The Korean Multi-Purpose Satellite 3A (KOMPSAT-3A) is equipped with a mid-wave infrared (MIR) sensor that acquires thermal imagery within the 3.3–5.2 μm spectral range. Unlike electro-optical (EO) sensors, which rely on solar reflection, MIR sensors detect thermal radiation emitted by objects and surfaces, enabling reliable operation under low-light and nighttime conditions (Mouats et al., 2015). This capability allows MIR sensors to operate reliably in diverse illumination environments, including nighttime, making them well-suited for thermal analysis in remote sensing applications. Consequently, MIR data have been extensively employed for detecting temperature variation, monitoring urban heat islands, assessing environmental changes, and identifying thermal anomalies in both natural and urban environments (Oh et al., 2020).

MIR imagery exhibits significant variations between daytime and nighttime acquisitions. Daytime images typically exhibit higher contrast due to the combined effects of solar reflection and thermal radiation, thereby enhancing object discrimination. In contrast, nighttime imagery relies solely on radiated surface energy, often resulting in reduced contrast and diminished object visibility compared to daytime images. Additionally, variations in thermal radiation and surface emissivity introduce radiometric inconsistencies, while differences in sensor attitude and acquisition time result in relative geometric dissimilarities, posing significant challenges for accurate image registration. Effectively aligning and integrating thermal information from both acquisitions necessitates a robust image registration framework capable of mitigating these discrepancies while preserving critical thermal features.

Image registration methods are generally classified into areabased and feature-based approaches, each with distinct advantages and limitations. Area-based methods operate by analyzing pixel intensity values between the reference and sensed images to identify corresponding points, typically using predefined templates (Chen et al., 2003). Commonly employed methods in this category include cross-correlation (CC), which measures the similarity of intensity patterns, and mutual information (MI), which evaluates the statistical dependency between intensity distributions. While effective under conditions with minimal geometric and radiometric distortions, these methods struggle with MIR imagery due to its inherently limited contrast and pronounced relative geometric dissimilarities. The limited contrast often obscures distinct intensity variations, making reliable template matching challenging. Additionally, temporal and environmental variations, such as fluctuations in thermal radiation and changes in surface emissivity, contribute to radiometric inconsistencies by altering the intensity distribution of MIR imagery, thereby reducing the reliability of area-based approaches.

In contrast, feature-based methods extract localized features, such as edges, corners, or key points, and encode them into descriptors, which represent these features in a format suitable for image similarity estimation (Zitova and Flusser, 2003; Huang et al., 2024). Widely used methods in this category include scaleinvariant feature transform (SIFT), speeded-up robust features (SURF), and oriented fast and rotated brief (ORB). These methods are inherently robust to geometric transformations, such as scaling, rotation, and affine distortions, as they rely on localized structural patterns rather than global pixel intensity consistency (Karami et al., 2017). As a result, feature-based approaches are well-suited for complex image registration tasks. However, applying feature-based methods to MIR imagery remains challenging due to its low contrast and homogeneous surface textures, which hinder the detection of robust features. Additionally, temporal disparities between daytime and nighttime acquisitions introduce radiometric variations, such as differences in thermal radiation and surface emissivity, as well as geometric variations resulting from changes in viewing conditions. These factors pose additional challenges to the effectiveness of conventional feature-based methods.

To address the challenges associated with radiometric inconsistencies, Li et al. (2018) proposed the radiant invariant feature transform (RIFT), a feature-based matching method designed to extract feature points that are robust to nonlinear radiometric distortions. RIFT leverages phase congruency (PC) in the frequency domain, which identifies perceptually significant features, such as edges and corners, independent of radiometric variations. This property makes RIFT particularly effective for multimodal image registration tasks, where differences in radiometric properties between datasets can hinder the performance of conventional feature-based methods. While RIFT provides robustness to radiometric variations, its application to daytime and nighttime MIR imagery remains challenging. The spectral and temporal characteristics of MIR images, such as low contrast and uniform textures, result in indistinct PC maps, regardless of contrast levels. Consequently, the original RIFT method struggles to extract reliable feature points in such cases, resulting in fewer robust correspondences. Furthermore, the lack of emphasis on prominent edge and corner features during PC map generation reduces its effectiveness when applied to MIR datasets with significant temporal disparities, such as daytime and nighttime image pairs.

