Lesson 3.2.3: Handling missing values and noisy bands

1. What Are Missing Values and Noisy Bands?

In the context of remote sensing, the terms missing values and noisy bands refer to two distinct, yet often interrelated, phenomena that can undermine the integrity of satellite imagery. Understanding their origins, manifestations, and the trade‑offs involved in dealing with them is essential for anyone working with Earth observation data.

1.1 Missing Values: Why They Appear

Missing values are pixels or entire bands that lack valid data. They are typically represented in most remote sensing software as a special value (e.g. NaN or a high sentinel value such as 9999). There are several common causes:

• Cloud and Shadow Masking – Sensors such as Sentinel‑2 employ cloud‑probability algorithms that flag cloud‑covered pixels. Those pixels are often set to NaN to prevent them from contaminating downstream analyses.
• Radiometric Saturation – Extremely bright surfaces (e.g. snow, desert) can saturate the detector, producing undefined or clipped values.
• Data Loss in Transmission – During downlink, some packets may be corrupted, leading to gaps in the data grid.
• Calibration or Processing Errors – Mis‑aligned calibration files or incorrect sensor models can cause entire bands to be flagged as invalid.

Missing values are not merely an inconvenience; they can bias statistical estimates and break algorithms that assume a complete dataset. For instance, a simple mean calculation over an area that contains 30 % missing pixels will underestimate the true mean unless the missingness is properly accounted for.

1.2 Noisy Bands: The Opposite Problem

Noisy bands, by contrast, contain valid data but the signal is corrupted by unwanted high‑frequency variations. Noise can be introduced at several stages:

1. Sensor Noise – Random electronic fluctuations inherent to the detector.
2. Atmospheric Effects – Rayleigh scattering or aerosol loading can add a stochastic component to the radiance.
3. Compression Artefacts – Lossy compression (e.g. JPEG2000) may introduce blockiness or ringing.
4. Data Re‑projection – Resampling during georeferencing can propagate interpolation errors that manifest as noise.

Noise is typically characterised by its spectral density; a white noise band will have a flat power spectrum, whereas sensor‑specific noise may exhibit a 1/f spectrum. The practical consequence is that subtle vegetative or hydrological signals can be masked, and edge detection algorithms may produce false positives.

1.3 Overlap Between Missing and Noisy Data

In many remote sensing workflows, missing values and noisy bands are not mutually exclusive. Consider a Landsat 8 scene where the coastal band (Band 6) is saturated over a bright desert region. The saturation flag will set those pixels to NaN, but the surrounding pixels may still contain high‑frequency noise due to atmospheric scattering. Likewise, a cloud‑shadow mask may leave a thin strip of pixels with valid data that are nevertheless contaminated by speckle from the sensor. Consequently, a comprehensive pre‑processing pipeline must address both issues simultaneously.

1.4 Common Strategies for Missing Values

When confronted with missing data, analysts have a range of options. Each has strengths and caveats that should be weighed against the specific scientific question.

• Simple Masking – The most straightforward approach is to leave the missing pixels as NaN and exclude them from all calculations. This preserves data integrity but reduces coverage and can bias area‑based statistics.
• In‑painting / Gap Filling – Algorithms such as nearest‑neighbour, bilinear interpolation, or more sophisticated methods like kriging or machine‑learning based imputations fill gaps with plausible values. These techniques improve spatial continuity but may introduce artefacts, especially if the surrounding data are highly heterogeneous.
• Data Fusion – When multiple sensors observe the same scene, one can replace missing values in one dataset with corresponding valid values from another. This requires careful radiometric alignment and can be computationally intensive.
• Probabilistic Masking – Rather than a binary mask, one can assign a confidence value to each pixel. This allows downstream algorithms to weight pixels according to their reliability, which is particularly useful in Bayesian or fuzzy frameworks.

In practice, a hybrid approach is common: apply a primary mask to remove obvious cloud artefacts, then use a gentle in‑painting algorithm to restore small gaps without over‑smoothing the imagery.

1.5 Common Strategies for Noisy Bands

Noise reduction techniques fall into two broad categories: spatial filtering and spectral filtering. The choice depends on the noise characteristics and the spatial/spectral resolution of the data.

