Editorial to Special Issue “Multispectral Image Acquisition, Processing and Analysis—2nd Edition”

1. Overview of the Issue: Multispectral Image Acquisition, Processing and Analysis—2nd Edition

Modern means of remote sensing (RS) in general and multispectral imaging in particular have experienced rapid development in recent years. Several factors facilitate the design and application of new imagers and methods for data processing. First, UAVs and drones have become possible carriers of such imagers, allowing one to operatively obtain data for desired regions from a wide range of altitudes and, respectively, with different resolutions. Second, convolutional neural networks and artificial intelligence approaches have sufficiently expanded the possibilities of data processing and the extraction of useful information. Third, more people and organizations are beginning to understand the capabilities and advantages of remote sensing and have began to exploit it for solving a wider set of practical tasks at the state and regional levels.

The chapters collected in this book and the 11 papers taken from the Special Issue “Multispectral Image Acquisition, Processing and Analysis—2nd Edition” of MDPI’s Remote Sensing journal reflect these tendencies. Below, we briefly describe these papers, focusing on the main benefits provided by the proposed solutions to different problems. Conditionally, the chapters can be divided into three groups: the first deals with more or less traditional approaches to multichannel RS data processing, the second includes papers considering neural network approaches, and the third relates to practical applications.

The chapter “Multiscale Fusion of Panchromatic and Multispectral Images Based on Adaptive Iterative Filtering” [1] considers the typical task of panchromatic and multispectral image fusion that has emerged, e.g., with the launch of the ZiYuan-3 (ZY-3) series of satellites. The core of the proposed fusion method is an adaptive smoothing filter that helps to solve a multiscale fusion task where conventional methods are often unable to handle the problems of spectral resolution loss and the reduced spatial resolution. The designed filter employs Gaussian convolution kernels instead of conventional mean convolution kernels to build a Gaussian pyramid. This enables one to adaptively construct convolution kernels of different scales for removing high-frequency information. The authors consider 15 known fusion methods and carry out a comparative analysis using experimental data acquired from a ZY-3 satellite. The authors demonstrate the benefits of the designed method, in particular, its ability to preserve good spatial information for image fusion at different scales with simultaneous appropriate preservation of spectral resolution. The proposed method produces a HQNR (hybrid quality with no reference) index of 0.9589 for the tested image and is able to fuse panchromatic and multispectral data of different scales for multispectral images acquired from various Chinese satellites. One more advantage is that the method is computationally efficient and can be applied onboard satellites based on embedded devices.

The chapter “BPG-Based Lossy Compression of Three-Channel Noisy Images with Prediction of Optimal Operation Existence and Its Parameters” [2] deals with the task of data volume reduction that is typical for many modern RS imaging systems. The peculiarity of this paper is that possible noise presence in acquired data is taken into account. In such a case, lossy compression is known to produce a specific noise filtering effect that can often produce a so-called optimal operation point (OOP), where a compressed image is closer to the corresponding true one than the original noisy image according to some quality criteria. The authors demonstrate that an OOP is possible if three-channel noisy images are compressed by a better portable graphics (BPG) encoder used in video processing. Moreover, the authors have proposed an approach to predict the existence of the OOP and the BPG encoder parameters in order to reach it. This allows one to simultaneously obtain a good quality compressed image and a rather large compression ratio in a fully automatic manner. The method is successfully tested for a set of three-channel images.

Relative radiometric normalization (RRN) of multi-temporal remote sensing images is essential for applications like change detection and image mosaicking. While existing sparse RRN methods have improved normalization accuracy, many approaches fail to account for different land use/land cover types or effectively handle nonlinear relationships between images. To address these limitations, the chapter “Automatic Relative Radiometric Normalization of Bi-Temporal Satellite Images Using a Coarse-to-Fine Pseudo-Invariant Features Selection and Fuzzy Integral Fusion Strategies” [3] proposes an automatic RRN technique utilizing clustered pseudo-invariant features (PIFs) through a coarse-to-fine strategy combined with a fusion-based modeling approach. The method introduces three innovations: a PIF selection strategy combining spectral correlation, a spectral angle mapper, and Chebyshev distance metrics; a cluster-wise robust linear regression (CRLR) approach; and a Choquet fuzzy integral fusion framework for reducing cluster discontinuities. The process involves two stages: a coarse stage generating a difference index to identify changed regions and a fine stage using histogram-based fuzzy c-means for image clustering and PIF selection through bivariate joint distribution analysis. The method then produces normalized images using both robust linear regression and CRLR methods, which are fused using the Choquet integral. Experiments on four bi-temporal satellite images and a simulated dataset demonstrate superior performance in accuracy and efficiency compared to existing algorithms.