These limitations highlight the need to enhance the original RIFT framework to effectively address the inherent challenges of daytime and nighttime MIR imagery. To address this, this study proposes a radiometric and feature contrast enhancement framework that integrates an improved RIFT methodology. The proposed approach includes a preprocessing step that applies gamma correction and Box-Cox transformation to normalize radiometric discrepancies, ensuring consistent input data for feature extraction. Furthermore, Gaussian weighting is applied during PC map generation to enhance contrast in edge and corner features, thereby improving feature detectability in lowcontrast regions. By addressing the limitations of the original RIFT method, the proposed framework offers a robust approach for the image registration of daytime and nighttime MIR imagery, enabling precise alignment across substantial spectral and temporal differences.

Various existing feature-based and area-based image registration methods have been explored for multi-modal image registration, including SIFT, SURF, and MI-based approaches. However, these methods struggle with low contrast and radiometric distortions inherent in MIR imagery, particularly in nighttime conditions. To address these limitations, this study proposes an enhanced registration framework that incorporates radiometric normalization and feature contrast enhancement. The framework aims to improve feature reliability and geometric alignment by applying radiometric correction techniques to stabilize pixel intensity distributions and refining feature extraction methods to enhance structural details in MIR imagery.

The proposed framework consists of two main stages (Fig. 1). The first stage focuses on radiometric normalization, where gamma correction and the Box-Cox transformation are applied to adjust intensity distributions and mitigate radiometric discrepancies. This preprocessing step ensures that feature extraction remains consistent across varying illumination conditions. The second stage enhances feature contrast and improves registration accuracy by refining the PC map using Gaussian weighting, which strengthens structural details. The RIFT algorithm is then used to extract robust feature points, followed by random sample consensus RANSAC filtering to remove outliers and improve matching reliability. Finally, an affine transformation is applied to align the nighttime MIR image with the daytime reference, ensuring precise spatial registration.

Figure 1. Overview of the proposed radiometric and feature contrast enhancement framework for daytime and nighttime MIR image registration.

2.1. Radiometric Contrast Enhancement

Daytime and nighttime MIR imagery exhibit significant radiometric discrepancies due to variations in illumination and radiometric properties. These differences result in inconsistent contrast and pixel intensity levels, which hinder effective feature detection, matching, and image registration accuracy. Nighttime MIR images, in particular, often suffer from low contrast due to their inherent radiometric characteristics, making the extraction of robust feature points challenging.

To address these issues, radiometric contrast enhancement was applied to mitigate nonlinear radiometric distortions and enhance feature visibility. This enhancement step compensates for radiometric inconsistencies, improves contrast between bright and dark regions, and ensures a more uniform pixel value distribution, thereby facilitating robust and accurate feature detection for image registration. The methodology involved two primary steps: gamma correction and the Box-Cox transformation. Initially, gamma correction was applied to compensate for the low contrast in the original image. The gamma correction function, shown in Eq. (1), modifies pixel intensity differences by adjusting the pixel value I_γ (x, y) based on the gamma value. As a nonlinear transformation, gamma correction compensates for radiometric inconsistencies, particularly enhancing feature visibility in low-contrast nighttime MIR imagery.

where I_γ (x, y) represents the gamma-corrected pixel value, I (x, y) is the original pixel value, I_max denotes the maximum pixel value, and γ is the gamma correction coefficient.

Gamma correction adjusts image contrast by modifying pixel intensity differences, particularly in darker and brighter regions, depending on the gamma value (Rahman et al., 2016). As a nonlinear transformation, gamma correction compensates for radiometric inconsistencies and enhances feature visibility, making it an effective preprocessing method for feature extraction, especially in low-contrast environments such as nighttime MIR imagery. When the gamma value is below 1, it reduces intensity variations in bright regions, preserving details in darker areas. However, this can further reduce contrast in low-intensity nighttime MIR imagery, complicating feature extraction. Conversely, a gamma value greater than 1 amplifies intensity differences, increasing contrast between dark and bright regions, enhancing object boundaries, and improving feature detection. For this reason, a gamma value exceeding 1 was applied in this study to enhance contrast and improve the distinction between image regions. This adjustment amplifies intensity variations, preserving critical structural details and enhancing nighttime MIR image registration.

