Nocturnal optical remote sensing in the visible and near infrared (VIS/NIR) of the electromagnetic (EM) spectrum is largely inferior both to its daytime counterpart as well as to nighttime remote sensing in the thermal infrared. Even if there is a large gap in terms of the amount and the diversity of available missions and products, there exists demand for such nighttime products. The interest in such products is growing as evident from the increasing number of applications [1
]. These include the monitoring of human settlements and urban dynamics, the estimation of demographic and socio-economic information, light pollution and its influence on ecosystems and human health and astronomical observations, energy consumption and demands, detection of gas flares and forest fires, natural disaster assessment, and the evaluation of political crises and wars [2
]. Most of these applications are derived from data linked to artificial lights which emit mainly in the VIS/NIR. A stronger focus on optical nighttime remote sensing is, therefore, well-founded. However, the aim of the first satellite sensor imaging low-light data, namely DMSP-OLS in 1976 [3
], was to collect global cloud cover data day and night, detecting nocturnal VIS/NIR emission sources was a widely used by-product. For example, to derive essential socio-economic information, the lighting type is, however, a much stronger indicator of economic growth than solely the intensity of light as used in most studies [4
]. In 2011, a considerable improvement in spatial resolution from 2700 to
m and detection limits from 5 × 10
to 2 × 10
was possible with the arrival of its follow-on, the NPP-VIIRS-DNB yet with a daily global coverage [5
]. In addition to these panchromatic (500–900nm, [5
], for NPP) space-based nighttime images, trichromatic ones come in the form of photographs with a spatial resolution between 10 and
m taken by astronauts aboard the ISS irregularly since 2003 [6
]. Other panchromatic data are acquired only over China frequently with a spatial resolution of
m by LJ1-01 since 2018 [7
] and of
m by EROS-B sporadically since 2013 [8
]. Other data with multiple spectral bands are sporadically acquired only with a spatial resolution of
m by AC-5 (AC-4 with similar spectral resolution) since 2013 [9
] and of
m by JL1-3B (JL1-07/08 with similar spectral resolution) since 2017 [10
]. Furthermore, sporadically acquired nighttime images of operational daytime missions reveal detection limits, e.g., for Landsat-8, of above
only for the multiple spectral bands [11
The need for finer spectral and spatial resolutions was expressed many times. For example, a high-pressure sodium (HPS) lamp is indistinguishable from a light emitting diode (LED) lamp in panchromatic images and a conversion from HPS to LED is even incorrectly observed as a decrease in radiant flux for typical panchromatic images ([12
], for Milan, Italy). For example, a street with one lamp every
m is indistinguishable from a street with two lamps every
m resolution according to the Nyquist sampling theorem. Despite a proposal for a Nightsat mission in 2007 [13
], however, there is still no space-based nighttime VIS/NIR mission up with spatial resolution less than
m, multiple spectral bands, and a global coverage.
For daytime imaging, for example, with four spectral bands typically blue (457–523 nm), green (542–578 nm), red (650–680 nm), and NIR1 (784–900 nm) are recommended ([14
], for Sentinel-2) and a panchromatic band (450–800 nm). Furthermore, spectral bands red edge (705–740 nm) and NIR2 (960–1040 nm) are suggested. For nighttime imaging typically the same sensors are used. In order to determine optimal spectral and radiometric characteristics for dedicated nighttime VIS/NIR imaging, it is important to note that available data, for example based on airborne campaigns, provide only panchromatic imagery with high radiometric resolution ([15
], for Berlin, Germany) or high spectral imagery with only low radiometric resolution ([16
], for Las Vegas, NV, USA). Hence, these sources do not satisfactorily determine the optimal sensor parameters and performances; instead, an end-to-end sensor simulation is required with controlled environments to perform a realistic and precise examination.
The objective of this article is to recommend spectral and radiometric nighttime sensor parameters that are needed to support the community’s requirements, as well as those of the lighting engineering community and the general public, with a main focus on urban environments and the detection and differentiation of artificial outdoor irradiance sources at necessarily high spatial resolution.