Lossy compression of remote sensing data plays a vital role in managing large volumes of satellite imagery but faces challenges in balancing compression ratios with data quality and privacy requirements. While existing methods achieve high compression ratios, they often struggle to maintain classification accuracy or ensure data privacy. To this end, the chapter “Discrete Atomic Transform-Based Lossy Compression of Three-Channel Remote Sensing Images with Quality Control” [4] proposes an approach using discrete atomic transform (DAT) for compressing three-channel remote sensing images. The method introduces three key features, controllable image quality through maximal absolute deviation settings, various transform depths for different performance requirements, and inherent privacy protection, without additional computational overhead. The approach involves setting compression parameters that directly relate to traditional metrics like root mean square error (RMSE) and peak signal-to-noise ratio (PSNR). Experiments on multispectral Sentinel images demonstrate that classification accuracy using maximum likelihood remains largely unchanged across a wide range of compression ratios, though individual class accuracies may vary based on feature distributions.

As mentioned above, the next four chapters in the second group address neural network approaches and deep learning.

The chapter “A New Multispectral Data Augmentation Technique Based on Data Imputation” [5] addresses the computational challenges of DL in hyperspectral and multispectral image classification, where processing large scenes with high spectral resolution often requires extensive computational resources and energy consumption. While traditional patch-based classification methods have shown success, their demanding nature has led to the need for more efficient approaches. To address these limitations, this chapter proposes a data augmentation scheme that combines segment-based classification with data imputation and matrix completion methods. The approach is based on three key components: implementing superpixel segmentation to reduce computational complexity, utilizing specific algorithms to generate synthetic training samples, and integrating both original and synthetic datasets for CNN model training. Experiments conducted on two high-resolution multispectral datasets demonstrate the effectiveness of the method in improving classification performance across all tested scenes, while significantly reducing computational overhead. In addition, the obtained results indicate particular promise for applications requiring efficient processing of large-scale remote sensing data, while maintaining high classification accuracy.

The chapter “Unified Interpretable Deep Network for Joint Super-Resolution and Pansharpening” [6] introduces a novel deep network, UIJSP-Net, designed to enhance the spatial quality of multispectral images while preserving spectral integrity. Addressing the challenge of balancing spatial and spectral qualities in joint super-resolution and pansharpening (JSP), UIJSP-Net models JSP as an optimization problem based on a physical model that includes multispectral and panchromatic images. The model incorporates two deep priors to capture latent distributions, enhancing accuracy, and uses the alternating direction method of multipliers, with iterative steps translated into network stages through unfolding. Extensive tests on datasets from GaoFen-2, QuickBird, and WorldView-3 satellites show UIJSP-Net’s superior performance compared to existing methods. Additionally, an experiment with NDVI highlights its potential in remote sensing, and an ablation study examines the importance of each module. Future work will explore more efficient SR networks for handling higher resolutions.

With the increasing availability of high-resolution (HR) remote sensing data, super-resolution (SR) reconstruction has become essential for enhancing low-resolution (LR) images used in monitoring vast areas. The chapter “IESRGAN: Enhanced U-Net Structured Generative Adversarial Network for Remote Sensing Image Super-Resolution Reconstruction” [7] introduces the Improved Enhanced Super-Resolution Generative Adversarial Network (IESRGAN), an upgraded model based on U-Net for 4 × scale SR of LR images, specifically using NaSC-TG2 remote sensing data. IESRGAN integrates Reflective Padding and Residual in-Residual Dense Blocks (RRDBs) to enhance edge details and applies U-Net with spectral normalization in the discriminator, focusing on semantic and structural accuracy. Extensive tests on real-world datasets demonstrate IESRGAN’s strong generalization and stability, excelling in PSNR, SSIM, and LPIPS metrics. The method shows promise for practical applications like feature recognition and land detection. Future directions include exploring higher magnification levels, using multi-source data fusion, and applying IESRGAN to real-world tasks, such as land monitoring and object classification.