Daytime and nighttime MIR imagery exhibit different pixel value distributions due to variations in illumination and radiometric properties, leading to inconsistencies in contrast and intensity levels. These discrepancies can hinder feature detection and matching, ultimately affecting image registration accuracy. Traditional intensity-based transformations, such as gamma correction, enhance contrast but do not effectively correct skewed pixel distributions. This can result in inconsistent feature extraction due to uneven variance in intensity values. To address this, the Box-Cox transformation is applied to stabilize variance, ensuring a more uniform pixel value distribution across the image. This step enhances the statistical reliability of feature extraction by minimizing radiometric disparities between daytime and nighttime imagery. The Box-Cox transformation mitigates these differences by reshaping the pixel value distribution, ensuring greater uniformity in intensity variations. This adjustment enables more consistent feature extraction, reduces radiometric disparities, and enhances the robustness of image registration under varying illumination conditions.

To address these challenges, the Box-Cox transformation was subsequently applied to stabilize variance in the image data and approximate a normal distribution. By adjusting nonlinear data using the parameter λ, the Box-Cox transformation reduces skewness and reshapes the data closer to normality (Box and Cox, 1964). This step complements gamma correction, which primarily enhances contrast but does not correct pixel value distribution biases. The mathematical formulation of the Box-Cox transformation is provided in Eqs. (2) and (3).

where I_γ represents the gamma-corrected image composed of multiple pixel values, I_γ, i denotes each individual pixel value within the image, expressed as I_{γ, 1}, I_{γ, 2}, …, I_{γ, n}. The parameter λ₁ controls the transformation, while λ₂ is an offset added to the pixel value. GM(I_γ + λ₂) represents the geometric mean of I_γ after adding λ₂, and I_BC(x, y) denotes the transformed pixel value after applying the Box-Cox transformation.

The parameter λ determines the nature of the Box-Cox transformation, controlling how pixel values are adjusted to stabilize the variance. By fine-tuning λ, pixel value distortions are minimized while maintaining a consistent variance, ensuring that the transformed data better conforms to a normal distribution. The optimal λ is estimated by maximizing the log-likelihood function, ensuring that the transformed pixel values approximate a normal distribution. This process improves the statistical properties of the data, enhancing the reliability of subsequent feature extraction. Daytime and nighttime MIR imagery exhibit different pixel value distributions due to variations in illumination and radiometric properties, leading to inconsistencies in contrast and intensity levels. These discrepancies can hinder feature detection and matching, ultimately affecting image registration accuracy. The Box-Cox transformation mitigates these differences by reshaping the pixel value distribution, ensuring greater uniformity in intensity variations. This adjustment enables more consistent feature extraction, reduces radiometric disparities, and enhances the robustness of image registration under varying illumination conditions. Fig. 2 illustrates the effects of preprocessing on daytime and nighttime MIR imagery, demonstrating improved uniformity in pixel value distribution after applying gamma correction and the Box-Cox transformation.

Figure 2. Daytime and nighttime MIR imagery before and after radiometric contrast enhancement. (a), (b) before preprocessing. (c), (d) after preprocessing.

2.2. Feature Contrast Enhancement

Although the RIFT method effectively extracts feature points that are resistant to radiometric distortions, its direct application to MIR imagery presents challenges. The resulting PC map often lacks sufficient clarity, making it difficult to distinguish between strong and weak features. This limitation hinders the ability to extract reliable key points, as important structural details are often obscured by noise or weak responses. The lack of welldefined feature contrasts reduces the overall robustness of feature detection, negatively impacting the accuracy of image registration. To overcome these limitations, an enhancement method is required to refine the PC map and improve feature detectability.

The conventional RIFT algorithm extracts feature points based on phase congruency, which is inherently robust to radiometric distortions. However, when applied to MIR imagery, the generated PC map often lacks sufficient clarity, making it difficult to distinguish strong and weak features. This results in unreliable keypoint detection, negatively impacting the accuracy of image registration. To overcome this limitation, Gaussian weighting is introduced to enhance strong features while suppressing weaker ones, thereby refining the PC map. This function, as described in Eq. (4), ensures that only the most structurally significant elements contribute to feature extraction, leading to improved feature matching between daytime and nighttime MIR imagery. By applying this weighting, the PC map becomes more refined, as illustrated in Fig. 3, enabling better differentiation between dominant and less significant features. The pixel intensity M(x,y) in the PC map was modified by applying a Gaussian weight, resulting in a transformed value (x,y), as defined in Eq. (5). This transformation enhances contrast in feature regions, making keypoints more distinguishable and improving the consistency of feature detection across varying radiometric conditions. Refining the PC map strengthens the robustness of the RIFTbased feature extraction process, leading to improved matching performance between daytime and nighttime MIR imagery.

