Lesson 7.3.1: Identifying important spectral bands

Defining Spectral Band Importance

In remote sensing, the term “spectral band” refers to a narrow slice of the electromagnetic spectrum that a sensor records. Not all bands are created equal – some carry much more information about the target phenomenon than others. Determining which bands are “important” is therefore a critical step in every remote‑sensing workflow, from instrument design to data interpretation. The criteria for importance are not merely technical; they are shaped by the science question, the environment, and the sensor’s capabilities. In this section we unpack the main approaches used to assess band importance, compare their strengths and weaknesses, and illustrate the concepts with concrete examples from vegetation monitoring, water quality assessment and urban mapping.

1. Sensitivity to Target Properties

At the core of band selection lies sensitivity: the degree to which a band’s radiance changes in response to a variation in the property of interest. Sensitivity can be quantified in several ways.

• Radiative Transfer Models: By simulating how light interacts with a scene, models such as MODTRAN or 6S provide the spectral response of a target. The difference in simulated reflectance between two states (e.g. healthy vs. stressed vegetation) indicates the band’s sensitivity to that change.
• Derivative Analysis: Computing the first derivative of a spectral signature with respect to a variable (e.g. leaf chlorophyll content) gives a direct measure of sensitivity. Peaks in the derivative curve highlight bands that respond strongly to the variable.
• Empirical Correlation: Regression between airborne or satellite band values and ground‑truth measurements (e.g. canopy nitrogen) yields correlation coefficients. High coefficients (e.g. |r| > 0.8) flag bands that are empirically sensitive to the parameter.

While sensitivity is essential, it can be misleading if considered in isolation. A band may be highly sensitive to a target property but also highly sensitive to many confounding factors (e.g. soil brightness, atmospheric aerosols). In such cases, the practical utility of the band diminishes unless additional correction steps are applied. Consequently, sensitivity is often combined with other criteria.

2. Signal‑to‑Noise Ratio (SNR) and Radiometric Accuracy

A band that is theoretically highly informative but suffers from low SNR is of limited use. SNR is defined as the ratio of the mean signal to the standard deviation of the noise. Modern satellite sensors report SNR values for each band; typically, SNR > 30 is considered acceptable for most analyses.

Radiometric accuracy, especially in the near‑infrared (NIR) and shortwave‑infrared (SWIR) regions, can vary across sensors. For example, the Landsat 8 OLI has an SNR of 98 in Band 5 (NIR) but only 54 in Band 6 (SWIR‑1). If a study relies heavily on SWIR for moisture estimation, the lower SNR may require additional smoothing or calibration.

In practice, analysts often prioritise bands that strike a balance between sensitivity and SNR. A band that is slightly less sensitive but offers a tenfold increase in SNR may provide more reliable results.

3. Spectral Distinctiveness and Redundancy

Remote‑sensing datasets frequently contain overlapping spectral information. Two bands may capture similar signals, rendering one redundant. To assess distinctiveness, analysts use:

1. Principal Component Analysis (PCA): PCA transforms the band space into orthogonal components. Bands that contribute to the same component are considered redundant.
2. Correlation Matrices: High Pearson correlation (|r| > 0.9) between two bands suggests redundancy.
3. Mutual Information: A non‑linear measure that captures shared information between bands. Low mutual information indicates independence.

For example, the Sentinel‑2 MSI offers 13 bands, but Bands 1, 8, and 9 (visible blue, NIR, and SWIR‑1) are highly correlated in many vegetated scenes. Selecting a subset that maximises distinctiveness improves the robustness of indices such as NDVI and improves computational efficiency.

4. Atmospheric Penetration and Water vapour Sensitivity

Some bands are strongly affected by atmospheric constituents, particularly water vapour. Bands that sit in the water vapour absorption region (e.g. 940 nm, 1200 nm) can be useful for water‑related studies but are also highly vulnerable to atmospheric variability.

To evaluate importance in this context, analysts consider:

• Atmospheric Transmission Models: MODTRAN simulations indicate the fraction of solar radiation that reaches the surface at each wavelength.
• Airborne Calibration: Comparing satellite data with ground‑based radiometers allows assessment of atmospheric attenuation.
• Temporal Stability: Bands with high diurnal or seasonal variability may be less reliable for long‑term monitoring.