Therefore, Section 2
analyzes the elements affecting the sensor signal, namely the natural (e.g., moon) and artificial (e.g., street light) nighttime radiation sources, their interactions with the surface (M2 and L1) and atmosphere (M1, M3, and L2), and the satellite sensor itself as illustrated in Figure 1
. For daytime similar considerations were widely investigated, e.g., by [14
]. For nighttime similar considerations were rarely investigated, e.g., by [17
] based on their findings on spectrometer measurements of outdoor lighting spectra. However, as only light source spectra were taken into account, a large part of the complexity is ignored by neglecting, for example, the variability in surface reflectances, atmospheric composition and sensor noise. For instance, two identical HPS or LED will look different when illuminating a patch of grass compared to a stretch of asphalt. Similarly, they will look different under hazy conditions compared to a clear night. Additionally, the number of spectral band combinations that the authors have considered was limited to eight and does not cover the full range of possibilities. Nevertheless, their recommendations are a reference for the considerations. Here, radiances are constructed which combines spectra from different lighting types, different intensities, different surface types, and different atmospheric compositions. In other words, it generates theoretical reference top-of-atmosphere (TOA) radiances for different conditions to know which signals arrive at a space-based sensor at night.
utilizes this data to answer the question, if it is possible to discriminate between different radiation sources from space-based images and at what spectral and radiometric resolutions. The complexity exceeds that of the traditional classification task, where the illumination source is known (e.g., sunlight) and the surface object types are unknown. In the nighttime case, the illumination source is also unknown and furthermore, irradiances produced by artificial lights are sometimes mixed with moonlight. It is, therefore, important to know how different moon characteristics affect TOA radiances and the discrimination of lighting types. Knowing the type of radiation source sheds light on a number of important light characteristics. Some essential considerations of lighting, however, are not linked to lighting type on a one-to-one basis. As a consequence, it is necessary to consider the dominant criteria in the planning of nighttime lighting, and how far these lighting quality indices, namely luminous efficacy of radiation (with radiant flux and luminous flux), spectral G index, and correlated color temperature, are derivable [18
]. Hence, the results of the simulations performed for various spectral and radiometric parameters and performances are analyzed, namely optimal spectral bands are derived (and typical TOA radiances are considered for the optimal radiometric resolution). While the radiances focus on homogeneous single-pixel environments, this approach does not, however, take into account any spatial information such as a lamp’s intensity distribution pattern or the overlapping of different lights. Therefore, also the spatial resolution is discussed.
Finally, Section 4
concludes the findings and recommendations for sensor parameters as well as the overview for future research.
As it is not effective to share full lamp spectral irradiances, technical descriptions of and management decisions on artificial lighting generally consists of only a limited number of performance parameters or indices. We investigate the most common spectral indices in lighting engineering, based on a report on road lighting and traffic signals of the European Commission [18
]. These define how much light is emitted, how much of the emitted light is seen by the human eye, or how much light is emitted in the blue part of the EM spectrum. In addition to the discrimination of lighting types, also the perceived color of the emitted light is assessed.
For optimization the MAE concerning the lighting parameters and F1 score concerning the lighting classification are minimized or maximized based on the at-sensor radiances only. Namely information on the radiation source, luminance, surface, or atmosphere are not considered while estimating the values.
3.1. Luminous Efficacy of Radiation
In designing artificial lighting, achieving a high luminous efficacy is crucial. Luminous efficacy rates the amount of visible light that is produced, in lumen, divided by the total amount electrical power that is required, in watt. Therefore, it is a measure for the efficiency of a particular luminaire system. Not only does luminous efficacy take into account emissions outside of the visual spectrum, but also, for example, decreased lumen as a result of dirt collections on the luminaires or electrical power losses in control gears. As luminous efficacy is impossible to be estimated without any ground-based information, it is often interchanged with luminous efficacy of radiation (LER), which is computed as the ratio between luminous flux
being the greatest luminous efficacy which can theoretically be achieved at
is the CIE photopic spectral luminous efficiency or the human eye’s relative sensitivity under well-lit conditions, and radiant flux
Thus, a choice of two bands is physically required for the estimation of LER. A panchromatic band B0 is used to estimate the irradiance emitted across the full EM spectrum (i.e., denominator of equation) and a (near) luminance band B1 is used to estimate the amount of irradiation which is visible to the human eye (i.e., numerator of equation).