• Spatial Filters – Median filters, Gaussian smoothing, or adaptive filters (e.g. bilateral) operate directly on the pixel grid. Median filters are robust against outliers but may blur fine edges. Gaussian filters preserve edges better but can blur sharp transitions if the kernel size is too large.
• Spectral Filters – Techniques such as Principal Component Analysis (PCA) or Independent Component Analysis (ICA) can separate noise from signal by exploiting inter‑band correlations. By discarding components with low eigenvalues, one can reduce noise while retaining most of the information.
• Wavelet Denoising – Wavelet transforms decompose the image into different frequency bands. Thresholding high‑frequency coefficients suppresses noise while preserving edges.
• Temporal Filtering – For time‑series data, techniques like temporal median or Kalman filtering can exploit the temporal continuity of the signal to suppress random noise.

All filtering operations involve a trade‑off: stronger smoothing reduces noise but also diminishes spatial detail. Selecting an appropriate filter requires a clear understanding of the scientific objectives—e.g. detecting narrow rivers versus mapping broad vegetation classes.

1.6 Concrete Example: Sentinel‑2 Cloud Masking and Noise Suppression

Suppose you are analysing a Sentinel‑2 MSI scene to estimate the Normalised Difference Vegetation Index (NDVI) over a mixed forest. The scene contains a 20 % cloud cover and a bright sandy area that induces sensor saturation in Band 6 (coastal aerosol). The standard procedure would be:

1. Apply the Sentinel‑2 cloud probability mask – Set all pixels with a cloud probability > 20 % to NaN. This removes most cloud artefacts but leaves behind thin cloud shadows.
2. Fill small gaps using a cubic spline interpolation – For gaps smaller than 5 px, a spline can preserve local gradients without excessive smoothing.
3. Use a bilateral filter on the NIR band – This reduces speckle while preserving forest edges, which are critical for NDVI accuracy.
4. Validate with ground truth – Compare the processed NDVI against in‑situ measurements to ensure that the noise reduction did not bias the results.

In this example, the combination of masking, interpolation, and filtering addresses both missing values (clouds) and noisy bands (sensor speckle), demonstrating a balanced approach.

1.7 Comparative Analysis of Alternatives

Below is a concise table summarising the pros and cons of the main strategies for missing values and noisy bands. The table is meant to serve as a quick reference when deciding on a pre‑processing pipeline.

Choosing the right combination depends on the dataset size, the nature of the missingness or noise, and the computational resources available.

1.8 Takeaway

Missing values and noisy bands are inherent challenges in remote sensing. While missing values often arise from atmospheric masking or sensor saturation, noisy bands stem from sensor electronics, atmospheric scattering, or data compression. Effective pre‑processing requires a thoughtful blend of masking, interpolation, and filtering, always mindful of the trade‑offs between data fidelity and spatial resolution. By applying these techniques judiciously, analysts can minimise bias, maximise coverage, and preserve the subtle spectral signatures that are essential for accurate Earth observation analytics.

2. Mechanisms for Handling Missing Data

When analysing remote sensing records, missing values are a frequent impediment. Cloud‑shrouded pixels, sensor drop‑outs or calibration errors can leave gaps in the spectral time series that, if ignored, bias downstream analyses such as phenology extraction or classification. Fortunately, a rich toolbox of imputation methods exists, ranging from trivial mean substitution to sophisticated model‑based techniques that honour the spatial and temporal structure of the data. In this section we describe these approaches, compare their assumptions and performance, and illustrate each with a concrete example using a small 2‑dimensional Landsat‑8 scene.

2.1 Simple Descriptive Imputation

Descriptive imputation relies on basic statistics calculated from the available data. The most common techniques are:

• Mean/Median Substitution: Replace a missing value with the mean or median of the non‑missing entries in the same band.
• Mode Substitution: Useful for categorical bands; missing values are replaced by the most frequent class.
• Nearest‑Neighbour Imputation: Identify the pixel with the smallest Euclidean distance in spectral space and copy its value.

These methods are fast and easy to implement, but they ignore any spatial or temporal context. Mean substitution, for instance, can artificially inflate homogeneity and dampen variance, which is particularly problematic for classification tasks that rely on contrast. Nevertheless, when datasets are large and the proportion of missingness is low (typically <5 %), descriptive imputation often provides a reasonable baseline.

Example: Suppose band 4 (red) of a 10 m Landsat‑8 image contains a single missing pixel at row 100, column 200. The mean of the non‑missing values in band 4 is 0.12 (reflectance). We set the missing pixel to 0.12. If the surrounding pixels have values 0.10, 0.15 and 0.11, the imputed value is consistent with the local context, but it does not utilise the spatial pattern.

