseudo-label noise, domain shifts, and the stability of consistency training, especially in complex and heterogeneous remote sensing environments.

3.3.3. Weakly-Supervised Models

Weakly-supervised stereo matching seeks to estimate disparities using indirect or imprecise supervisory signals, such as semantic segmentation, geometric layout, or camera configuration priors, rather than dense ground-truth disparity labels. This paradigm is especially appealing for domains like remote sensing, where acquiring accurate pixel-wise annotations at scale is impractical.

Tulyakov et al. [111] pioneered this direction by integrating coarse scene and optical flow priors to guide disparity estimation from stereo image pairs, effectively mitigating label noise during training. Following this work, additional weakly-supervised methods have been proposed for general-purpose stereo vision [112], demonstrating the feasibility of learning disparity under limited annotation regimes. SUW-Stereo [113] jointly combined supervised learning, unsupervised learning, and weakly-supervised learning for disparity estimation.

In the remote sensing domain, Albanwan and Qin [114] addressed the high cost of acquiring ground truth for diverse global locations and the limited transferability of deep learning models by introducing a weakly-supervised strategy. To circumvent the dependence on precise labels, this approach utilizes filtered disparity maps derived from the traditional Semi-Global Matching (SGM) algorithm as weak supervision signals. Specifically, the framework employs a confidence-based selection scheme that integrates SGM energy maps with image texture analysis to extract reliable disparity measurements, effectively mitigating the noise inherent in traditional matching results. By fine-tuning well-established models—such as GC-Net, PSMNet, and LEAStereo—on these high-confidence pseudo-labels, the method significantly enhances model robustness across 20 diverse geographical sites without requiring any ground-truth information, as shown in Figure 10.

Despite its potential, weakly-supervised stereo matching remains underexplored in remote sensing. This is largely due to the absence of suitable indirect labels and standardized benchmarks tailored to satellite imagery. Consequently, progress in this area has stalled, highlighting an urgent need to construct remote sensing-specific datasets and evaluation protocols that facilitate weakly-supervised learning.

3.4. Advantages and Limitations

Different categories of stereo matching algorithms present distinct strengths and weaknesses. For clarity, we summarize the main advantages and limitations of each representative approach in

Table 2. Advantages and Disadvantages (A&D) of different categories of stereo matching algorithms.

Category	Advantages	Disadvantages
Hybrid methods	Combine robustness of priors with flexibility of deep learning; improved cross-domain generalization	Complex pipeline; limited improvement compared to fully learning networks
2D convolution-based models	Efficient and lightweight; suitable for real-time and edge devices	Limited accuracy and robustness in complex or textureless regions
3D convolution-based models	Strong spatial regularization; higher disparity accuracy	High computation and memory cost; difficult for large-scale or real-time use
Iterative optimization-based models	Accurate and memory-efficient; support progressive refinement	Sequential inference introduces latency, especially on high-resolution RS data
Transformer-based models	Capture long-range dependencies; strong semantic awareness	High computational cost; limited applications in RS domain
Multi-task learning models	Leverage semantic/geometry cues; benefit downstream tasks	Task balancing is challenging; increased annotation and computation cost
Vision Foundation Model Integration	Strong generalization; effective cross-domain adaptation via large-scale pretraining	Integration into stereo is still preliminary; training/fine-tuning is resource-demanding
Self/semi/weakly supervised methods	Reduce dependence on dense GT disparity; more practical in RS	Sensitive to seasonal/illumination changes; accuracy gap with fully supervised remains

, which provides a concise comparison to highlight their applicability in remote sensing scenarios. Beyond category-specific designs, several symmetry/consistency-based refinements (e.g., left–right semantic/photometric consistency) have shown tangible gains and are directly applicable to RS scenarios.

4. Acceleration of Stereo Matching Models

End-to-end deep stereo matching models have substantially improved disparity estimation accuracy. However, their reliance on computationally intensive operations, particularly large 3D convolutions, often demands high-performance GPUs, thereby limiting their deployment in real-time or resource-constrained environments. To overcome these challenges, recent research has increasingly emphasized model acceleration, aiming to reduce runtime and memory consumption while maintaining competitive accuracy.

4.1. Lightweight Design

Although 3D convolution-based architectures offer superior accuracy, they incur significant computational overhead due to the high complexity of cost volume regularization. This results in parameter redundancy and constrains their practical application in real-world scenarios. Consequently, a significant number of studies have focused on designing lightweight network architectures to reduce computational overhead while preserving accuracy.

