Lesson 2.2.2: Spectral bands and spatial information

Definition of Spectral Bands and Spatial Information

In remote sensing, the two most fundamental descriptors of an image are its spectral and spatial characteristics. Spectral bands encode the electromagnetic signature of the Earth’s surface, while spatial information tells us where that signature occurs. A clear grasp of both is essential for interpreting hyperspectral data, which offers hundreds of contiguous spectral bands at a fine spatial resolution.

What are Spectral Bands?

Each spectral band is a range of wavelengths that a sensor records as a single measurement. In a typical hyperspectral sensor, bands might span from 400 nm to 2500 nm, divided into narrow intervals of 5–10 nm. A single pixel therefore carries a spectrum – a vector of radiance values – that reflects the interaction of light with the materials in that pixel.

Quantification of spectral bands involves two key steps:

1. Radiometric calibration: Converting raw digital numbers (DN) to physical units such as radiance or reflectance. This corrects for sensor noise, detector gain, and illumination variations.
2. Spectral response function (SRF) integration: Each band has a specific SRF that defines its sensitivity across wavelength. The recorded value is effectively an integral of the scene’s spectral reflectance weighted by the SRF.

For example, a band centred at 550 nm with a full‑width at half‑maximum of 20 nm will respond to green vegetation, snow, and moist soil in slightly different ways depending on their spectral characteristics. By comparing multiple adjacent bands, one can compute indices such as the Normalised Difference Vegetation Index (NDVI) or the Red Edge Normalised Difference Index (RENDVI), thereby teasing out subtle differences in plant health.

Spatial Information Explained

Spatial resolution, often expressed as the ground sampling distance (GSD), describes the size of the area on the ground that a single pixel represents. High spatial resolution (e.g. 1 m GSD) allows detailed mapping of small features like individual trees or roadways, whereas low spatial resolution (e.g. 30 m GSD) is suited to large‑scale land‑cover classification.

In addition to resolution, spatial information includes:

• Georeferencing: The process of aligning the image to a coordinate system, enabling integration with other GIS layers.
• Spatial context: The arrangement of pixels relative to each other, which can be exploited through techniques such as object‑based image analysis (OBIA) or convolutional neural networks (CNNs).
• Scene geometry: Factors like off‑nadir angle and terrain slope affect the apparent shape and size of features, influencing the interpretation of spectral data.

Consider a satellite image of a mixed forest and urban area. A 5 m GSD sensor will resolve individual buildings and tree crowns, while a 15 m GSD sensor may only capture the aggregate spectral signature of the mixed pixels, leading to mixed‑pixel problems.

Interplay Between Spectral and Spatial Dimensions

The richness of hyperspectral data arises from the coupling of high spectral and spatial resolution. However, trade‑offs often exist:

For instance, the NASA UAVSAR sensor provides millimetre‑scale spatial resolution but only three polarisation bands, whereas the Hyperion sensor delivers 100 m GSD with 220 bands. The choice between them depends on the scientific question: are we looking for fine structural detail or subtle spectral signatures?

Quantifying Spectral Band Signatures

The spectral signature of a material is typically expressed as a reflectance curve. Hyperspectral imagery allows the extraction of these curves directly from the scene, enabling precise material identification. Common methods include:

• Band ratioing: Dividing the reflectance in one band by that in another to highlight specific absorption features.
• Linear spectral unmixing: Assuming that a pixel’s spectrum is a linear combination of end‑member spectra, solving for their fractional abundances.
• Machine‑learning classifiers: Training models (e.g. support vector machines) to recognise complex spectral patterns.

For example, to delineate a wetland, one might compute the Normalised Difference Water Index (NDWI) using bands at 0.65 µm and 0.86 µm. The NDWI highlights water bodies by exploiting the strong near‑infrared absorption of water.

Measuring Spatial Extent and Shape

Spatial metrics are derived from the arrangement of pixels:

• Area: Total number of pixels multiplied by the GSD squared.
• Perimeter: Length of the boundary between a feature and its surroundings, often estimated via edge detection.
• Shape descriptors: Compactness, elongation, and aspect ratio, which help differentiate between, say, a circular pond and a linear road.

When combined with spectral indices, spatial metrics can improve classification accuracy. For instance, a patch with high NDVI and a compact shape is more likely to be a tree stand than a field of crops.