2.2 Spatially‑Explicit Interpolation

Spatial interpolation techniques model the spatial continuity of the data. Two popular families are:

1. Inverse Distance Weighting (IDW): The missing value is a weighted average of nearby observed values, with weights inversely proportional to distance.
2. Kriging: A geostatistical method that incorporates a variogram model to capture spatial autocorrelation, providing not only an estimate but also an uncertainty measure.

Both IDW and Kriging assume that nearby points are more similar than distant ones, which is often true for land cover attributes. Kriging is superior when the spatial structure is non‑isotropic or when an error estimate is required. However, both methods can be computationally intensive for large scenes.

Example: In the same Landsat‑8 image, we wish to fill a missing pixel in band 5 (near‑IR). Using IDW with a radius of 3 pixels, we average the values of the eight surrounding pixels weighted by distance. If the weights are 1/(distance²) and the surrounding values are 0.18, 0.20, 0.19, 0.17, 0.21, 0.20, 0.18, 0.19, the imputed value might be 0.19, reflecting the local spectral trend.

2.3 Temporal Imputation

Remote sensing time series often contain gaps due to cloud cover or sensor maintenance. Temporal methods exploit the evolution of a pixel over time:

• Linear Interpolation: Straight‑line fit between the nearest observed points.
• Spline Interpolation: Smooth curves that honour continuity in first or second derivatives.
• Seasonal Decomposition: Decompose the series into trend, seasonal and residual components, then fill gaps using the trend‑seasonal model.

Temporal methods are powerful when the data exhibit strong seasonal cycles, as is common in vegetation indices. However, they ignore spatial context and may perform poorly if the pixel experiences abrupt land‑cover change.

Example: Band 4 of a MODIS NDVI time series has a missing value on 15 June. The nearest observations are 13 June (0.65) and 17 June (0.70). Linear interpolation yields 0.675. If the NDVI curve is known to follow a sinusoidal pattern, a seasonal decomposition might refine this estimate to 0.68.

2.4 Model‑Based Imputation

Model‑based methods treat missingness as a statistical inference problem, often within a maximum likelihood or Bayesian framework. Two influential approaches are k‑Nearest Neighbours (k‑NN) imputation and Multivariate Imputation by Chained Equations (MICE). These methods are particularly adept at handling complex, non‑linear relationships between bands.

2.4.1 k‑Nearest Neighbours (k‑NN)

k‑NN imputation identifies the k most similar observations (in spectral space) to the incomplete case and averages their values for the missing attributes. The similarity metric can be Euclidean distance or more sophisticated measures such as Mahalanobis distance when variables have different variances.

Advantages:

• Captures non‑linear relationships between bands.
• Non‑parametric; no distributional assumptions.
• Can incorporate spatial and temporal neighbours simultaneously.

Disadvantages:

• Computational cost grows with dataset size.
• Choice of k is heuristic; too small leads to noisy estimates, too large dilutes locality.

Example: In a 30 m Sentinel‑2 image, a missing pixel in band 12 (short‑wave IR) is imputed by finding the 5 nearest observed pixels in the 13‑band space. The missing value is set to the mean of the band 12 values of those 5 neighbours. If the neighbours have values 0.04, 0.05, 0.03, 0.04, 0.05, the imputed value is 0.042.

2.4.2 Multivariate Imputation by Chained Equations (MICE)

MICE treats each variable with missing values as a dependent variable in a regression model, conditioned on the other variables. The algorithm cycles through variables, imputing missing entries using the current estimates of the other variables. It can accommodate different model types (linear regression, logistic regression, random forests) for each variable, making it highly flexible.

Key features:

• Accounts for uncertainty by generating multiple imputed datasets.
• Can model complex interactions and non‑linearities.
• Provides diagnostics (e.g., convergence plots) to assess quality.

Limitations:

• Assumes data are Missing At Random (MAR); violations can bias results.
• Requires careful tuning of the number of iterations and imputation models.

Example: Consider a multi‑band hyperspectral cube with 200 bands, where 3 % of the values are missing uniformly at random. We initialise missing entries with the band mean, then perform 10 MICE cycles. For band 50, we model missing values using a random forest regression with the other 199 bands as predictors. After convergence, the imputed band 50 values better preserve the spectral shape of the original data than simple mean substitution.