Figure 3. Comparison of PC maps before and after enhancement. (a) PC map result from the original RIFT. (b) PC map results from the proposed method.

where G(x, y) denotes the value of the Gaussian function, where x and y are the spatial coordinates, σ represents the standard deviation of the Gaussian function, M′(x, y) is the pixel value after applying the Gaussian weight, and M′(x, y) is the original pixel value.

The RIFT method, which utilizes a PC map in the frequency domain, was employed to extract robust feature points resistant to nonlinear radiometric distortions, enabling the registration of preprocessed daytime and nighttime MIR imagery. Unlike conventional feature-based methods that depend on intensity gradients, the RIFT utilizes phase information, which is invariant to radiometric variations, making it particularly effective for multimodal image registration. This method begins by computing directional moments from the PC map to identify key structural features. Corner points are derived from the minimum moment map highlighting localized regions of significant intensity variation, while edge points are extracted from the maximum moment map capturing structural features. These moment maps ensure that distinct key points and edges are effectively detected, providing a detailed representation of structural elements. To further enhance feature characterization, Log-Gabor filters are applied at each feature point, generating responses across multiple orientations and scales. This multi-scale representation ensures that feature descriptors remain robust to geometric transformations. These responses are then used to compute the maximum index map (MIM), which encodes the dominant orientation and scale information for each feature point. By capturing both local and global structures, the RIFT enhances the reliability of feature matching between daytime and nighttime MIR imagery.

To construct the feature descriptor, patches surrounding each feature point are divided into sub-grids, where orientation histograms are computed to capture local directional properties. These histograms represent the distribution of dominant orientations, preserving critical structural information within each sub-region. The computed histograms are then concatenated to form the final feature vector, encoding both local and global structural characteristics of the image. This structured representation enhances feature robustness against radiometric variations and geometric transformations, ensuring reliable image registration under varying illumination conditions. By integrating fine-grained local details and broader spatial patterns, this method improves feature distinctiveness, enabling more accurate matching between daytime and nighttime MIR imagery.

For image registration, the proposed method was employed to extract matching points from the preprocessed daytime and nighttime MIR imagery. To ensure robust correspondences, the RANSAC algorithm was applied to remove outliers and retain only the most reliable inlier points, improving the accuracy of feature matching by minimizing errors from mismatches and noise. Using the filtered inlier points, an affine transformation model was estimated to align the nighttime MIR image with the spatial coordinates of the daytime MIR image. This transformation corrects geometric distortions and ensures that both images share a common reference frame, thereby completing the registration process and facilitating further comparative analysis.

3.1. Data and Experimental Setup

The accuracy and performance of the proposed image registration method were evaluated using four pairs of KOMPSAT-3A Level 1G daytime and nighttime MIR images acquired from different geographic locations: Australia, China, the Netherlands, and Russia. The specifications of the KOMPSAT-3A MIR images used in the experiment are provided in Table 1. Each image pair consists of a daytime image and a corresponding nighttime image captured on different dates. The spatial resolution of all images is 30 meters. In the registration process, the daytime MIR images served as reference images, while the nighttime MIR images were treated as sensed images.

Table 1 . Specification of KOMPSAT-3A MIR images used in the experiment.

No.	Location	Type	Acquisition date	Image size (pixels)
Site 1	Australia	Daytime	2016.11.01	518 × 519
		Nighttime	2016.11.29	517 × 519
Site 2	China	Daytime	2017.04.24	515 × 542
		Nighttime	2017.05.17	530 × 593
Site 3	Netherlands	Daytime	2022.03.09	551 × 699
		Nighttime	2022.02.28	569 × 640
Site 4	Russia	Daytime	2019.09.12	528 × 539
		Nighttime	2019.09.01	547 × 544

In this experiment, the proposed registration method was applied to align nighttime MIR images with their corresponding daytime reference images. The preprocessing step included gamma correction and the Box-Cox transformation to normalize radiometric discrepancies and refine intensity distributions, as described in Section 2. The enhanced PC map, generated with Gaussian weighting, was used for feature extraction using the RIFT algorithm. To ensure robust correspondences, outlier filtering was performed using the RANSAC algorithm, and an affine transformation was estimated for final alignment.