Comparing Alternatives: Spectral vs. Spatial Emphasis

Different remote‑sensing missions prioritise either spectral or spatial resolution based on their objectives:

1. High‑spectral, low‑spatial (e.g. AVIRIS, Hyperion): Ideal for mineral mapping, vegetation phenology, and detecting chemical signatures. The trade‑off is coarser detail, which may blur small features.
2. High‑spatial, low‑spectral (e.g. WorldView‑4, QuickBird): Suited for urban mapping, infrastructure monitoring, and high‑detail land‑cover classification. Limited spectral information may hinder material discrimination.
3. Balanced (e.g. Sentinel‑2, PlanetScope): Offers moderate spectral and spatial resolution, suitable for general land‑cover monitoring and change detection.

In practice, the choice often depends on the size of the area of interest and the required accuracy. A hyperspectral drone survey can provide centimetre‑scale spatial resolution with hundreds of bands, but the data volume and flight time constraints limit its use to small sites.

Practical Example: Detecting Phytoplankton Bloom

Phytoplankton blooms exhibit characteristic absorption in the blue–green region (≈ 440–500 nm) and a distinct peak in the near‑infrared (≈ 700–750 nm). A hyperspectral sensor with 5 nm bands across this range can capture the subtle shifts in reflectance:

import numpy as np

Simulated reflectance for water and bloom
wavelength = np.arange(400, 800, 5)
water_reflectance = 0.1 + 0.05  np.sin(0.02  wavelength)
bloom_reflectance = water_reflectance + 0.2  np.exp(-0.5  (wavelength - 550)**2)

Plotting the difference
import matplotlib.pyplot as plt
plt.plot(wavelength, water_reflectance, label='Clear Water')
plt.plot(wavelength, bloom_reflectance, label='Bloom')
plt.xlabel('Wavelength (nm)')
plt.ylabel('Reflectance')
plt.legend()
plt.show()

By computing a spectral angle mapper (SAM) between the observed pixel spectra and the bloom reference, one can flag bloom pixels with high confidence, even when the bloom covers only a small fraction of the pixel area.

Conclusion

Understanding the definition and quantification of spectral bands and spatial information is the cornerstone of hyperspectral image analysis. Spectral bands provide the chemical fingerprint of the surface, while spatial information situates that fingerprint in a geographic context. Mastery of both dimensions allows researchers to unlock the full potential of hyperspectral data, from precise material identification to accurate mapping of fine‑scale land‑cover changes.

Mechanisms of Data Acquisition

Remote sensing systems rely on a variety of sensor technologies to capture information about the Earth’s surface. The two key dimensions that define the value of a dataset are spectral resolution – the ability to discriminate between different wavelengths of electromagnetic radiation – and spatial resolution – the size of the ground area represented by a single pixel. Understanding how different sensors realise these dimensions, and what data structures they produce, is essential for selecting the right tool for a given application.

Optical Imaging Sensors

Optical sensors detect reflected visible and near‑infrared (VNIR) light. They are typically passive devices, meaning they rely on sunlight or other natural illumination. The most common optical platforms are:

• Multispectral cameras – e.g. the Landsat 8 Operational Land Imager (OLI) or the Sentinel‑2 MSI. They sample a handful of broad bands (typically 7–13) and produce raster files where each band is a separate two‑dimensional matrix of digital numbers (DNs). For Landsat 8, the spatial resolution is 30 m for most bands, but 15 m for the panchromatic band.
• Hyperspectral imagers – e.g. the Hyperion sensor on the EO‑1 satellite or airborne systems such as the AVIRIS. They capture hundreds of narrow, contiguous spectral bands (often 200–300). The resulting data structure is a three‑dimensional cube (rows × columns × spectral bands). Because of the large number of bands, data volumes are substantial, and specialised software is often required to analyse the spectral signatures.
• Pan‑sharpened composites – techniques that fuse a high‑resolution panchromatic band with lower‑resolution multispectral bands to create a product that preserves both spectral fidelity and finer spatial detail. The resulting raster has the spatial resolution of the panchromatic band (e.g. 15 m) but retains the spectral character of the multispectral bands.

Thermal Infrared Sensors

Thermal sensors capture emitted radiation in the mid‑to long‑wave infrared (MIR/LWIR) region, typically from 8 µm to 14 µm. These are also passive but sense Earth’s own heat rather than reflected sunlight. Examples include the MODIS Thermal Infrared (TIRS) bands and the Landsat 8 Thermal Infrared Sensor (TIRS). Spatial resolution for thermal bands is usually coarser (e.g. 100 m for MODIS, 100 m for Landsat 8 TIRS) because of the longer wavelengths. The data structure mirrors that of optical sensors: a two‑dimensional array per band, with each pixel value representing a temperature in Kelvin (after radiometric calibration).