SCV-Net [129] introduced sparse cost volume construction to reduce computational demands, enabling stereo models to retain high accuracy with lower memory usage and faster inference. To mitigate the high GPU memory footprint of the Guided Aggregation Network (GA-Net), Xia et al. [130] proposed a pyramid optimization module that progressively refines disparity predictions. Their method significantly reduced memory overhead and improved runtime efficiency, while effectively accommodating large disparity ranges commonly encountered in remote sensing imagery.

Current lightweight approaches mainly center around architectural refinements and computational optimizations. While some models have demonstrated real-time performance on GPUs, achieving similar efficiency on embedded platforms or mobile devices remains an open challenge for future exploration.

4.2. Compression

4.2.1. Model Pruning

Model pruning aims to eliminate redundant parameters by identifying and removing less informative weights, thereby reducing network size and computational complexity while maintaining inference accuracy. Importance criteria for pruning can be based on weight magnitudes, gradient sensitivity, or activation statistics.

Xiang et al. [131] introduced a fast iterative shrinkage-thresholding algorithm to learn soft pruning masks for 2D convolutional layers, enabling more efficient stereo reconstruction. Despite the success of pruning in other vision domains, its application to stereo matching remains limited. Nonetheless, these initial efforts suggest that tailored pruning strategies could significantly enhance the computational efficiency of stereo models. In the context of remote sensing, where high-resolution data and large disparity ranges are common, designing pruning techniques that preserve spatial detail and domain-specific robustness remains a valuable direction for future research.

4.2.2. Model Quantization

Model quantization aims to compress deep neural networks by reducing numerical precision, thereby decreasing model size and accelerating inference. The primary challenge lies in minimizing performance degradation resulting from reduced bit width and dynamic range. Although quantization has been extensively applied to tasks such as image classification and object detection, its use in stereo matching remains relatively underexplored.

Bi3D [47] proposed a binary disparity estimation network, demonstrating a linear trade-off between precision and computational complexity. Chen et al. [132] developed a stereo matching framework using binary descriptors, which achieved real-time performance on FPGA hardware. More recently, Wang et al. [133] introduced an 8-bit quantized stereo model optimized for deployment on wearable edge devices such as smart glasses.

While quantization offers significant advantages in terms of hardware deployment and runtime efficiency, it often entails accuracy degradation. Furthermore, quantized stereo models tailored for remote sensing imagery remain largely unexplored. Given the unique challenges posed by large disparity ranges, repetitive textures, and environmental noise in satellite imagery, developing robust quantized stereo methods is both an urgent and impactful research direction.

5. Remote Sensing Stereo Matching Datasets

5.1. Development of Stereo Datasets

Public datasets have played a pivotal role in advancing stereo matching models, especially in the deep learning era. Numerous benchmark datasets have been developed for terrestrial scenarios. For instance, in 2012 and 2015, Geiger et al. [20,134] introduced the KITTI 2012 and KITTI 2015 datasets, which leverage LiDAR data to provide sparse ground truth for autonomous driving applications. Scharstein et al. [135] presented the Middlebury dataset in 2014, using structured light to obtain dense disparity maps in indoor environments. In 2016, Mayer et al. [9,136] proposed SceneFlow, a large-scale synthetic dataset created using Blender 4.2, which became widely used for pre-training stereo networks. This contributed to the now-standard training paradigm: pre-training on synthetic datasets followed by fine-tuning on smaller, real-world datasets. That same year, Huang et al. introduced Apolloscape [137], a high-resolution street-view dataset. More recently, DrivingStereo [138] and CRE-Stereo [139] further enriched large-scale stereo datasets with diverse weather conditions and synthetic scenarios.

In contrast to the abundance of stereo datasets available for natural image scenarios, benchmark datasets for remote sensing stereo matching remain scarce. This limitation primarily stems from the substantial cost and technical challenges associated with generating accurate disparity ground truth at large scales. To address this gap, the IEEE GRSS Data Fusion Contest (DFC) in 2019, organized by Le et al. [21], introduced the Urban Semantic 3D (US3D) dataset, providing annotated data for both stereo matching and semantic segmentation. As summarized in

Table 3. Evaluation of the stereo matching methods on the US3D datasets. EPE: End-Point Error. D1: percentage of erroneous pixels. The downward arrow (↓) next to a metric indicates that lower values correspond to better performance. * denotes self-supervision, ^† represents baseline results reported in [21]. Note: MRFCNet is named after the multi-receptive-field contextual network proposed in [81]; the original paper does not provide an explicit acronym.