3.2. Evaluation and Validation

To evaluate the effectiveness of the proposed image registration method, 70% of the extracted matching points were used to estimate the transformation model, while the remaining 30% were designated as checkpoints for validation. While using a subset of matching points as checkpoints provides a practical approach for assessing algorithm performance, it may introduce biases toward specific transformation models or localized regions. To address this limitation, an additional evaluation was conducted using manually selected checkpoints from invariant regions within the images. This approach ensures that the assessment captures overall registration accuracy, minimizing the influence of localized variations.

To highlight the effectiveness of the proposed method, we compare its performance with the original RIFT method. RIFT was selected as the baseline comparison method because it is a feature-based image registration approach that is inherently robust to radiometric variations. However, RIFT alone does not explicitly enhance feature contrast or optimize keypoint extraction, leading to limitations in MIR image registration, especially under low-contrast nighttime conditions. Our proposed method integrates radiometric contrast enhancement (gamma correction and Box-Cox transformation) and feature contrast enhancement (Gaussian weighting on the PC map) to improve feature extraction and matching reliability. This comparison allows us to assess how these additional preprocessing steps enhance the robustness of registration results in multimodal MIR imagery.

A quantitative comparison was conducted between the original RIFT and the proposed method using key performance indicators to evaluate feature-matching accuracy and registration quality. The number of matching points (NMP) represents the total number of extracted correspondences, indicating the method’s ability to establish feature matches between the two images. The number of true matching points (NTMP) refers to the correctly aligned correspondences, serving as a measure of feature matching reliability. The relative root mean square error (RMSErelative) quantifies the registration error relative to the image scale and is computed as the RMSE of the transformed test points against their reference locations. The RMSEabsolute represents the absolute positional error in pixels and is calculated using manually selected checkpoints, providing a measure of spatial accuracy. Lastly, the 90% circular error (CE90) defines the radius within which 90% of the registration errors fall, offering an assessment of overall registration precision.

Table 2 presents the quantitative evaluation metrics for image registration using the original RIFT and the proposed method. Across all test sites, the proposed method resulted in a higher NMP and NTMP, indicating improved feature distinctiveness and a more reliable set of correspondences. Additionally, both RMSErelative and RMSEabsolute values were lower, signifying a reduction in registration error, while the CE90 values showed a substantial decrease, highlighting improved spatial alignment precision. Among the evaluated sites, Site 3 exhibited the most significant improvement, with the number of matching points increasing from 58 to 80 and CE90 decreasing from 3.456 pixels to 0.914 pixels. This indicates that the proposed method is particularly effective in areas with well-defined structural features, where feature contrast enhancement further refines keypoint extraction. In contrast, Site 2 showed the least improvement, likely due to the presence of featureless ocean regions in the imagery.

Table 2 . Quantitative evaluation metrics for image registration using the original RIFT and the proposed method.

Original RIFT	Site 1	Site 2	Site 3	Site 4
NMP	50	36	58	55
NTMP	26	21	35	29
RMSErelative (pixels)	2.981	2.779	2.880	2.956
RMSEabsolute (pixels)	2.761	3.248	1.841	2.949
CE90 (pixels)	5.019	5.666	3.456	6.174

Proposed method	Site 1	Site 2	Site 3	Site 4
NMP	61	42	80	74
NTMP	35	25	41	36
RMSErelative (pixels)	2.425	2.239	1.976	2.167
RMSEabsolute (pixels)	0.863	1.795	0.466	1.336
CE90 (pixels)	1.588	3.529	0.914	2.505

While the proposed method enhances feature detection and registration performance in MIR images, it still faces limitations in low-texture environments, where feature detection remains challenging. In particular, images with uniform textures, such as oceanic regions, tend to exhibit reduced keypoint detection performance. Similar findings have been reported in previous studies on feature-based image registration in low-texture environments. Zhao et al. (2024) demonstrated that traditional feature-based methods, such as SIFT and SURF, encounter significant challenges in ISAR image registration when applied to homogeneous regions with sparse structural information. These methods rely on distinct key points, which are scarce in featureless areas like large water bodies. This limitation aligns with our experimental results, where the proposed method exhibited reduced performance in the Chinese coastal region. The extensive oceanic areas in the MIR images likely contributed to the lower number of detected feature points and the increased registration error. These findings suggest that further improvements in feature extraction for low-texture environments—such as integrating deep learning-based adaptive feature detectors or multiscale keypoint selection strategies—could enhance registration accuracy in such challenging conditions.