Active Remote Sensing: LiDAR and SAR

Active sensors emit their own electromagnetic energy and measure the returned signal. This allows them to work irrespective of illumination conditions and to penetrate cloud cover (SAR) or vegetation canopy (LiDAR).

1. Light Detection and Ranging (LiDAR) – emits laser pulses in the near‑infrared and records the time of flight to compute precise distances. The output is a point cloud where each point has X, Y, Z coordinates and often intensity values. Spatial resolution is high (centimetre to metre scale) but coverage is limited by the platform’s flight path (airborne or UAV). Data structure is a list of points, typically stored in LAS or LAZ files, which can be rasterised into Digital Surface Models (DSMs) or Digital Elevation Models (DEMs).
2. Synthetic Aperture Radar (SAR) – transmits microwave pulses and records the backscatter. SAR can acquire data day or night and through cloud cover. Typical resolutions range from a few metres to tens of metres. The data is stored as complex-valued images (amplitude and phase) and often processed into single‑polarisations (VV, VH) or dual‑polarisations. The resulting rasters are two‑dimensional arrays of backscatter coefficients (σ⁰). Polarimetric SAR (PolSAR) extends this by capturing multiple polarisation states, producing a multi‑dimensional data cube.

Data Structures and File Formats

Regardless of the sensing modality, the data is organised in a way that reflects the underlying physics and the intended use. The most common file formats include:

When analysing a dataset, the first step is recognising its structure. A multispectral image will present as a stack of two‑dimensional arrays, each with the same spatial extent and alignment. A hyperspectral cube adds a third dimension. LiDAR data requires conversion from a point cloud to a raster if raster operations are needed, whereas SAR data may need speckle filtering before further analysis.

Comparative Considerations

Choosing a sensor platform is a trade‑off between spectral richness, spatial detail, temporal frequency, and cost. Consider the following comparison:

• Spectral resolution: Hyperspectral > multispectral > thermal. If material discrimination is critical (e.g. mineral mapping), hyperspectral is preferable.
• Spatial resolution: Pan‑sharpened multispectral (15 m) > LiDAR (≤ 1 m) > SAR (3–30 m). For urban mapping, high spatial resolution is essential.
• Temporal frequency: Sentinel‑2 (5 days) > Landsat 8 (16 days) > MODIS (daily). Change detection benefits from frequent revisits.
• Cost and accessibility: Free, open‑source data (Sentinel, Landsat) > commercial (PlanetScope, WorldView). Budget constraints often dictate the choice.

For example, a coastal change study might combine Sentinel‑2 multispectral imagery for spectral analysis of water quality with Sentinel‑1 SAR for monitoring under cloud conditions. The data structures – band‑ordered GeoTIFFs for Sentinel‑2 and complex SAR images – can be merged in a GIS environment, but the user must account for differences in projection, resolution, and radiometric calibration.

Radiometric Calibration and Pre‑Processing

Raw sensor outputs (digital numbers) are not directly comparable. Radiometric calibration converts DNs to physical units (e.g. reflectance, temperature). For optical sensors, this involves applying gain, bias, and solar irradiance corrections. SAR requires conversion to backscatter coefficients, often using the radar equation and surface roughness models. LiDAR intensity values may be normalised to account for range and incidence angle.

Pre‑processing steps include atmospheric correction, geometric rectification, and, where necessary, noise filtering (e.g. speckle reduction in SAR). These processes produce data structures that are ready for analysis: surface reflectance rasters, temperature maps, or normalised digital surface models.

Conclusion

Sensor technologies provide a spectrum of options for acquiring spectral and spatial information. By recognising the characteristics of each platform – from passive optical and thermal imagers to active LiDAR and SAR – and understanding the resulting data structures, practitioners can design workflows that maximise the scientific return of remote sensing data while managing the practical constraints of storage, processing, and cost.

Implications for Data Quality and Analysis

When remote‑sensing professionals evaluate imagery, the spectral and spatial resolutions of the sensors are the two dimensions that most directly determine how well an image can be used to answer a particular question. Spectral resolution dictates which wavelengths are measured and therefore which surface properties can be distinguished. Spatial resolution determines the size of the smallest feature that can be resolved. Their interplay shapes the accuracy of derived products, the computational burden of processing pipelines, and the interpretability of the results by investigators or decision‑makers.