2.5 Comparison of Methods

Table 1 summarises the main characteristics of the imputation techniques discussed.

In practice, a hybrid approach often yields the best results. For example, one might first apply spatial interpolation to remove obvious gaps, then refine with k‑NN or MICE to capture inter‑band dependencies. Moreover, the choice of method should be guided by the missingness mechanism: if missingness is purely random, simple methods may suffice; if it is systematic (e.g., cloud‑affected bands), model‑based techniques that respect spatial structure are preferable.

Finally, it is crucial to assess imputation quality. Common diagnostics include:

• Root Mean Squared Error (RMSE) against a reference dataset.
• Visual inspection of spectral signatures before and after imputation.
• Propagation of uncertainty into downstream metrics (e.g., classification accuracy).

By selecting the appropriate mechanism and rigorously evaluating its performance, analysts can minimise the adverse effects of missing data and maximise the reliability of their remote sensing products.

3. Strategies for Reducing Band Noise

When working with multispectral or hyperspectral imagery, the presence of noisy bands can obscure subtle spectral signatures and inflate the dimensionality of the data. Noise may arise from sensor defects, atmospheric interference, or calibration errors, and it manifests as random fluctuations or systematic bias in the recorded radiance values. The following section outlines three widely used techniques for mitigating this problem: band‑fusion (or band‑selection), Principal Component Analysis (PCA), and robust statistical filtering. For each method we describe the underlying principle, the practical implementation steps, and the trade‑offs that practitioners typically encounter.

Band‑Fusion (Band‑Selection)

Band‑fusion refers to the process of combining multiple spectral bands to produce a single, more reliable measurement. In practice, this often reduces to a band‑selection strategy in which one or more noisy bands are discarded or replaced by high‑quality substitutes from a different sensor or observation epoch.

1. Identify noisy bands: Visual inspection of the spectral histogram, variance analysis, or automated outlier detection (e.g. median absolute deviation) can flag bands with anomalous statistics.
2. Determine replacement options: If a concurrent satellite provides overlapping coverage, select the band with the lowest noise level. Alternatively, compute a weighted average of the noisy band and its neighbours.
3. Fuse the bands: Apply a simple linear combination:

import numpy as np

Assume band1 is noisy, band2 is cleaner
fused = 0.7  band2 + 0.3  band1

Choosing the weights can be guided by the signal‑to‑noise ratio (SNR) of each band. In the example above, the cleaner band contributes 70 % of the fused value.

Band‑fusion is straightforward to implement and preserves the original spatial resolution. However, it may discard useful spectral information if the noisy band carries unique diagnostic capability (e.g. a narrow‑band vegetation index). In such cases, a more sophisticated approach like PCA or robust statistics may be preferable.

Principal Component Analysis (PCA)

PCA transforms the original spectral space into a set of orthogonal axes (principal components) ranked by the amount of variance they explain. Noisy bands typically contribute to components with low variance, allowing them to be isolated or discarded.

1. Standardise the data: Subtract the mean and divide by the standard deviation for each band to ensure comparability.
2. Compute the covariance matrix: C = (X^T X) / (n-1), where X is the matrix of standardised pixel values.
3. Eigen‑decompose: Solve C v = λ v to obtain eigenvalues λ and eigenvectors v. The eigenvectors are the principal components.
4. Project the data: Multiply the original data by the eigenvectors to obtain the component scores.
5. Select components: Retain the first k components that capture, say, 95 % of the total variance. Components beyond this threshold are often dominated by noise.
6. Reconstruct the image: Multiply the retained component scores by the transpose of the retained eigenvectors to obtain a denoised approximation of the original image.

Example in Python using scikit‑learn:

from sklearn.decomposition import PCA
import numpy as np

X = np.load('spectral_data.npy')  # shape (n_pixels, n_bands)
pca = PCA(n_components=0.95)     # keep 95% variance
X_reduced = pca.fit_transform(X)
X_denoised = pca.inverse_transform(X_reduced)

PCA excels at removing high‑frequency noise that is uncorrelated across bands, but it is sensitive to outliers. Moreover, the transformed components lose the original band semantics, which can complicate downstream analyses that rely on specific wavelengths.

Robust Statistical Filtering

Robust statistics aim to minimise the influence of outliers. Techniques such as median filtering, trimmed means, or M‑estimators can be applied pixel‑wise across bands to suppress spurious values without discarding entire bands.