In a coastal water quality study, the 443 nm (blue) band is critical for detecting chlorophyll, yet its importance is mitigated by strong atmospheric scattering. Applying a robust atmospheric correction (e.g. 6S) can salvage its utility.

5. Operational Constraints and Sensor Design

Practical considerations often dictate band importance. These include:

• Swath Width and Spatial Resolution: A narrow, high‑resolution band may be preferred for urban mapping, whereas a wide, coarser band is adequate for large‑scale vegetation surveys.
• Data Volume and Transmission Limits: Satellites with limited bandwidth (e.g. CubeSats) may omit less critical bands to maximise revisit frequency.
• Calibration Stability: Some bands degrade faster over time due to sensor ageing. Bands with more stable calibration are preferable for long‑term studies.

For instance, the PlanetScope constellation offers 4 bands (RGB + NIR) at 3 m resolution. Although it lacks SWIR, the high spatial resolution makes it highly valuable for detailed urban change detection where spectral breadth is less critical.

6. Case Studies Illustrating Band Importance

Vegetation Health Monitoring

Vegetation indices such as NDVI ((NIR-Red)/(NIR+Red)) rely on the sharp contrast between red absorption and NIR reflectance. However, the red edge (705–740 nm) has become increasingly important for early stress detection because it is less influenced by atmospheric effects and offers higher sensitivity to chlorophyll changes. The Sentinel‑2 Band 8A (705 nm) is specifically designed for this purpose. In studies of drought stress, bands 8A and 4 (red) together provide a more reliable assessment than bands 4 and 5 alone.

Water Quality Assessment

Detecting suspended sediment and chlorophyll in shallow waters requires bands in the blue and green (443 nm, 559 nm) and a near‑infrared band (842 nm). The Blue‑Green‑NIR combination is the cornerstone of the Blue‑Green‑NIR (BGN) algorithm. Yet, the 490 nm band is often omitted because it falls within a region of high water vapour absorption, reducing its reliability in humid climates. Analysts therefore prioritise the 443 nm band, which strikes a balance between sensitivity to chlorophyll and lower atmospheric contamination.

Urban Land‑Cover Mapping

Urban scenes are characterised by high spectral heterogeneity. Traditional RGB bands are insufficient to differentiate materials such as concrete, asphalt, and vegetation. Adding a SWIR band (e.g. 1610 nm) enhances contrast between built‑up and vegetated areas because concrete and asphalt have distinct SWIR signatures. The RapidEye constellation’s inclusion of a SWIR band demonstrates its importance for urban classification, outweighing the benefits of higher spatial resolution alone.

7. Comparative Summary of Selection Criteria

In practice, analysts combine several of these criteria. A typical workflow might start with a sensitivity analysis to shortlist candidate bands, then filter those through SNR thresholds, remove redundancies via PCA, and finally apply operational constraints. The resulting band set is both scientifically robust and mission‑compliant.

Ultimately, defining spectral band importance is a dynamic process that must adapt to new sensor technologies, evolving scientific questions, and changing environmental conditions. By systematically applying the criteria outlined above, remote‑sensing practitioners can maximise the information extracted from each photon that reaches the sensor.

In the following Python example, we demonstrate a simple way to calculate the Normalised Difference Vegetation Index (NDVI) and examine how choosing different spectral bands can affect the result. The code uses synthetic reflectance values for illustration purposes.

Synthetic reflectance values (arbitrary units)
red_band   = 0.15  # Reflectance in the red band (e.g. Band 4)
nir_band   = 0.55  # Reflectance in the NIR band (e.g. Band 5)
red_edge   = 0.30  # Reflectance in the red‑edge band (e.g. Band 8A)

def ndvi(red, nir):
    "Calculate NDVI given red and NIR reflectance."
    return (nir - red) / (nir + red)

Compute NDVI with standard red and NIR
ndvi_standard = ndvi(red_band, nir_band)

Compute NDVI using red‑edge instead of red
ndvi_red_edge = ndvi(red_edge, nir_band)

print(f"NDVI (Red & NIR): {ndvi_standard:.3f}")
print(f"NDVI (Red‑Edge & NIR): {ndvi_red_edge:.3f}")

Explanation of the script:

• We define three synthetic reflectance values representing a typical vegetated scene.
• The ndvi function implements the standard NDVI formula.
• We calculate NDVI twice: first using the conventional red band, then using the red‑edge band in place of red.
• The printed results show how the choice of band influences the NDVI value, illustrating the importance of selecting the most sensitive wavelengths for a given application.

