Review of Noise Removal in Passive Remote Sensing: Comparative MSI vs HSI Perspectives with Indian Case Studies

Abstract

The proliferation of passive remote sensing (RS) missions particularly multispectral (MSI) and hyperspectral (HSI) sensors has enabled wide-ranging applications in agriculture, water resources, urban monitoring, and climate studies. However, the presence of noise arising from atmospheric effects, sensor calibration errors, striping, Gaussian disturbances, and mixed pixels significantly degrades image quality, limiting the accuracy of downstream tasks such as feature extraction, classification, and spectral unmixing. While substantial research has focused on classification and dimensionality reduction, comparatively fewer efforts have systematically reviewed noise mitigation strategies. This review provides a comprehensive analysis of noise removal techniques in passive RS, spanning statistical filtering, band selection, ensemble learning, and advanced machine learning (ML) and deep learning (DL) approaches, including recent transformer-based architectures. We emphasize the comparative impact of noise in MSI versus HSI data as well as highlight Indian case studies with open datasets (Resourcesat, Cartosat, HySIS, Sentinel-2), and analyze emerging trends such as semi-supervised learning and multimodal fusion. By integrating classical methods with modern ML pipelines, this review offers a structured perspective on challenges, progress, and future research directions for robust and noise-resilient passive RS applications.

Introduction

• Passive Remote Sensing (RS) systems, including Multispectral Imaging (MSI) and Hyperspectral Imaging (HSI), are critical for Earth observation, particularly in agriculture, forestry, water management, and urban studies.

• MSI sensors (e.g., Sentinel-2, Resourcesat) offer broader bands and are good for large-scale mapping.

• HSI sensors (e.g., NASA’s AVIRIS, India’s HySIS) provide hundreds of narrow, contiguous bands—ideal for fine material discrimination but highly sensitive to noise.

• Despite the usefulness of RS data, noise from atmospheric interference, sensor errors, and pixel mixing significantly hampers their effectiveness.

2. MSI vs. HSI: Capabilities and Noise Sensitivity

• MSI is more robust to noise but offers lower spectral detail.

• HSI offers higher spectral discrimination but is more affected by noise and redundancy.

3. Main Noise Sources in Passive RS

1. Atmospheric Effects: Scattering, absorption, path radiance distort spectral signals.

2. Sensor Errors: Calibration drift, striping, thermal noise, and dead pixels.

3. Random Noise: Gaussian noise from electronic fluctuations.

4. Mixed Pixels: Pixels capturing multiple materials due to limited spatial resolution.

4. Noise in Indian RS Datasets

5. Noise Removal Techniques

A. Statistical Filtering

• Simple smoothing filters.

• SVMs and Random Forests offer inherent robustness to noisy, high-dimensional data.

B. Band Selection & Feature Reduction

• Techniques like PCA, ICA, and spectral band pruning reduce noise by discarding redundant/noisy bands.

C. Ensemble Learning

• Aggregates outputs of multiple classifiers to increase stability and reduce error sensitivity.

D. Deep Learning Approaches

• Spectral–spatial deep networks and transformers handle high-dimensional data, improving denoising.

• Learn robust features that suppress noise without losing important spectral information.

E. Semi-Supervised & Multimodal Fusion

• Leverages unlabeled data to learn noise-resilient features.

• Fusion with LiDAR (e.g., Rehman et al., 2025) improves robustness by integrating structural data.

6. Research Gap & Contribution

• While classification and feature extraction are well-studied, systematic approaches to denoising remain underexplored.

• This review fills the gap by:

  • Comparing MSI and HSI noise behaviors.

  • Surveying both classical and modern noise mitigation techniques.

  • Highlighting Indian RS datasets and case studies.

  • Proposing future directions in semi-supervised learning and transformer-based models.