Method	Year	Publication	EPE (px) ↓	D1 (%) ↓	Times (ms) ↓	Equipment
SGM † [141]	2005	CVPR	10.34	43.00	-	Xeon
iResNet-i2 † [142]	2018	CVPR	3.05	33.00	-	NVIDIA Titan X
DenseMapNet † [143]	2018	ICRA	3.51	35.00	-	NVIDIA GTX 1080Ti
MRFCNet [81]	2019	IGARSS	1.34	9.01	670	-
BGA-Net [12]	2020	TGRS	1.20	7.20	1572	NVIDIA 1080Ti
RAFT-Stereo [14]	2021	3DV	1.29	8.01	575	NVIDIA RTX 6000
HMSM-Net [58]	2022	ISPRS	1.19	7.16	511	NVIDIA RTX 3090
S2Net [80]	2023	TGRS	1.439	10.051	-	NVIDIA Tesla V100
MaskCRNet [18]	2024	ISPRS	1.12	7.01	724	NVIDIA RTX 3090
S3Net [144]	2024	IGARSS	1.403	9.579	-	NVIDIA Tesla V100
MEMF-Net [96]	2025	JSTARS	1.18	6.64	-	NVIDIA RTX 4090
HDSM-Net * [89]	2025	ISPRS	2.09	18.6	-	NVIDIA RTX 4090
SRCV-Net [62]	2025	JSTARS	1.387	8.615	-	NVIDIA RTX A6000
DBMSMNet [63]	2025	JSTARS	1.48	7.41	490	NVIDIA RTX 4090

, a number of representative methods have been evaluated on this benchmark. For reference, Figure 11 presents the qualitative results of [18] on the US3D dataset, as reported in the original paper. In addition, Patil et al. [140] proposed the SatStereo dataset, constructed from WorldView-3 satellite imagery, further facilitating research in satellite-based stereo reconstruction.

Subsequent efforts have progressively expanded the availability of remote sensing stereo benchmarks through diverse annotation strategies. In 2021, Wu et al. [145] constructed the ISPRS 2021 dataset by integrating 3D reconstructions with LiDAR point clouds to generate sparse disparity annotations. In 2022, He et al. [58] leveraged GaoFen-7 stereo image pairs and airborne LiDAR measurements to obtain dense disparity ground truth, further improving annotation completeness. To alleviate the difficulty of real-world annotation, Reyes et al. [146] proposed SyntCities, a synthetic remote sensing dataset generated using CityEngine and specifically designed for stereo vision research. More recently, Li et al. [19] introduced WHU-Stereo, a benchmark constructed from airborne optical imagery and LiDAR point clouds. As summarized in

Table 4. Evaluation of the stereo matching methods on the WHU-Stereo datasets. The downward arrow (↓) next to a metric indicates that lower values correspond to better performance. EPE: End-Point Error. D1: percentage of erroneous pixels. *: Denotes self-supervision.

Method	Year	Publication	EPE (px) ↓	D1 (%) ↓	Times (ms) ↓	Equipment
SGM [141]	2005	CVPR	4.88	50.79	506	Xeon
StereoNet [148]	2018	ECCV	2.45	25.12	238	NVIDIA Titan X
PSMNet [17]	2018	CVPR	2.48	24.81	614	NVIDIA Titan-Xp
RAFT-Stereo [14]	2021	3DV	1.77	14.35	575	NVIDIA RTX 6000
HMSM-Net [58]	2022	ISPRS	1.67	12.94	511	-
MaskCRNet [18]	2024	ISPRS	1.66	12.87	724	NVIDIA RTX 3090
MEMF-Net [96]	2025	JSTARS	1.524	11.6	-	NVIDIA RTX 4090
HDSM-Net * [89]	2025	ISPRS	3.310	29.0	-	NVIDIA RTX 4090
SRCV-Net [62]	2025	JSTARS	1.7523	14.15	-	NVIDIA RTX A6000
DBMSMNet [63]	2025	JSTARS	2.13	13.46	-	NVIDIA RTX 4090

, this dataset has become an important benchmark for evaluating representative remote sensing stereo methods. For reference, Figure 12 presents the qualitative results of [18] on the WHU-Stereo dataset, as reported in the original paper. Furthermore, Zhang et al. [147] proposed UAVStereo, a multi-resolution and multi-scene dataset collected from low-altitude UAV platforms, extending stereo benchmarks to aerial imaging scenarios.