Band B0 estimates the amount of emitted irradiance, namely the radiant flux and is itself a light parameter playing one of the most important roles that corresponds to a certain luminaire system. The estimation of the radiant flux from band B0, however, is not straightforward, as it depends on a number of different parameters, e.g., surface reflection, atmospheric transmittance, and the ratio of emitted power within the measured spectrum. As in typical situations surface reflection and atmospheric transmittance are essentially constant and similar factors to bands B0 and B1, however, they basically cancel for LER. As band B0 covers the full EM spectrum, the signal is high and sensitive to all considered radiation types; it also serves as a panchromatic band and is used for normalization purposes. Whereas it is possible to estimate radiant flux of lamps, uncertainties remain relatively high. Usually, rather than the radiant flux, it is the required electrical power that is of interest. However, estimating the latter is further complicated by the need for data on electrical power efficacy, which describes the ability to transform electrical power into optical power.
Band B1 estimates the amount of visible light, namely the luminous flux and is also an important light parameter by itself, its detector bandpass function shall closely resemble the photopic spectral luminous efficiency curve . However, the form of a detector bandpass function differs from the form of . The optimal spectral band B1 to resemble is reached for nm and nm, theoretically. Note that the CW differs slightly from the wavelength at which reaches its maximum, i.e., nm. This is ascribed to the asymmetrical form of having a slightly positive skew. It is expected that this shift to higher wavelengths will be less significant in practice, as most lamps have almost no emissions around the larger base of the function, i.e., nm.
The estimated LER, namely , is not only achieved by as these ratios do not yet represent LER values, but they need to be adjusted by applying a multiplication factor a. It is theoretically approximated by a . A practical approximation based on the ratios and LER using least-squares estimation is preferred.
In order to select the optimal parameters for B0 and B1, uniformly distributed sampling is used for the CW and FWHM. Based on the discussions, for B0, CW ranges between 500 and
nm and FWHM ranges between 300 and
nm, and for B1, CW ranges between 500 and
nm and FWHM ranges between 80 and
nm. Considering all possible combinations, the optimum is reached for the combination
nm (panchromatic) as well as
nm (green). Note that, as expected, B1 deviates slightly from those reached from approximating the photopic luminous efficiency function directly. An MAE for
is reached. For comparison, the mean LER for the test data equals
with a standard deviation of
. It is important to note that these LER are slightly higher than they will be in reality, since they are only based on the 350-900nm range, with data outside this range missing for the spectra. The theoretical estimation of a
results in 174, while the practical estimation yields
. Applying these analyses to the band B1 recommended by [13
nm, results in an MAE for
, indicating that the proposed bands offer a significant improvement for the estimation of the efficiency of artificial lighting.
Let us consider instead of the optimized the typical panchromatic band B0 with nm and nm. Here, the range 865–900 nm is covered, but it is relevant for fire and incandescent lamps only, where the radiant flux is only marginally emitted in VIS/NIR at all. However, using the optimized or typical panchromatic band, estimations on the radiant flux are derivable taking the lighting type into account. For all other lighting types there is almost no emission in this range. Here, the range 374–499 nm is not covered, but it is relevant for more than half of the considered lighting types, e.g., LED lamps produce strong emissions, however, HPS lamps do not.
3.2. Spectral G Index
Light pollution, especially in the blue part of the spectrum, has gained substantial attention. Examples are the disruptive effect of artificial lights on the nocturnal behavior of different species as well as on human health [27
]. For a long time correlated color temperature has been the principal indicator for the amount of emitted blue light, despite its inability to sufficiently describe the spectrum of a lamp. However, the European Commission has published a report in which it recommends the use of the so-called spectral G index instead [18
], which is computed as the total amount of luminous flux divided by the amount of radiant power emitted between 380 and
nm with high values corresponding to low blue light emissions, namely
. Note the similarities between the enumerator of this equation and the one of LER. Thus, a choice of two bands is physically required for the estimation of G, with a focus on the amount of blue light. However, the enumerator is estimated by the same spectral band B1. The denominator, on the other hand, comprises the sum of emissions between 380 and
nm, and therefore needs an additional band B2. This sum equals applying a rectangular detector bandpass with CW of
nm and FWHM of
nm. The optimal band for a super-Gaussian detector bandpass function will not, in practice, differ much from these values.
The estimated G, namely , first ignores the logarithmic form, in order to more accurately estimate the multiplication factor a.