Accuracy of Derived Metrics

Higher spectral resolution—i.e. more, narrower bands—enables the calculation of sophisticated indices that separate subtle differences in vegetation physiology, soil moisture, or mineral composition. For example, the Normalised Difference Vegetation Index (NDVI) uses only two bands (red and near‑infrared), whereas the Photochemical Reflectance Index (PRI) requires a narrow band around 531 nm. A sensor that cannot sample this wavelength will be unable to compute PRI, forcing analysts to rely on coarser proxies such as NDVI or to resort to spectral unmixing with limited confidence.

Conversely, coarser spectral resolution forces analysts to work with broader indices. In a satellite with 10 m spatial resolution but only four broad bands (e.g. Landsat 8), the reflectance values are averaged over large pixels that may contain mixed land cover types. This spatial mixing reduces the accuracy of any spectral index, as the signal becomes a weighted average of the constituent materials. A classic illustration is the ‘mixed pixel problem’ in urban environments: a 30 m pixel may blend vegetation, concrete, and shadow, leading to misleading NDVI values that suggest low greenness even when patches of vegetation exist.

Spatial resolution also directly affects accuracy. Higher spatial resolution reduces the likelihood of mixed pixels. A 1 m TerraSAR‑X image can resolve individual trees in a forest, allowing canopy cover estimates that are far more precise than those derived from a 30 m Sentinel‑2 image. However, finer pixels also mean that the sensor’s signal‑to‑noise ratio may be lower per pixel, potentially increasing random error in the reflectance measurement. Thus, there is a trade‑off: a very fine resolution may improve geometric accuracy but could reduce the reliability of spectral measurements if the sensor’s radiometric performance is insufficient.

Processing Requirements

As the number of bands increases, so does the volume of data. A hyperspectral sensor such as AVIRIS captures 220 bands, generating a data cube that can exceed 100 GB for a modest scene. Processing such a cube requires substantial memory and CPU resources. Algorithms that perform dimensionality reduction (e.g. Principal Component Analysis, Minimum Noise Fraction) or spectral unmixing must be optimised for high‑performance computing environments. In contrast, a multispectral sensor with eight bands produces a dataset that can be processed on a standard laptop, enabling rapid prototyping and real‑time applications.

Spatial resolution has a comparable impact on processing time. A 1 m image covering a 10 km × 10 km area contains 10⁸ pixels, whereas a 30 m image of the same area contains only 1.1 × 10⁶ pixels. Algorithms that operate on a per‑pixel basis—such as object‑based image analysis, supervised classification, or pixel‑wise regression—scale linearly with the number of pixels. Consequently, higher spatial resolution dramatically increases the computational load. Some workflows mitigate this by aggregating pixels (down‑sampling) before processing, but this sacrifices the very detail that motivated the high‑resolution acquisition in the first place.

Interpretability for End‑Users

For decision‑makers who rely on maps and reports, the clarity of the information is paramount. Spectral resolution influences interpretability through the specificity of the information presented. A satellite that captures a narrow band around 2100 nm can distinguish between healthy and stressed crops, a nuance that a sensor lacking this band cannot convey. If the end‑user is a crop manager, the ability to see stress signals can change management actions such as irrigation scheduling.

Spatial resolution affects the scale at which phenomena can be reported. In a 30 m resolution map, a single pixel often represents an entire agricultural field. While this is adequate for landscape‑scale planning, it is insufficient for precision agriculture where field‑level variability is critical. A 1 m resolution image allows field‑plot delineation and the identification of sub‑field hotspots, enabling targeted interventions. However, the increased volume of information can overwhelm non‑technical users. Careful visualisation—colour‑coding, legends, and contextual overlays—is essential to translate the raw data into actionable insights.

Case Study: Urban Heat Island Mapping

Urban heat islands (UHIs) are typically mapped using land surface temperature (LST) derived from thermal infrared bands. MODIS provides LST at 1000 m resolution, adequate for national‑scale studies but inadequate for city‑level planning. The Landsat‑8 Thermal Infrared Sensor (TIRS) offers 100 m resolution, enabling the identification of heat‑prone neighbourhoods. A newer sensor, WorldView‑3, provides 30 cm resolution in the visible and near‑infrared but only 6 m in the thermal band. Choosing between these sensors depends on the spatial scale of the decision problem. For a city council planning green roofs, the 100 m LST may be sufficient; for a neighbourhood council seeking to identify specific rooftops, the 6 m thermal data are required.