5.2. Dataset Characteristics

Notably, remote sensing datasets exhibit several unique characteristics that substantially influence stereo matching performance. The main characteristics are listed in follows:

(1)

Ground Sampling Distance (GSD). Remote sensing datasets exhibit diverse spatial resolutions—for instance, US3D at approximately 0.3 m, WHU-Stereo at around 0.8 m, GaoFen-7 at about 0.8 m, and UAVStereo ranging from 0.05 m to 0.2 m. The GSD defines the spatial resolution of a remote sensing image, representing the real-world distance between two adjacent pixels on the ground. In satellite stereo imagery, GSD is determined by the imaging altitude H, focal length f, and sensor pixel size p, following the relation $\mathrm{GSD}=H\times p/f$. A smaller GSD provides finer spatial detail, allowing more accurate disparity estimation and 3D reconstruction, but also increases the data volume and computational burden. Conversely, larger GSD values reduce spatial precision and may obscure small objects such as vehicles or trees, thereby complicating stereo correspondence.

(2)

Acquisition mode and processing methods. A further source of variability arises from acquisition strategies. Certain datasets (e.g., US3D, SatStereo) are constructed from incidental multi-temporal views of the same geographic location. These inevitably introduce seasonal and illumination changes, thereby creating radiometric inconsistencies between the left and right images. By contrast, other datasets (e.g., WHU-Stereo, GaoFen-7, ISPRS 2021) are based on near-simultaneous acquisitions, which mitigate such inconsistencies but still retain geometric and atmospheric challenges.

(3)

Disparity range. Remote sensing datasets also exhibit distinctive disparity distributions. Unlike computer vision benchmarks such as KITTI (0–192) or SceneFlow (0–256), the disparity range in RS data is often narrower and may even include negative values (e.g., −64 to 64 in US3D). This property has direct implications for algorithm design: methods based on fixed disparity assumptions become less suitable, and instead adaptive cost volume construction, hierarchical search, or multi-scale disparity estimation strategies are required.

Table 5. Stereo matching dataset for remote sensing images.

Datasets	Year	Resolution	Mode	Total Data	Training Size	Testing Size	Label Type
US3D [21]	2019	$1024\times 1024$	RGB	4292	4242	50	Dense
SatStereo [140]	2019	$500\times 500$	Panchromatic	72	–	–	Dense
ISPRS 2021 [145]	2021	$1024\times 1024$	IRRG	1092	585	507	Sparse
GaoFen-7 [58]	2022	$1024\times 1024$	Panchromatic	490	400	90	Dense
SyntCities [146]	2022	$1024\times 1024$	RGB	8100	6480	1620	Dense
WHU-Stereo [19]	2023	$1024\times 1024$	Panchromatic	1757	1220	537	Dense
UAVStereo [147]	2023	$960\times 540$∼$8192\times 5460$	RGB	38,781	30,924	7757	Dense

provides a summary of commonly used RS stereo datasets, highlighting their resolutions, acquisition modes, and ground-truth types. For example, SatStereo and US3D rely on temporally displaced image pairs from the same region, thus introducing seasonal differences, whereas GaoFen-7, WHU-Stereo, and ISPRS 2021 leverage simultaneous image capture, which helps to avoid such inconsistencies.

5.3. Dataset Challenges

Based on these dataset-level properties, remote sensing stereo faces additional domain-specific challenges. We have summarized the following main challenges:

(1)

Limited data volume. Most RS stereo datasets are relatively small compared to terrestrial benchmarks, which restricts the training of deep stereo networks and limits their ability to learn complex spatial and semantic patterns.

(2)

Seasonal and environmental variations. In datasets with multi-temporal acquisition, differences in season, weather, and illumination between left and right images substantially complicate disparity estimation and require models with stronger generalization capability.

(3)

Distinct disparity statistics. The disparity distributions in RS imagery differ markedly from those of everyday scenes, diminishing the effectiveness of pre-training on large-scale synthetic datasets designed for CV applications.

(4)

High frequency of small targets. RS scenes often contain numerous fine-scale objects, such as buildings, trees, and vehicles. Accurate disparity estimation for such targets requires stereo models to maintain high spatial resolution and fine-grained matching precision.

(5)

Irregular and repetitive textures. Surfaces such as vegetation, water bodies, and urban facades frequently exhibit repetitive or ambiguous patterns. These characteristics create challenges for traditional correspondence search and necessitate algorithms that integrate semantic and geometric context.

Taken together, these characteristics make stereo matching in remote sensing a uniquely challenging problem, distinct from terrestrial scenarios. Addressing them calls for continued development of large-scale and well-annotated benchmark datasets, as well as the design of novel model architectures and learning strategies tailored specifically for remote sensing imagery.