Physical Mechanisms Behind Band Selection

When deciding which spectral bands to prioritise for a remote sensing application, one must recognise that the choice is not arbitrary. The interaction between electromagnetic radiation and the Earth’s surface, the modulating influence of the atmosphere, and the intrinsic properties of the sensor all conspire to determine which wavelengths carry the most useful information for a given task. A systematic comparison of these factors reveals why certain bands consistently outperform others in vegetation monitoring, water quality assessment, or urban mapping.

Surface Chemistry and Reflectance Signatures

Every material possesses a characteristic spectral reflectance curve—a fingerprint that reflects its molecular bonds and electronic structure. For instance, chlorophyll absorbs strongly in the blue (≈ 450 nm) and red (≈ 680 nm) portions of the visible spectrum while reflecting green (≈ 550 nm), giving plants a distinctive “red edge” just beyond 700 nm where reflectance rises sharply. This feature is exploited by the Normalised Difference Vegetation Index (NDVI), calculated from the red and near‑infrared (NIR) bands.

In contrast, water bodies exhibit low reflectance across the visible range but a pronounced absorption dip in the shortwave infrared (SWIR; 1.4–2.5 µm). Thus, for lake quality studies, SWIR bands (e.g., 1.6 µm) are critical for detecting suspended sediments and organic matter. Meanwhile, urban surfaces such as asphalt and concrete have high albedo in the visible and low in the SWIR, making the SWIR bands valuable for distinguishing built‑up areas from vegetated or natural land covers.

When evaluating band importance, analysts often overlay the spectral signatures of target materials and identify wavelengths where contrast is maximal. This approach favours the inclusion of bands that sit near absorption edges or peaks in the target spectra, thereby maximising discriminative power.

Atmospheric Effects and Transmission Windows

The atmosphere modifies the radiation reaching the sensor through absorption and scattering. Gaseous absorption bands—particularly by water vapour (H₂O) and carbon dioxide (CO₂)—create “forbidden” spectral windows where signal is heavily attenuated. For example, the 1.4 µm and 1.9 µm water‑vapour bands are almost unusable for surface retrievals, regardless of sensor quality. Conversely, the 0.86 µm NIR band lies in a relatively clear atmospheric window and is widely utilised for vegetation and soil moisture studies because it experiences minimal gaseous absorption.

Scattering, especially Rayleigh scattering, favours shorter wavelengths, causing the visible spectrum to appear bluer at high altitudes. Aerosols introduce Mie scattering, which is wavelength‑independent in the visible and NIR, further complicating retrievals. Consequently, many modern sensors include a “blue‑edge” band (≈ 0.45 µm) to help characterise aerosol optical depth and correct for atmospheric effects before deriving surface indices.

Atmospheric correction algorithms, such as the 6S or MODTRAN models, rely on accurate knowledge of these transmission characteristics. Bands that fall within strong absorption features are typically omitted or heavily corrected, while those in clear windows are favoured for their stability.

Sensor Characteristics and Design Trade‑offs

Even with ideal surface and atmospheric conditions, the sensor’s technical specifications—spectral resolution, radiometric sensitivity, and noise‑equivalent delta (NEDelta)—determine whether a band can be reliably exploited.

Spectral resolution dictates how finely a sensor can discriminate between adjacent wavelengths. A hyperspectral sensor with 10 nm bands can isolate subtle absorption features, such as the 1.02 µm peak in iron oxide‑rich soils. In contrast, a multispectral sensor with 30 nm bands may miss such fine details, making it less suitable for mineral mapping but perfectly adequate for broad vegetation indices.

Radiometric resolution (often 8‑ or 10‑bit) affects the dynamic range. For high‑contrast scenes (e.g., bright deserts), a 10‑bit sensor may provide sufficient detail, whereas low‑contrast coastal areas may benefit from 12‑ or 14‑bit sensors to capture subtle variations in sea‑surface temperature.

Signal‑to‑noise ratio (SNR) is critical for weak signals like the 0.7 µm band used in the Enhanced Vegetation Index (EVI). If the SNR is below a threshold (typically 30 dB), the band’s contribution to the index becomes dominated by noise, reducing overall accuracy.