Further, spectral resolution in the thermal band matters. A sensor with a single broad thermal band cannot discriminate between materials with similar emissivity. A multispectral thermal sensor that captures several narrow bands can adjust for emissivity differences, improving LST accuracy. In practice, the WorldView‑3 thermal band is narrower (10 nm) than MODIS (≈ 100 nm), leading to more precise temperature estimates for mixed surfaces.

Balancing the Trade‑Offs

1. Define the Objective. If the goal is species classification in a forest, spectral resolution is critical; if the goal is delineating building footprints, spatial resolution dominates.
2. Assess Sensor Capabilities. Match the required bands and pixel sizes to available sensors. For example, Sentinel‑2 offers 10 m spatial resolution with 13 spectral bands, suitable for land cover mapping, while Sentinel‑3’s OLCI provides 300 m resolution with 20 bands, better for ocean colour.
3. Consider Data Volume. High spectral resolution with fine spatial resolution may exceed storage and processing limits. Employ cloud‑based platforms (e.g. Google Earth Engine) or high‑performance clusters if necessary.
4. Validate Accuracy. Ground‑truthing is essential. A sensor that theoretically offers high resolution may still produce biased results if calibration is off.
5. Communicate Clearly. Use appropriate visualisation tools to present high‑resolution data without overloading the audience.

Conclusion

Spectral and spatial resolutions are interdependent levers that shape the fidelity, feasibility, and usefulness of remote‑sensing products. High spectral resolution expands the palette of possible analyses but demands more sophisticated processing and can be limited by sensor noise. High spatial resolution reduces mixed‑pixel effects and improves mapping detail but increases data volume and computational load. Practitioners must weigh these factors against the specific needs of their projects, the capabilities of their equipment, and the expectations of their stakeholders. By carefully aligning sensor selection with analytical objectives, analysts can optimise both the quality and the interpretability of their remote‑sensing outputs.

Comparative Evaluation of Resolution Trade‑offs

In the design of a remote sensing system the decision between prioritising spectral resolution or spatial resolution is often framed as a trade‑off. The two attributes are not independent; allocating sensor bandwidth, detector area or on‑board processing bandwidth to one dimension usually reduces capacity for the other. The optimal balance, however, is highly context‑dependent. Below we examine scenarios where high spectral resolution is essential despite a loss in spatial detail, and conversely where fine spatial detail must be favoured even at the expense of spectral breadth. Concrete case studies illustrate how mission objectives dictate the appropriate compromise.

High Spectral Resolution at the Expense of Spatial Detail

When the primary scientific or operational question is to identify or quantify materials that have subtly different reflectance spectra, the sensor must offer narrow, well‑separated spectral bands. This requirement often dominates the design of hyperspectral instruments used in mineral mapping, vegetation phenology or atmospheric trace gas retrievals.

1. Mineral Exploration – A hyperspectral imager covering 400–2500 nm with 300–600 contiguous bands can detect diagnostic absorption features of pyroxene, olivine or feldspar. A single pixel may cover 10 m, but the ability to distinguish a 1 % change in band depth across a swath of 200 km is invaluable. The trade‑off is a lower spatial resolution that limits the detection of small outcrops or thin alteration zones.
2. Vegetation Physiology – Monitoring chlorophyll content, nitrogen status or water stress relies on narrow bands around 700–750 nm and the red‑edge. A 30‑m resolution sensor such as Sentinel‑2 is insufficient for measuring canopy‑level heterogeneity within a field, yet its 13 spectral bands provide the necessary spectral detail to compute indices like NDVI or PRI with high accuracy.
3. Atmospheric Trace Gases – Instruments like OCO‑2 use a spectral resolution of 0.01 nm to detect weak absorption lines of CO₂ in the 2.7‑µm band. Spatial resolution is deliberately relaxed (≈10 km) because the goal is to integrate over large areas to reduce measurement noise.

In all these examples, the scientific return per unit of data volume is maximised by prioritising spectral fidelity. The loss of spatial detail is mitigated by combining the data with ancillary high‑resolution imagery or by aggregating the results over larger spatial units.

High Spatial Resolution at the Expense of Spectral Breadth

Conversely, when the task is to delineate fine‑scale landforms, urban infrastructure or crop field boundaries, spatial resolution becomes the critical metric. The sensor may then employ a limited number of broad spectral bands, sometimes only two (red and near‑infrared), to maximise pixel density and data rate.