In order to select the optimal band B2, uniformly distributed sampling is used for the CW between 420 and
nm and FWHM between 100 and
nm based on the discussions. As expected, the optimal band B2 closely resembles the mentioned rectangular with
nm (blue). For a factor
an MAE for
of 0.081 is obtained. By comparing this result to the mean 1.875 and the standard deviation 1.923 of the data, adding a band in the blue part of the spectrum proves to be beneficial. The recommended bands by [13
], occasionally criticized for its lack of a dedicated blue band, only reaches an MAE of
of 0.569 with its scotopic band
nm. With common criteria suggesting
], an error of 0.081 is acceptable in most cases. Note that these accuracy values additionally depend on the characteristics of the sensor, i.e., detection limits, saturation, and the number of bits used for radiometric sampling. Hence, the error will be slightly larger in reality, but remains acceptable for a sensible choice of sensor parameters.
With most of modern streetlights being non-Planckian radiators more emphasis is placed on the estimation of the spectral G index, as opposed to estimating correlated color temperature.
In order to classify different radiation sources into their respective type, sensor data is compared to a spectral library with the means of a k-nearest-neighbor (KNN) classification. Put simply, a particular radiation source is labeled with the same class as the majority of its k nearest neighbors in feature space. KNN is used as it is robust. The goal is not to find the best possible classifier, but the classification method serves as a comparison measure to judge the usefulness of a particular band combination. When applying KNN, adding features might deteriorate the classification performance, even if it improves the classification of one of the classes. For example, if adding a particular band improves the classification of one lighting type, it might generate large distances for other lighting types, because of large variances in this band. To overcome these effects, with normalized bands considered as features, namely its signal is divided by the signal of the panchromatic band, KNN is applied to each possible combination of features individually. For each lighting type the best feature combination is determined by withholding the combination with the highest classification performance. To combine the resulting binary one-versus-all classification results, a weighted voting is performed, with the classification performance used as weights. In other words, if a particular lamp is classified as multiple types, the one with the best performing classifier will be the deciding. For the case where a lamp is not classified as any lighting type, it is labeled as no class.
Based on Figure 4
, the spectral library includes two selected representative spectra for each of the eight considered lighting types combined with eight typical surface reflectances and normalized to luminance of
. Furthermore, four fires with temperatures equally distributed between 400 and
K are included. Finally, all these spectral radiances are combined with the five stated atmospheric profiles at 20 and
km visibility each.
For each lighting type, KNN searches for the best possible combination of bands, including bands B1 and B2 that have been fixed. In order to select the optimal additional bands, uniformly distributed sampling is used for the CW between 350 and nm and FWHM between 5 and nm.
As a reference, the classification performance of the case is given, where no additional bands are added, i.e., by only making use of bands B0 for normalization, B1, and B2. Here, a mean F1 score of
is reached. The four bands suggested by [13
] instead of B1 and B2, however, reached a mean F1 score of
. Thus, it is evident that improvements are possible, and required, by including more bands. Table 1
gives an overview of the F1 scores for individual radiation source classes, for optimal band selections of 0, 1, and 2 additional bands. An examination of the optimal combination of 3 additional bands proves that improvements are minimal with a mean F1 score of
for 2 additional bands. One additional band results in a mean F1 score of
. Hence, the addition of two bands is recommended in order to allow lighting type identification.
An optimum is reached for
nm (yellow-orange), in the case of one additional band B3. Corresponding mean F1 scores significantly increase to a similar mean F1 score than for the bands suggested by [13
], but with one band less. For example, the largest improvement is seen for mercury vapor, which is clearly differentiated from other lighting types.
Although it is expected that the best combination of two additional bands includes a band identical, or at least similar, to mentioned band B3, this is not necessarily the case. The reason for this is that some parts of the spectra might possess high correlations with other parts of the spectra. In other words, joining the two best scoring spectral bands does not necessarily result in a better classification performance if they contain similar, correlated, information. The determination of two additional bands, therefore, needs to start from the situation with B0, B1, and B2 fixed, and with detector bandpasses for B3 and B4 being generated. In the case of two additional bands B3 and B4, an optimum is reached for nm and nm (yellow-orange) as well as nm and nm (near infrared). B3 does not, as is expected, differ much from the one optimal in the case of one additional band. The largest improvements are seen for HPS and LED classes as well as only two classes generate F1 scores lower than , i.e., fire and fluorescent.