A qualitative evaluation was conducted to visually assess the alignment of daytime and nighttime MIR imagery before and after registration. Fig. 4 presents a qualitative comparison of the registration results for four sites in a mosaic format, allowing for a direct visual analysis of the improvements achieved through the proposed method. The left column displays the overlaid images before registration, where noticeable misalignment is observed due to geometric distortions and radiometric inconsistencies. These discrepancies are particularly pronounced along highcontrast structural features, such as roads and river boundaries, where shifts in alignment are most visible. The middle column presents the results obtained using the original RIFT method. While partial improvements in alignment are evident, residual misalignment remains, particularly in low-contrast regions where subtle intensity variations hinder accurate feature matching.

Figure 4. Qualitative comparison of image registration results. The left column: before registration. The middle column: the original RIFT. The right column: the proposed method. Yellow circles highlight key feature points where alignment differences are most evident between methods. (a) Site 1. (b) Site 2. (c) Site 3. (d) Site 4.

The right column illustrates the registration results using the proposed method, which integrates radiometric and feature contrast enhancement to improve overall registration accuracy. Radiometric contrast enhancement, incorporating gamma correction and the Box-Cox transformation, ensures consistent intensity distributions between the two images, improving feature reliability. Additionally, feature contrast enhancement strengthens key structural elements, facilitating more accurate and stable correspondences. These improvements are particularly beneficial in challenging regions, such as areas with weak texture or low contrast, where the original method struggled. Compared to the original RIFT method, the proposed approach achieves more precise alignment along high-contrast regions and structural edges, significantly reducing feature displacement and improving geometric consistency. These qualitative results further confirm the effectiveness of the proposed method in overcoming the limitations of the original RIFT approach. The enhanced alignment, especially in regions with challenging radiometric conditions, highlights the advantages of integrating radiometric and feature contrast enhancement in strengthening the robustness of the registration process.

While the original RIFT method effectively detects robust feature points, the proposed method enhances its feature extraction performance in nighttime MIR images by incorporating additional contrast enhancement techniques such as gamma correction, Box-Cox transformation, and Gaussian weighting. These additional steps ensure more precise feature matching and better overall registration accuracy. By integrating Gaussian weighting, the proposed approach enhances the PC map, leading to improved keypoint localization and stronger feature extraction. By addressing both radiometric and geometric inconsistencies, the proposed method provides a more comprehensive solution than the original RIFT approach. The experimental results demonstrate that these enhancements contribute to more reliable feature matching, reduced registration errors, and improved spatial alignment accuracy.

In this study, we proposed a radiometric and feature contrast enhancement framework for the image registration of daytime and nighttime MIR imagery from KOMPSAT-3A. The approach consists of a preprocessing step to normalize radiometric differences and an enhanced feature extraction method to improve registration accuracy. Gamma correction was applied to enhance contrast, while the Box-Cox transformation stabilized variance and normalized the pixel distribution, effectively reducing discrepancies caused by varying radiometric characteristics and lighting conditions. To further improve feature detection and matching, the RIFT method was refined through feature contrast enhancement, which suppressed weak features while emphasizing prominent structural details. This enhancement improved the PC map, leading to more reliable feature extraction and robust correspondence matching. The proposed method was validated across four geographically diverse sites, demonstrating higher featurematching accuracy, reduced registration errors, and improved spatial alignment compared to the original RIFT method. These results confirm that integrating radiometric and feature contrast enhancement effectively addresses the challenges of daytime and nighttime image registration, leading to more consistent and precise alignment under varying environmental and radiometric conditions.

This study was supported by the Research Program funded by the SeoulTech (Seoul National University of Science and Technology).

No potential conflict of interest relevant to this article was reported.