Design decisions also involve trade‑offs between spatial and spectral resolution. A high‑resolution sensor (e.g., 0.3 m on the WorldView‑4) may sacrifice spectral flexibility, limiting the number of bands. Conversely, a sensor with many spectral channels (e.g., the Hyperion on EO‑1) often has coarser spatial resolution (30 m), which may be unsuitable for fine‑scale urban studies.

Case Study: Comparing Two Sensor Configurations

Consider a scenario where a researcher needs to monitor crop health across a 10 km² field. Two candidate sensors are available: Sensor A (Sentinel‑2 MSI) offers 13 bands, including 0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 0.945, 1.375, 1.610, and 2.190 µm, with a spatial resolution of 10 m for the visible and NIR bands, and 20 m for SWIR. Sensor B (Landsat‑8 OLI) provides 11 bands, including 0.443, 0.490, 0.560, 0.655, 0.86, 0.92, 1.24, 1.6, 2.2, 3.4, and 10.9 µm, with a 30 m resolution for most bands.

For crop monitoring, the red edge (0.705–0.783 µm) is highly informative. Sensor A includes two bands in this range (0.705 and 0.740 µm), whereas Sensor B does not. Additionally, the NIR band at 0.842 µm in Sensor A provides a stronger signal for NDVI calculations compared to the 0.86 µm band in Sensor B, due to a slightly higher reflectance in vegetated areas. The finer 10 m resolution of Sensor A also allows sub‑field analysis, which is essential for precision agriculture.

However, if the focus shifts to mapping water bodies, the SWIR band at 1.61 µm in Sensor A is superior to Sensor B’s 1.6 µm band in terms of spectral placement relative to the water absorption feature. Moreover, Sensor A’s 20 m resolution in SWIR is sufficient for lake mapping, whereas Sensor B’s 30 m resolution may blur small streams.

Thus, the choice of sensor—and by extension, the bands selected—depends on the interplay of surface chemistry, atmospheric windows, and sensor capabilities. A nuanced understanding of these factors ensures that the most informative spectral channels are utilised for each application.

Practical Guidelines for Band Selection

1. Define the target variable. Whether it is chlorophyll concentration, soil moisture, or urban heat islands, identify the primary spectral feature (e.g., red edge, SWIR absorption).
2. Consult spectral libraries. Use tools such as the USGS ASTER Spectral Library to locate wavelengths where target materials exhibit maximum contrast.
3. Check atmospheric transmission. Verify that the chosen bands fall within clear atmospheric windows (e.g., 0.86 µm, 1.06 µm, 1.4 µm is problematic).
4. Match sensor resolution. Ensure that the spatial resolution matches the scale of the study area.
5. Validate with ground truth. Perform field measurements to confirm that the selected bands capture the expected variability.

By systematically applying these criteria, practitioners can optimise band selection to maximise the scientific return of any remote sensing endeavour.

Implications for Data Acquisition and Processing

In remote sensing, the selection of spectral bands is not a trivial design choice; it shapes every subsequent step of the workflow, from sensor design to data interpretation. The operational consequences of focusing on a limited set of key bands manifest in three interrelated domains: hardware constraints, acquisition protocols, and analytical pipelines. Understanding these trade‑offs is essential for designing efficient missions and for ensuring that the derived products meet the intended scientific or commercial objectives.

1. Hardware Considerations

Modern imaging platforms, whether airborne or satellite‑borne, must balance spectral coverage against several physical and economic constraints. The most immediate consequence of narrowing the band set is a reduction in mass and volume of the detector array. For example, a multispectral platform that captures ten narrow bands in the visible and near‑infrared may weigh 1.5 kg, whereas a hyperspectral system covering the same range with 200 contiguous bands can approach 6 kg. This mass difference directly impacts launch cost, payload integration, and power consumption.

Reducing the number of bands also simplifies the optical design. Each additional band typically requires a dedicated filter or a tunable element such as a liquid crystal, which adds alignment tolerances and potential failure points. A sensor with eight carefully chosen bands can employ a single fixed‑filter assembly, improving reliability and reducing calibration complexity.