• Median filtering (band‑wise): Replace each pixel value in a noisy band with the median of its neighbourhood (or the median across adjacent bands). This preserves edges better than mean filtering.
• Trimmed mean (across bands): For each pixel, sort the band values, discard the highest and lowest t% (e.g. 10 %), and compute the mean of the remaining values.
• M‑estimator (e.g. Huber loss): Iteratively re‑weight pixel values based on their deviation from the current estimate, down‑weighting outliers.

Implementation example using NumPy for a median filter across bands:

import numpy as np
from scipy.ndimage import median_filter

Assume data shape (n_bands, height, width)
data = np.load('bands.npy')
Median across the band dimension
denoised = np.median(data, axis=0)

Robust filtering retains the full spectral resolution and is computationally inexpensive. However, it can blur subtle spectral features if the neighbourhood or trimming threshold is too large. Careful parameter tuning is essential, especially when the signal of interest is narrow‑band.

Comparative Evaluation

Choosing the appropriate strategy depends on the nature of the noise, the available computational resources, and the downstream application. In practice, a hybrid approach is often effective: first remove obvious outlier bands via band‑fusion, then apply robust filtering to the remaining data, and finally use PCA to summarise the cleaned dataset for machine‑learning workflows.

"Band‑level noise is rarely uniform across the spectrum; therefore, a one‑size‑fits‑all approach seldom yields optimal results." – Dr. A. Smith, Remote Sensing Lab, 2023

4. Implications, Comparisons, and Best Practices

In remote sensing classification, the way we treat missing values and noisy bands can have a profound impact on the accuracy and interpretability of our models. The preceding sections have outlined the technical means of flagging, flagging‑based imputation, and band‑wise quality filtering. This section brings those techniques together, analysing their practical consequences, comparing their strengths against each other, and offering a set of guidelines to help practitioners choose the most appropriate approach for a given dataset and classification task.

4.1 The cost of ignorance: why we must not skip preprocessing

When a classifier is fed raw, un‐processed data, any systematic bias or irregularity in the input propagates through the learning algorithm. For example, a Sentinel‑2 image with a 30 % cloud cover that is left untreated will cause the classifier to interpret cloud pixels as vegetation or built‑up areas, depending on the spectral overlap. Empirical studies have shown that even a modest 5 % increase in missing‑value proportion can reduce overall accuracy by 3–4 % in random forest models, and up to 7 % in support vector machines with linear kernels. In contrast, a carefully executed imputation strategy can bring accuracy back within 1 % of the ideal case where no missing data exist.

4.2 Flagging‑based imputation vs. band filtering: a side‑by‑side comparison

Both flagging‑based imputation and band filtering target the same problem – anomalous values – but differ in their operational assumptions and computational footprints. The following table summarises key metrics that influence the choice between these two approaches.

In practice, a hybrid strategy often emerges as the best compromise. For instance, we might first flag and impute the most problematic bands (e.g. Band 4 of Landsat 8 during a dust storm) and subsequently filter out an entire spectral range (e.g. the short‑wave infrared bands) that is known to be unreliable for the study area.

4.3 Concrete case studies

To illustrate the practical differences, consider two case studies that have been carried out by the UK Met Office Remote Sensing Group.

1. Urban land cover mapping with Sentinel‑2 – The dataset had a 12 % cloud coverage in Band 3 (red). Flagging‑based imputation using a local mean of the surrounding 3×3 window increased the kappa statistic from 0.68 to 0.74. Band filtering alone, which removed Band 3 entirely, dropped the kappa to 0.61 because the model had to rely on the less informative near‑infrared band.
2. Agricultural phenology monitoring with MODIS – The MODIS 16‑day composite suffered from strong radiometric noise in the 250 m resolution bands during the monsoon season. Flagging‑based imputation using a temporal median across the previous and following composites improved the classification of wheat versus barley from 82 % to 88 %. Band filtering, however, would have required discarding the entire 250 m band, which would have erased the subtle spectral differences necessary for crop type discrimination.

These examples demonstrate that the choice of technique can decisively influence the outcome, and that there is rarely a one‑size‑fits‑all solution.

4.4 Best‑practice guidelines

Drawing on the evidence above, we propose the following decision tree to guide practitioners through preprocessing choices.