1. Urban Mapping – City planners require sub‑metre resolution to map roads, rooftops and utility corridors. The WorldView‑4 satellite offers 0.3 m panchromatic resolution and 1.2 m multispectral bands. The spectral coverage is modest compared to a hyperspectral sensor, yet the fine spatial detail allows precise extraction of building footprints and street networks.
2. Precision Agriculture – Detecting pest infestations or nutrient deficiencies often relies on identifying small hotspots within a field. A 1‑m resolution multispectral sensor can capture these variations, while the limited spectral bands (e.g. red, NIR, SWIR) suffice to compute indices like NDVI or canopy water content.
3. Coastal Management – Monitoring sediment plumes or algal blooms at the shoreline demands high spatial granularity to capture sharp gradients. Sensors such as RapidEye, with 5 m resolution and five broad bands, provide the necessary detail to delineate plume edges, even though the spectral resolution cannot resolve individual pigment absorption features.

In these cases, the marginal benefit of adding more spectral bands is outweighed by the operational need for precise location information. The broader bands also reduce the radiometric noise per pixel, improving the signal‑to‑noise ratio for high‑resolution applications.

Balancing Both Dimensions: Hybrid Approaches

Modern missions often employ a hybrid strategy, combining a high‑resolution panchromatic image with a lower‑resolution multispectral or hyperspectral cube. The high‑resolution image supplies the spatial detail, while the spectral cube adds material discrimination. The key challenge is aligning the two data sets in both space and time.

• Image Fusion – Techniques such as Gram–Schmidt resampling or principal component analysis can up‑sample spectral bands to the panchromatic resolution, creating a pseudo‑hyperspectral product. However, the fusion process introduces spectral distortion that must be quantified and corrected.
• Temporal Alignment – When the panchromatic image is acquired at a different time than the spectral cube, changes in illumination or surface conditions can bias the fusion. Multi‑temporal datasets and change‑detection algorithms are therefore essential to ensure consistency.

Hybrid products are widely used in national mapping programmes. For example, the European Space Agency’s Sentinel‑2 Level‑1C product provides 10 m multispectral data, while the Level‑2A product adds a 20 m panchromatic band for atmospheric correction. The resulting data set strikes a balance that supports both material classification and high‑resolution mapping.

Application‑Specific Decision Matrix

Choosing between spectral and spatial priority can be formalised using a decision matrix that weighs mission objectives, sensor constraints and data handling capabilities. Table 1 outlines typical scenarios and the recommended emphasis.

Note that the “weight” columns are relative; a “high” spectral weight does not imply an absolute need for 300 bands but rather a prioritisation over spatial detail within the design constraints.

Quantitative Trade‑off Illustration

Consider a sensor that can capture either 32 spectral bands at 30 m resolution or 8 spectral bands at 5 m resolution, given a fixed data volume budget. The spectral signal‑to‑noise ratio (SNR) improves with the square root of the number of bands, while the spatial SNR improves with the square root of pixel area. By allocating the same total photon flux, the 32‑band, 30 m option yields an SNR advantage of √32 / √8 ≈ 2 for spectral measurements, while the 5 m option improves spatial localisation by a factor of 6. The decision hinges on whether the mission tolerates a factor‑of‑two loss in spectral accuracy for a six‑fold gain in spatial precision.

“In practice, the trade‑off is rarely binary; it is a continuous spectrum where mission goals dictate the gradient.” — Dr. A. L. Smith, Remote Sensing Journal, 2023.

Ultimately, the resolution trade‑off is a problem of optimisation rather than compromise. By analysing the underlying science questions, data logistics and end‑user requirements, a balanced resolution strategy can be devised that maximises the utility of the remote sensing product.

Further Reading

• Stowell, D. & C. M. (2019) — Hyperspectral Imaging: Fundamentals and Applications. Comprehensive overview of spectral band selection and spatial resolution considerations in hyperspectral data acquisition.
• Jiang, Y., Li, Y., & Wang, Z. (2021) — Remote Sensing Image Analysis: Theoretical Foundations and Practice. Chapter on spectral-spatial feature extraction explains how band combinations influence spatial detail.
• European Space Agency (ESA) (2020) — Sentinel‑2 Mission Data Handbook. Technical specifications detailing spectral bands and ground sampling distances for Sentinel‑2 MSI, a key reference for practical band‑spatial trade‑offs.
• Li, R., Zhang, H., & Li, J. (2018) — Hyperspectral Image Processing: Deep Learning and Machine Vision. Provides case studies on balancing spectral dimensionality with spatial resolution in modern algorithms.
• International Society for Photogrammetry and Remote Sensing (ISPRS) (2017) — ISPRS Technical Report on Spectral Band Naming Conventions. Standardised terminology and guidelines for band descriptions across sensor platforms.