A closer look at the confusion matrix in Table 2
reveals the reasons for these low values and details which types of misclassifications are to be expected. For example, fire is sometimes wrongly classified as an incandescent lamp or as a mercury vapor lamp. The confusion with mercury vapor lamps is probably solved by adding a narrow band around
nm, where mercury vapor lamps have one of their emission peaks. However, the relatively low occurrences of both classes in urban areas, and the relatively small improvement of introducing such a band aiming solely to distinguish between these two types, do not justify its consideration. Likewise, there is a high correlation between fluorescent and mercury vapor lamps. The reason for this lies in the manufacturing process of a fluorescent lamp. Similarly to mercury vapor lamps, fluorescent lamps make use of mercury gas, resulting in nearly identical emission spectra. Adding a narrow band around
nm probably solves this issue. However, again its consideration is not justified.
3.4. Correlated Color Temperature
In order to assess the perceived color of the light emitted by a particular lamp, its spectrum is compared to a range of blackbody radiators, which follow Planck’s law. The absolute temperature of the blackbody that most closely resembles the spectrum of the lamp, defines the so-called correlated color temperature (CCT) [18
]. It needs to be noted that, while the computation of CCT values is relevant for lamps that closely resemble the spectrum of a Planckian source, e.g., in the case of incandescent lamps, it is no longer relevant for other lighting technologies such as gas discharge or LED lamps. Despite its limited ability to describe a lamp spectrum, CCT remains a widely applied indicator, as it is relatively straightforward to grasp its meaning.
Another frequently cited parameter to describe a light source spectrum is that of the color rendering index (CRI), which expresses a lamp’s ability to faithfully reproduce different colors along the spectrum, compared to a blackbody radiator with the same CCT. Typically, incandescent lamps have high CRI values close to the maximum value of 100. LPS lamps, on the other hand, have only one narrow peak in its spectrum and, therefore, yield low CRI values, near 0. As estimating CRI requires a very high spectral resolution its estimation is not considered, here.
Although it will most likely lose its value as a lighting metric, as it is limited in properly describing a lamp’s characteristics, CCT remains a valuable and frequently mentioned specification. For a proper estimation of CCT, a good distribution of spectral bands along the visible spectrum is required. Looking at the bands that are already fixed for LER, G, and classification of lighting types, B1 and B2 seem to be good candidates to cover the green and blue part of the spectrum. The part of the spectrum that is not covered by the existing bands is located in the red part of the spectrum. It is, therefore, expected that adding a single band in that wavelength range will significantly decrease the estimation error of CCT.
As expected, the optimum is reached for a band that covers the red part of the spectrum, namely
nm (red). By adding this band, MAE for the estimated CCT, namely
which is derived based on the estimated tristimulus values
], is significantly improved from
K. An additional advantage of including B5 is that it offers the possibility of generating true color imagery, with B2, B1, and B5 corresponding to the blue, green, and red band.
3.5. Performance Analyses
The recommended spectral bands are illustrated in combination with some typical lamp spectra in Figure 5
without the panchromatic band with 374–864 nm. What is immediately seen is the ability of bands B3 and B4 to distinguish between different lighting types. Additionally, there is a good spread of the different bands across the VIS/NIR spectrum, except for the wavelengths 650–800 nm. This unsurprisingly coincides exactly with that part of the spectrum where lamps typically emit no light. Due to the nature of nocturnal radiation sources, the choice of spectral bands differs significantly from the typical daytime optical sensors. This nighttime focus results in rather atypical bands, e.g., the narrow yellow-orange band B3 around
A performance comparison for the selected bands with respect to other available band combinations, i.e., the proposal for Nightsat,
m bands of Sentinel-2, AC-5, and JL1-3B is given in Table 3
, where the bands are illustrated in Figure 6
. As the characteristics of the green band of JL1-3B and photographs taken by astronauts aboard the ISS are the same, the estimated LER are also the same. The blue bands are shifted by
nm only and therefore, the estimated spectral G indices are also similar. As the table reveals, a performance improvement was achieved for all relevant indices with the optimized bands. The Nightsat mission proposal, which is the standard reference with respect to nighttime VIS/NIR missions, does score relatively well in certain aspects. For example, the classification of radiation source types reaches similar results as the proposal with three multispectral bands. However, it does not succeed in estimating emissions in the blue part of the spectrum, as is reflected by the large MAE for the spectral G index.