Semantic segmentation in high-resolution remote sensing images is crucial for applications like precision agriculture and disaster assessment. While convolutional neural networks have improved segmentation, many models fail to refine segmentation maps fully or exploit contextual dependencies. To address these issues, the chapter “HRRNet: Hierarchical Refinement Residual Network for Semantic Segmentation of Remote Sensing Images” [8] proposes the Hierarchical Refinement Residual Network (HRRNet), which utilizes ResNet50 as a backbone, along with attention blocks and decoders. HRRNet incorporates a channel attention module (CAM) and pooling residual attention module (PRAM) to capture contextual dependencies, enhancing segmentation precision. Feature maps from each ResNet50 layer are processed through attention blocks, with multi-scale fusion improving the segmentation of various ground object types. HRRNet shows improved results on the ISPRS Vaihingen and Potsdam datasets, outperforming other models. Future directions include refining segmentation for categories like low vegetation and trees, especially considering challenges such as large intra-category differences and minimal inter-category differences.

Finally, the third group of chapters relates to practical applications.

The Inner Niger Delta (IND) in Mali, a vital floodplain system supporting over three million people, relies heavily on seasonal flooding for agriculture, fishing, and livestock. Monitoring flood variations is challenging due to data scarcity and the large area. The chapter “Inner Niger Delta Inundation Extent (2010–2022) Based on Landsat Imagery and the Google Earth Engine” [9] utilized Google Earth Engine and Landsat imagery (2010–2022) to analyse flood extent trends, employing multiple water and vegetation indices in a supervised classification approach. The results indicate a positive trend in annual flood extent, with a minimum of 15,209 km² in 2011 and a peak of 21,536 km² in 2022. Upstream water extraction reduced the inundated area by about 6–10%. The method, confirmed by river discharge data, proves effective for flood monitoring in large, data-limited areas like the IND. Future work should incorporate radar and laser altimetry to enhance accuracy, study climate change impacts, and explore sustainable irrigation practices to balance wetland conservation with agriculture.

The chapter “UCTNet with Dual-Flow Architecture: Snow Coverage Mapping with Sentinel-2 Satellite Imagery” [10] introduces UCTNet, an efficient dual-flow architecture integrating CNN and Transformer models for snow and cloud classification in remote sensing (RS) multispectral imagery. Traditional methods struggle with snow-cloud differentiation due to similar spectral properties, and existing deep-learning approaches lack optimal feature extraction. UCTNet’s structure includes a CNN–Transformer Integration Module (CTIM) for enhanced data fusion and additional fusion modules for superior accuracy. Tested on a four-band RS dataset, UCTNet outperformed prior models (e.g., U-Net, Swin, CSDNet), achieving 95.72% accuracy and a mean IoU of 91.21%, with the smallest model size of 3.93M. It demonstrated a 2.30% increase in mIoU over CNN-based methods and outperformed Transformer-based models by up to 2.54%. Although effective, UCTNet faces challenges with Sentinel-2’s 10m spatial resolution, which limits classification at class boundaries and for small snow areas. Future improvements may include training on higher-resolution datasets to address these limitations for enhanced practical applicability.

The chapter “Applying Artificial Cover to Reduce Melting in Dagu Glacier in the Eastern Qinghai-Tibetan Plateau” [11] investigates using geotextile covers to mitigate glacier melt at Dagu Glacier No. 17 in China. From August 2020 to October 2021, covering a small glacier area (0.0005 km²) with geotextiles reduced ice mass loss by an average of 15% annually compared to uncovered areas. The effectiveness is attributed to the high albedo of geotextiles, reflecting more solar energy than glacier surfaces, which helps lower melt rates. Initial protection was higher, achieving a 27% reduction, but decreased to 8% over time due to aging and a decline in the cover’s albedo from 70% to 40%. Geotextiles may offer effective short-term protection for small glaciers or glacier termini but are limited by high costs, making large-scale applications challenging. Future research should explore the relationship between geotextile albedo and glacier ablation rates, assess the method’s effectiveness at different elevations, and consider integrating this approach with other glacier-protection techniques under various climate conditions.

2. Conclusions

The eleven papers featured in this Special Issue highlight the continued dedication of the remote sensing community to advancing innovative techniques that address the needs of end-users. These developments and enhancements are expected to further support the growth and refinement of remote sensing applications for operational purposes. We hope readers find this Special Issue both insightful and engaging.