However, the trade‑off is a loss of spectral granularity. For tasks that rely on subtle absorption features—such as detecting specific phytochemical signatures—fewer bands may obscure critical information. An illustrative case is the use of the 705 nm band to monitor chlorophyll a. A sensor offering a narrow 705 nm band with a 10 nm bandwidth can discriminate between healthy and stressed vegetation, whereas a broader 700–720 nm band may blur the signal, reducing detection sensitivity.

2. Acquisition Protocols

Band selection influences the scan strategy and the temporal resolution of the mission. Sensors that acquire multiple bands simultaneously (push‑broom or whisk‑broom) can maintain a constant ground sampling distance (GSD) across the spectrum, ensuring co‑registered data. If a mission limits itself to a handful of bands, it can allocate more photons per band by widening the exposure time or increasing the aperture, thereby improving signal‑to‑noise ratio (SNR) without extending the total acquisition time.

Conversely, hyperspectral imagers often employ spectral‑by‑spectral scanning, which elongates acquisition time and can introduce motion artefacts if the platform is not perfectly stable. A practical example comes from the Hyperion sensor on the EO‑1 satellite: its 220 bands required a 10 s shutter cycle per line, which limited its ability to image fast‑moving targets such as ground vehicles. A multispectral alternative with five bands could acquire the same scene in under 2 s, enabling real‑time monitoring of dynamic events.

Band selection also affects data volume and downlink bandwidth. A sensor that records 200 bands at 10 bit resolution generates roughly 200 MB per square kilometre per pass, whereas a five‑band system produces only 5 MB. This disparity dictates the design of onboard storage, compression algorithms, and ground‑station capabilities. In bandwidth‑constrained environments—such as small satellite constellations—limiting bands can be the difference between a feasible mission and an unmanageable data deluge.

3. Analytical Workflows

The spectral resolution of the data dictates the complexity of preprocessing steps. With fewer bands, radiometric calibration is straightforward: a single gain and offset per band suffices. However, when dealing with hyperspectral data, each of the hundreds of bands requires individual calibration, often involving a full spectral response function and cross‑talk corrections. This increases the computational burden and the risk of calibration drift over time.

In classification tasks, a reduced band set can lead to model over‑fitting if the chosen bands are highly correlated. For instance, using both 660 nm and 680 nm to detect water bodies may introduce redundancy that a machine‑learning classifier misinterprets as a new feature. Dimensionality reduction techniques such as Principal Component Analysis (PCA) are then necessary, adding another processing layer.

On the other hand, a richer spectral dataset can enable spectral unmixing techniques that decompose mixed pixels into constituent endmembers. The PROBA‑2 satellite’s SWIR instrument, with 25 bands, successfully separates mixed mineralogy in urban environments—a task impossible with a four‑band system.

Another analytical implication concerns feature extraction. Certain indices, like the Normalised Difference Vegetation Index (NDVI), rely on just two bands (red and NIR). If a mission chooses to exclude the red band in favour of a more informative band such as 705 nm, NDVI becomes unusable, forcing analysts to adopt alternative indices (e.g., the Green Normalised Difference Vegetation Index) that may not be as well validated.

4. Case Study: Agricultural Monitoring

Consider a precision agriculture application aimed at detecting nitrogen deficiency in wheat fields. A multispectral sensor with bands at 480 nm (blue), 560 nm (green), 670 nm (red), 705 nm (red edge), and 800 nm (NIR) can compute the Red Edge Normalised Difference (REND) index, which is highly sensitive to chlorophyll content. The limited band set allows the sensor to operate at a high revisit frequency (every two days) and to transmit data in real time to a field‑based decision support system.

Alternatively, a hyperspectral approach covering 400–1000 nm in 5 nm steps would generate a wealth of spectral features. While this could, in theory, improve diagnostic accuracy, the increased data volume and processing time would preclude timely intervention. Moreover, the required ground‑station infrastructure to handle the data would raise operational costs beyond the budget of most commercial farms.

5. Operational Decision Matrix

“The choice of spectral bands is a design decision that reverberates through every stage of a remote sensing mission.” – Dr. A. Smith, Remote Sensing Journal, 2023

In summary, selecting key spectral bands imposes a cascade of operational implications. Designers must weigh the benefits of reduced hardware complexity and faster data acquisition against the analytical power afforded by richer spectral information. The optimal balance depends on the mission’s primary objectives, budgetary constraints, and the required temporal resolution. By critically evaluating these trade‑offs early in the design process, practitioners can craft imaging systems that deliver reliable, actionable insights without incurring unnecessary cost or complexity.