3.6. Radiometric Resolution
As the conversion from sensor signal to a digital number as illustrated in Figure 4
is not considered, unlimited dynamic ranges and quantization are assumed. Detectors, however, have a detection limit, saturation, and bit depth. Assuming detectors have a linear response, the conversion for band B from an incoming band-averaged spectral radiance
is computed by
represent the detection limit and the saturation of the sensor as well as
is the maximum digital number that can be attained, e.g., 255 for 8-bit images.
Recommendations for the detection limits and saturation of multispectral nighttime VIS/NIR sensors have been made by [13
]. For a photopic spectral band with 510–610 nm detection limit and saturation of
are recommended. Here, band-averaged spectral radiances for the corresponding band B1 range between
for the lamps. Note that the lower limit considers land environments and challenges with the limited reflected light in aquatic environments are not covered. Especially the upper limit is significantly lower than the saturation recommended by [13
], where the Luxor Sky Beam in Las Vegas, NV, USA, is used as reference. With a 42 billion candela tunnel of light, it is the strongest light beam in the world. Taking a linear response into account and ignoring less than 2% of all TOA radiances, for band B1 detection limit and saturation of
are recommended in the case of 16-bit as well as
in the case fewer bits are available. The detection limit recommended by [13
], on the other hand, more or less conforms to the computed values, here. However, setting the detection limit to at most
extends the area of operation to the lighting of pedestrian and cycle zones.
For the other bands, the distribution of band-averaged spectral radiances follow that of band B1 for the lamps, with marginally lower values for the narrow band B3 and with the exception of two bands, i.e., the blue band B2 and the near infrared band B4. The rather low TOA radiances of band B2 between and W msrnm is explained by the fact that some lamps, e.g., high- and low-pressure sodium lamps, barely emit blue light. Ignoring less than 12% of the smallest TOA radiances and with the importance of blue light emissions in mind, even for relatively low TOA radiances, it is recommended that the blue band has a slightly lower detection limit by a factor . Another difference is seen in band B4, where again some lamps, e.g., low-pressure sodium, fluorescent, and LED lamps, barely emit near infrared light, but also rather high TOA radiances of W msrnm are computed, belonging to some of the high-pressure sodium lamps. Therefore, by ignoring less than 33% of the smallest and 2% of the largest TOA radiances, it is recommended that this band has a higher dynamic range, with a saturation in the case of 16-bit at and W msrnm in the case fewer bits are available. Increasing the saturation of band B4 is not only useful for high-pressure sodium lamps, but additionally increases the detection rate of fire.
Although the main focus is on urban areas, the detection of fire is an interesting by-product of a dedicated nighttime VIS/NIR sensor. Performing a similar analysis for fire spectra with temperatures between 400 and K illustrates that the highest TOA band-averaged radiances are not surprisingly located in band B4, in the near infrared part of the spectrum. However, even this band is not able to detect all fires, given the detection limit and saturation recommended for lamps. For a standard atmosphere a detection limit of W msrnm for B4 roughly corresponds to fires of K, while a saturation at W msrnm for B4 roughly correspond to a temperatures of K. However, the saturation is less of an issue, as TOA radiance values are lower in the panchromatic band B0, for example, thereby not exceeding the saturation threshold. As a consequence, most fires with temperatures exceeding K are detectable by the recommended sensor. These temperatures cover most of the forest fires, meaning that the proposed spectral bands, with their detection limits and saturation, serve as an additional tool for fire detection programs. Although the VIS/NIR part of the spectrum does not cover the radiation peak of fires, as given by Wien’s displacement law, there is an important difference with respect to daytime optical sensors. The lower detection limits that are required for nighttime VIS/NIR sensors offer an opportunity to detect the lower TOA radiances that are emitted by fires in the VIS/NIR region, typically not visible to daytime VIS/NIR sensors.