Conclusion

This review synthesized developments in noise removal for passive remote sensing, highlighting sources of noise, comparative impacts on MSI and HSI, and the evolution of mitigation strategies from classical filtering to deep learning and multimodal fusion. A comparative perspective demonstrated that MSI, with fewer broad bands, is less sensitive to narrowband distortions and benefits primarily from preprocessing and robust classifiers, while HSI, with hundreds of contiguous channels, demands advanced ML/DL approaches to suppress amplified noise and redundancy. Indian case studies using Resourcesat, Cartosat, HySIS, and Sentinel-2 underscored that noise removal is not a peripheral step but a central requirement for reliable agricultural monitoring, environmental assessment, and disaster management. These studies illustrate how atmospheric corrections, semi-supervised learning, and deep models have improved resilience to noise in real-world applications. Looking forward, future research must move beyond algorithmic denoising to sensor-level innovations, including noise-aware detectors and in-sensor calibration routines that enhance data quality at acquisition. Coupled with physics-informed deep learning and multimodal fusion, such approaches will create noise-resilient pipelines that are both efficient and physically reliable. By combining algorithmic advances with sensor hardware integration and benchmark development, the remote sensing community can move toward robust, reproducible, and operationally impactful noise mitigation, particularly in the Indian context where open satellite data and applications are rapidly expanding. In this context, the contribution of this review is threefold: (i) to systematically survey noise sources and mitigation strategies in passive RS, spanning statistical filtering, band selection, ensemble learning, and advanced ML/DL methods; (ii) to provide a comparative perspective on how noise impacts MSI and HSI differently, drawing insights from both classical and modern approaches; and (iii) to ground the discussion in the Indian context by highlighting case studies that employ open datasets such as Resourcesat, Cartosat, HySIS, and Sentinel-2. By addressing these research gaps, the paper seeks to consolidate fragmented knowledge on noise handling, highlight emerging trends such as transformer-based architectures and multimodal fusion, and outline future directions for building noise-resilient remote sensing pipelines.

References

[1] Goetz, A. F. H., Vane, G., Solomon, J. E., & Rock, B. N. (1985). Imaging spectrometry for Earth remote sensing. Science, 228(4704), 1147–1153. [2] Richards, J. A., & Jia, X. (2006). Remote sensing digital image analysis: An introduction. Springer. [3] Plaza, A., et al. (2009). Recent advances in techniques for hyperspectral image processing. Remote Sensing of Environment, 113(S1), S110–S122. [4] Melgani, F., & Bruzzone, L. (2004). Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 42(8), 1778–1790. [5] Ham, J., Chen, Y., Crawford, M. M., & Ghosh, J. (2005). Investigation of the random forest framework for classification of hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 43(3), 492–501. [6] Briem, G. J., Benediktsson, J. A., & Sveinsson, J. R. (2002). Multiple classifiers applied to multisource remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 40(10), 2291–2299. [7] Waske, B., & Benediktsson, J. A. (2007). Fusion of support vector machines for classification of multisensor data. IEEE Transactions on Geoscience and Remote Sensing, 45(12), 3858–3866. [8] Zheng, Q., Wang, Q., & Yuan, Q. (2021). Hyperspectral image denoising by low-rank and sparse representations with spectral–spatial total variation regularization. IEEE Transactions on Geoscience and Remote Sensing, 59(1), 1–15. [9] Maxwell, A. E., Warner, T. A., & Fang, F. (2018). Implementation of machine learning classification in remote sensing: An applied review. International Journal of Remote Sensing, 39(9), 2784–2817. [10] Grewal, N., Yadav, M., & Singh, S. (2023). Deep learning for hyperspectral image classification: A review. Information Fusion, 91, 230–251. [11] Patel, N. R., et al. (2015). Assessing agricultural drought in Gujarat using AWiFS data of Resourcesat-2. Current Science, 108(3), 524–531. [12] Panigrahy, S., Manjunath, K. R., Ray, S. S., & Parihar, J. S. (2010). Mapping of cropping system for the Indo-Gangetic Plain using multi-date SPOT NDVI–A case study for Punjab. Current Science, 98(4), 502–508. [13] Roy, P. S., et al. (2016). Comparison of Sentinel-2 and Landsat-8 data for land cover mapping. Remote Sensing Letters, 7(5), 324–333. [14] Chavan, R. S., & Dhekane, A. M. (2018). Hyperspectral imaging satellite (HySIS): Opportunities for Earth observation in India. Journal of the Indian Society of Remote Sensing, 46(9), 1431–1440. [15] Dhekane, A. M., et al. (2021). Semi-supervised learning for hyperspectral image classification with limited labeled samples. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14, 883–896. [16] Kambli, N., & Palkar, R. (2024). Crop-specific classification of sugarcane using deep learning on Indian remote sensing data. International Journal of Remote Sensing. (in press). [17] Jha, C. S., et al. (2022). Forest fire monitoring in India using Sentinel-2 MSI: Challenges and opportunities. Environmental Monitoring and Assessment, 194(8), 558.

Copyright

Copyright © 2025 Rashmi Nitwane, Dr. Vaishali Bhagile . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.