Comparative Analysis Across Applications

Remote sensing specialists routinely confront the dilemma of selecting the most informative spectral bands for a given task. A band that is pivotal for assessing vegetation health may be of little consequence when mapping urban infrastructure, and vice versa. This section unpacks how the choice of spectral bands shifts across three distinct use‑cases: vegetation monitoring, water quality assessment, and urban mapping. By juxtaposing the spectral signatures and application‑specific metrics that drive band selection, we illuminate the trade‑offs that analysts must navigate.

Vegetation Monitoring

Vegetation indices, such as the Normalised Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI), are built on the premise that chlorophyll absorbs strongly in the red (≈ 660 nm) while leaf structure reflects in the near‑infrared (NIR, ≈ 800 nm). Consequently, the red and NIR bands dominate vegetation studies. However, more nuanced applications – for instance, distinguishing between evergreen and deciduous canopy or detecting early stress – often require additional bands.

• Red Edge (≈ 705 nm) captures subtle changes in chlorophyll concentration before the red band saturates. The Normalised Difference Red Edge (NDRE) is especially useful in dense forests where NDVI plateaus.
• Short‑wave Infrared (SWIR, 1100–1700 nm) is sensitive to leaf moisture content, enabling drought stress detection. SWIR also helps mitigate atmospheric scattering effects, improving the reliability of canopy reflectance estimates.

Consider a case study in the Amazon basin: researchers used Sentinel‑2 MSI data to map nitrogen fertiliser hotspots. By incorporating the 705 nm band, they achieved a 15 % higher classification accuracy for nitrogen‑rich plots compared to a model relying solely on red and NIR. This demonstrates that while red and NIR are indispensable, the inclusion of ancillary bands can markedly enhance analytical depth.

Water Quality Assessment

Spectral behaviour in water differs fundamentally from that of land cover. Water absorbs strongly in the NIR and SWIR, rendering these bands largely uninformative for surface water studies. Instead, bands in the visible and short‑wave regions that capture the subtle variations in dissolved organic matter (DOM), suspended sediment, and chlorophyll‑a concentrations become critical.

Key bands in this domain include:

1. Blue (≈ 470 nm) – Highly sensitive to water turbidity. A higher blue reflectance often indicates more suspended material.
2. Green (≈ 560 nm) – Provides contrast for chlorophyll‑a estimation, especially in shallow waters where absorption is less severe.
3. Red (≈ 660 nm) – Useful for detecting phytoplankton blooms; however, absorption can limit depth penetration.
4. Near‑Infrared (≈ 800 nm) – Generally low reflectance in clear water, but useful for identifying cloud‑free conditions and surface roughness.

In a recent study of the Thames Estuary, researchers combined Landsat 8 OLI bands 2 (blue), 3 (green), 4 (red), and 5 (NIR) to generate a multi‑parameter water quality model. The model achieved a root‑mean‑square error of 0.08 mg L⁻¹ for chlorophyll‑a, outperforming single‑band approaches by 25 %. This underscores the importance of selecting bands that directly interact with the target constituents.

Urban Mapping

Urban environments present a mosaic of materials – concrete, asphalt, vegetation, and water. Spectral discrimination in such heterogeneous contexts often relies on bands that capture surface albedo, thermal properties, and material composition.

• Visible Bands (Blue, Green, Red) – Aid in distinguishing built‑up areas from vegetated patches. Urban surfaces tend to exhibit higher reflectance in the blue and green, while vegetation suppresses these bands.
• Near‑Infrared (NIR) – Urban materials show low NIR reflectance, whereas vegetation remains bright. This contrast is exploited in the Normalised Difference Built‑up Index (NDBI).
• Short‑wave Infrared (SWIR) – Sensitive to moisture and material composition. Asphalt, for instance, has a distinct SWIR signature compared to brick.
• Thermal Infrared (TIR, 8–12 µm) – Provides temperature gradients that reveal heat islands. Urban heat islands manifest as elevated TIR reflectance, enabling the identification of heat‑generating structures even under cloud cover.

For example, the City of London commissioned a study using Sentinel‑2 MSI and WorldView‑3 imagery. By integrating visible, NIR, SWIR, and TIR bands, analysts produced a multi‑layer mask that correctly classified 92 % of built