For a standard atmosphere and a constant albedo of 20%, typical for road surfaces, TOA band-averaged spectral radiances for full moon conditions range between for band B2 and W msrnm for band B5. As these values exceed the recommended detection limit, it is necessary to model out moonlight in most cases. With typical albedo values for snow and clouds around 95 and 70%, their respective TOA radiances range between and W msrnm. Thus, under full moon conditions, it is possible to detect both clouds and snow cover. As both effects are extended in size; a spatial binning results in detection capabilities even with reduced moonlight. However, in comparison to, e.g., NPP-VIIRS-DNB its ability to detect such phenomena is limited.
The computed performance metrics for LER, spectral G index, classification, and CCT are based on sensor signals before being converted into DN. This means that the results are slightly worse in a realistic setup, since certain small TOA radiance differences will be lost as a result of radiometric sampling. With the above-mentioned spectral bands and their recommended detection limits and saturation levels, additional analyses are carried out for different bit depths in Table 4
For most bit depths, classification results are more or less stable, with the exception of an 8-bit conversion, which produces significantly deteriorated mean F1 scores. While the conversion to 10-bit still succeeds at classifying most of the radiation sources, the ability to estimate the luminous efficacy of radiation and spectral G index has drastically declined with respect to larger bit depths, with its values unacceptable for proper use. It is, therefore, recommended to apply a radiometric sampling of at least 12-bit, with higher bit depths not considerably better.
For each of the bands, typical TOA band-averaged spectral radiance values are calculated based on the selected lamp spectra, surface types, luminance recommendations, and atmospheric conditions. It is important to note here that some of these parameters are uniformly distributed between a minimum and a maximum value. Therefore, rather than covering a realistic distribution of values, it reflects a range of possibilities that is evenly distributed.
3.7. Spatial Resolution
The light emitted by artificial lighting sources does not only vary with wavelength, but also depends on the direction in which the light is emitted. The spatial resolution that is required for a VIS/NIR nighttime sensor depends completely on the objective of such a mission. It depends especially on the scale of the objects that need to be detected. For example, if the focus of a mission is on single-lamp level, a different spatial resolution is required, compared to city block level such as NPP-VIIRS-DNB. However, for such spatial resolutions, the multispectral approach makes little sense, since the signal that arrives at the sensor consists of a multitude of lamp signals and lamp types, turning the estimation of LER, spectral G index, and radiation type meaningless. Therefore, the focus will be on the single-lamp level here. It is sufficient to consider the panchromatic band only as reducing the spectral resolution as such is not likely to change the detectability of lamps.
To arrive at recommendations concerning the spatial resolution of a nighttime VIS/NIR sensor which focuses on artificial lighting, the spacing between different lamps, in combination with their mounting height, plays an important role. Typical values for these variables were derived from lighting engineering standards [21
]. This lead to three different cases, i.e., a spacing of
m and mounting heights of
m for residential roads (Figure 7
a); a spacing of
m and mounting heights of
m for roads with a mixed function (Figure 7
b); and a spacing of
m and mounting heights of
m for major roads (Figure 7
c). Lower spacing distances than
m do occur, but are not frequent. A typical luminous intensity distribution pattern is considered, which shows the intensity of emitted light for different directions.
According to the Nyquist sampling theorem, the sampling frequency shall be at least twice the highest frequency contained in a signal. Applying this logic here means that the required ground sampling distance should equal half of the spacing, or less, between neighboring lamps. With a minimum spacing of
m, this results in
m or less. Moreover, neighboring lamps usually possess similar characteristics, which makes the detection of individual lamps not necessarily required in all cases and a reduced ground sampling distance for multispectral bands than for the panchromatic band are feasible. For each of these cases, some sensible spatial resolution options are investigated in Figure 7
. Notwithstanding the above-mentioned prediction, road lighting does not behave like a regular point source. Intensity distributions have a major influence on the positioning estimation of lamps. The results predicted by the Nyquist theorem, however, are confirmed by visual inspection that a spatial resolution of
m is feasible. As a target of lighting engineering is to create more homogeneous illumination patterns on the surface and the atmosphere blurs the shape, a spatial resolution less than
m is in this case required to distinguish lamps and between public and private lighting.
For larger-sized roads, such as dual-carriageways, not only the most common one-sided arrangement and single central arrangement occur that are sufficiently covered by the investigated single-row arrangement, but also a twin central arrangement (Figure 7
d), a two-sided opposite arrangement (Figure 7
e), and a two-sided staggered arrangement (Figure 7
f) are common. Lamp spacings of
m are considered as a spacing of
m is not relevant, as it corresponds to relatively narrow roads in a residential area, for which a single-row arrangement is the preferred option.