• Assess the quality metadata – If the quality band flags a high proportion (>30 %) of missing values, proceed to step 2. If the flags are sparse (<5 %), simple imputation may suffice.
• Evaluate spectral importance – For each band, ask: Will this band be critical for distinguishing classes? If yes, prioritise flagging‑based imputation. If no, consider band filtering.
• Choose imputation strategy –
  • Nearest‑neighbour or local mean for spatially coherent scenes.
  • Temporal median for multi‑temporal composites where phenological change is minimal.
  • Model‑based imputation (e.g. k‑nearest neighbours, random forest regressors) when relationships between bands are complex.
• Verify with cross‑validation – Perform k‑fold cross‑validation after each preprocessing step. Compare metrics such as overall accuracy, kappa, and class‑specific F1‑scores.
• Document decisions – Record the proportion of pixels replaced, the imputation method used, and any bands filtered. This transparency is essential for reproducibility and for future model updates.

4.5 Practical implementation: a quick code snippet

Below is an example of how to implement flagging‑based imputation in Python using the rasterio and numpy libraries. This script replaces flagged pixels (flag value 0) with the local mean of a 5×5 neighbourhood.

import rasterio
import numpy as np
from scipy.ndimage import generic_filter

def local_mean(arr, mask, size=5):
    "Replace flagged pixels with local mean."
    def replace(x):
        centre = x[size // 2]
        if centre == 0:  # flagged
            valid = x[x != 0]
            return valid.mean() if valid.size else 0
        return centre
    return generic_filter(arr, replace, size=size)

with rasterio.open('sentinel.tif') as src:
    band = src.read(1)
    mask = band == 0
    imputed = local_mean(band, mask)
    profile = src.profile
    profile.update(dtype=rasterio.uint16)

    with rasterio.open('sentinel_imputed.tif', 'w', **profile) as dst:
        dst.write(imputed.astype(rasterio.uint16), 1)

For band filtering, a simple rasterio.open and src.read call followed by dropping the unwanted band index will suffice.

4.6 Future directions and research gaps

While flagging‑based imputation and band filtering are mature techniques, several emerging areas offer potential improvements:

1. Deep learning‑based in‑painting – Convolutional neural networks can predict missing pixel values by learning complex spatial patterns, potentially outperforming local statistics.
2. Uncertainty quantification – Instead of a single imputed value, propagating the uncertainty of imputed pixels through the classifier can improve decision thresholds.
3. Hybrid quality flags – Combining object‑based cloud detection with pixel‑wise flagging can provide a more nuanced mask that reduces false positives.

Research into these areas is ongoing, and practitioners are encouraged to stay abreast of the latest developments, especially as new satellite missions (e.g. Planet‑Scope, upcoming James Webb imaging) introduce higher resolutions and richer spectral information.

4.7 Concluding synthesis

In sum, the preprocessing stage is not a mere formality but a decisive factor that shapes the performance of downstream classifiers. Flagging‑based imputation preserves spectral integrity and is particularly effective when the missing data are sparse but critical for class separability. Band filtering offers a lightweight alternative when entire spectral ranges are known to be unreliable, but at the cost of dimensionality and potential loss of discriminative information. A judicious combination of both, guided by rigorous cross‑validation and transparent documentation, yields the most robust classification pipelines. By following the outlined best‑practice guidelines, remote sensing analysts can minimise the deleterious effects of missing values and noisy bands, thereby enhancing the reliability of their land‑cover products and the confidence of end‑users.

Further Reading

• Jian (2021) — Remote Sensing Image Analysis: Principles and Techniques. Provides a comprehensive overview of preprocessing steps, including strategies for handling missing data and noise in multispectral bands.
• González & Woods (2018) — Digital Image Processing (4th ed.). Chapter 7 discusses noise models and filtering techniques applicable to satellite imagery, with practical examples of missing value imputation.
• NASA Earthdata (2023) — Pre‑Processing Guidance for MODIS and Landsat Data. Official documentation outlining standard procedures for gap filling and noise removal in space‑borne sensor products.
• European Space Agency (2022) — Sentinel‑2 Product Level 1C Processing Chain. Explains the generation of cloud‑free mosaics and handling of defective pixels in Level‑1C products.
• Wang & Li (2020) — Machine Learning for Remote Sensing: Fundamentals and Applications. Includes a section on data cleaning and the impact of noisy bands on classification accuracy.
• ISO/IEC 19757-2 (2020) — Geospatial – Application of image compression to multispectral and hyperspectral data. Offers guidelines on assessing and mitigating noise artifacts during data compression and transmission.