6. Future Work and Suggestions

Despite the remarkable progress achieved in recent years, stereo matching for remote sensing imagery still faces critical challenges in accuracy, generalization, and computational efficiency. To address these limitations, future research should focus on two complementary aspects: (1) model design, to enhance robustness and adaptability across diverse remote sensing scenarios; and (2) dataset construction, to provide more representative and standardized benchmarks for algorithm development and evaluation.

6.1. Model Design

Future stereo matching models for remote sensing should be both accurate and efficient, capable of adapting to diverse imaging geometries and environmental conditions. Several directions are particularly promising:

(1)

Development of high-accuracy and remote sensing oriented algorithms. Remote sensing imagery exhibits distinctive characteristics such as seasonal variation, small object size, and repetitive textures across large areas. Current models, often adapted from terrestrial benchmarks, struggle to capture these patterns effectively. Future work should focus on designing specialized architectures that explicitly model the geometric, spectral, and temporal properties of remote sensing imagery, improving both robustness and precision in complex scenes. Meanwhile, the integration of large-scale vision foundation models and multi-modal satellite data (e.g., SAR–optical fusion) will likely define the next phase of RS stereo research.

(2)

Lightweight and efficient stereo matching networks. Although deep learning-based methods have achieved remarkable accuracy, their high computational cost limits practical deployment. With increasing demands for onboard processing on UAVs, satellites, and embedded systems, future models must achieve an optimal balance between accuracy, complexity, and runtime. Lightweight multi-scale architectures, efficient cost aggregation, and adaptive attention mechanisms could facilitate real-time inference in resource-constrained environments. Although 3D reconstruction tasks emphasize accuracy, edge-based inference is essential for onboard satellite or UAV applications requiring immediate geometric feedback.

(3)

Intelligent and adaptive learning under label scarcity. The limited availability of annotated stereo pairs and large domain gaps between terrestrial and aerial imagery restrict supervised learning. Future research should explore adaptive and self-evolving learning paradigms, such as self-supervised, weakly supervised, or meta-learning frameworks, integrated with foundation models and domain adaptation to enhance generalization across sensors and conditions.

(4)

Unified multi-task stereo matching frameworks. Stereo matching in remote sensing is often coupled with tasks like semantic segmentation, object detection, and 3D reconstruction. Developing unified architectures that enable mutual task reinforcement while maintaining interpretability and computational efficiency could lead to comprehensive and scene-aware 3D understanding in remote sensing.

6.2. Dataset Construction

Progress in remote sensing stereo matching also depends on the availability of diverse, high-quality benchmark datasets. Future datasets should aim to capture the complexity of real-world scenarios and support fair, scalable evaluation. Two key directions are outlined below:

(1)

Large-scale and diverse benchmarks. Future datasets should encompass a broad range of Ground Sampling Distances (GSDs)—from centimeter-level UAV imagery to meter-level satellite data—and include both simultaneous and multi-temporal acquisitions to assess robustness under illumination, seasonal, and geometric variations. Integrating negative disparity ranges and metadata such as viewing geometry or land-cover categories will further enhance their utility.

(2)

Task-driven and standardized dataset design. Beyond disparity estimation alone, future benchmarks should be constructed with downstream applications in mind, such as 3D reconstruction, digital surface modeling, and change detection. Building datasets with consistent annotation standards, unified evaluation metrics, and open-source protocols will foster fair comparison, reproducibility, and progress across the community.

Taken together, these future directions emphasize that advancing stereo matching in remote sensing requires synergistic efforts in both algorithmic innovation and data foundation—balancing accuracy, efficiency, and generalization toward practical and scalable deployment.

7. Conclusions

This survey has reviewed binocular stereo vision in remote sensing, encompassing traditional algorithms, deep learning-based models, acceleration strategies, and datasets. While deep stereo networks developed in the computer vision domain have achieved remarkable success, they face clear limitations in remote sensing, including poor cross-domain generalization, high computational costs for very high-resolution images, sensitivity to seasonal and illumination variations, difficulty in capturing small objects, and reliance on scarce labeled data. Future progress requires models tailored to remote sensing, with lightweight and multi-scale architectures, integration of geometric and semantic priors, self- and weakly supervised training to mitigate annotation scarcity, and improved robustness through attention mechanisms and domain adaptation. Addressing these challenges will enable more accurate and efficient stereo matching for applications such as 3D reconstruction, terrain modeling, and environmental monitoring.