A similar pattern as for single central arrangement with double the number of lamps is achieved for the twin central arrangement, as neighboring twin lamps are that close to each other that they are almost identical to single lamps. However, under certain conditions a thorough pattern analysis at a minimum spacing of at most half the distance between the centers of the light cones of the twin lamps allows the two lamps to be differentiated. Furthermore, as the light cones face each other, the task gets more complicated. For the two-sided opposite arrangement a similar situation occurs, the light cones are oppositely directed, which allows for a slightly coarser spatial resolution still enabling the discrimination of the lamps. Moreover, since the lamps exhibit identical characteristics, it is acceptable to classify them as a single lamp. It is considerably easier to detect single lamps for a two-sided staggered arrangement. Once more, as all lamps have a distance of more than m to each other due to the carriage width, a comparable situation as for the single-row arrangement with a spacing of m is present.
Given the recommended spatial resolution of
m and the relatively low detection limits, it is also possible to combine two panchromatic bands. The first band combines a high spatial resolution of
m with relaxed detection limits, thereby only focusing on lamp detection, while the second band combines lower detection limits with a relaxed spatial resolution, e.g., 20–25 m. This is also sufficient for the multispectral bands, for which a reduced spatial resolution of 40–50 m is still acceptable as lighting parameters typically do not change from lamp to lamp, but potentially from street to street. However, in this case a stronger mixture of the lamps spectra with spectra of residential and industrial lighting as well as vehicle lights have to be considered also in combination with skyglow [29
With a focus on spectral characteristics, but also considering radiometric and spatial resolutions, the article performed and analyzed simulations to recommend performance parameters for nocturnal multispectral satellite imagery for urban areas as summarized in Table 5
. The simulations accounted for all major contributions to the signal, namely typical theoretical fire spectra, lamp spectral libraries, standard luminance values for road surfaces, a surface reflectance library, estimations of atmospheric effects, and the sensor. Future research shall generate and consider fire spectral libraries in visible and near infrared (VIS/NIR) to thermal infrared to exploit the capabilities of enhanced nighttime satellite imagery for users of fire products. For urban areas the most important lighting parameters, namely luminous efficacy of radiation (LER) (with radiant flux and luminous flux), spectral G index (G), classification of lighting types (fire, incandescent lamps, high-pressure sodium lamps, low-pressure sodium lamps, mercury vapor lamps, metal halide lamps, fluorescent lamps, warm-white LED lamps, cold-white LED lamps) (Classif.), and correlated color temperature (CCT) are considered.
Reference radiances represent the mean of all considered TOA radiances and the corresponding Signal-Noise-Radio is derived according to Section 2.6
, whereas all other parameters are already directly derived in Section 3
The next step in this process of simulation is to generate more complex imagery based on measured lamps, measured reflectance spectra, digital surface models, cadastral maps, modeled cloud conditions and adding factors such as the moon, residential and industrial lighting as well as vehicle lights. Therefore, further ground, air-, and spaceborne measurements need to be integrated. Such more realistic data allow investigating how mixtures of different radiation sources and spatial resolutions influence, e.g., classification results and radiant flux estimations considering cloud parameters. Finally, the accuracies of the models and the parameters are to be covered more detailed which is of major importance for applications, too.
Nighttime images with high spectral and high spatial resolutions are a relatively unexplored field. The options for future research are, therefore, plentiful. One such area of interest is the estimation of radiant flux also considering inter- and intra-night changes of emissions, which enables recognition of changes in human activities, where research has already scratched the surface of this topic. Time-series of imagery at different scales are investigated to cover the dynamics of the urban lightscape [30
]. Another such area of interest is the estimation of cutoff of lamps, i.e., non-cutoff, semi-cutoff, cutoff, or full-cutoff, to rate the amount of wasted light and glare which is a major lighting parameter not covered, here. Imagery acquired under different tilting angles are investigated.
With the mentioned areas of research only touching a part of the countless opportunities, simulating nocturnal imagery is a major step towards a deeper understanding of nighttime VIS/NIR remote sensing (missions) to reveal insights in the human activities shaping and changing the Earth.