A NEW APPROACH TO IDENTIFY ON-GROUND LAMP TYPES FROM NIGHT-TIME ISS IMAGES

Artificial night-time light (NTL), emitted by various on-ground human activities, becomes further intensive in many regions worldwide. Its adverse effects on humans’ and ecosystems’ health crucially depend on the light spectrum, making the remote discrimination between different lamps a highly important task. However, such studies remain extremely limited, and none of them exploits freely available satellite imagery. In the analysis, the possibility to remotely assess the relative contribution of different lamp types into outdoor lighting is tested. For this sake, the radiometrically calibrated ISS RGB image is used. Spatial resolution of the image is ~20 meters, implying that each pixel may represent a mixture of different lamp types. Unmixing analysis to the detailed spectral signatures of the corresponding in situ measurements is performed, with ‘pure’ lamps’ signatures as the endmembers. Afterwards, statistical models to reproduce the The paper as a non-peer reviewed EarthArXiv preprint 2 results of unmixing based on the broad-band RGB image from the ISS are run. The built models predict well (with R reaching ~0.87) the contribution of high-pressure sodium (HPS) and metalhalide (MH) lamps, the most spread in the study area (Haifa, Israel). The restored map for HPS allocation demonstrates high concordance with the network of municipal roads, while that for MH shows notable coincidence with the industrial facilities and the airport.

The adverse effects of NTL are known to depend crucially on the light spectrum (Brainard et al., 2001;Gaston et al., 2014;Papamichael et al., 2012;Schroer et al., 2016). For instance, the shortwavelength light stronger suppresses melatonin production and distort circadian rhythms in mammals (Haim and Portnov, 2013;Hatori et al., 2017;Lockley et al., 2003), while the long-wavelength light stronger disrupts the magnetic orientation of migratory birds (Wiltschko et al., 1993). As far as adverse effects of NTL become more recognized to depend on light spectrum, multispectral satellite imagery analysis becomes more numerous (Guk and Levin, 2020;Huang et al., 2021;Levin and Duke, 2012;Rybnikova and Portnov, 2017;Zheng et al., 2018). Such studies usually aim at revealing associations between light of certain spectra and land-use types or economic activities. The assumption (either explicit or implicit) behind these studies is that different land use types or economic activities use predominantly certain lighting source type. In the meantime, direct discriminating between different lamp types based on satellite imagery would allow a more fine-tuned analysis of adverse health effects associated with NTL. In turn, this would contribute to elaborating more precise policies for diminishing light pollution.
Such studies, aiming to directly link NTL spectra with onground lamp types, remain, however, extremely limited (Elvidge et al., 2010;Hale et al., 2013;Sánchez de Miguel et al., 2019;Zheng et al., 2018). In two of them (Elvidge et al., 2010;Sánchez de Miguel et al., 2019), the authors tested the principal possibility to identify lamp type, proceeding from corresponding spectral signature, either detailed or aggregated. In the first study, Elvidge with co-authors analysed spectral signatures of 43 different lamps representing nine the most widespread lamp types, using ASD spectroradiometer, implying measuring the signatures from 400 to 2500 nm with 10 nm band width (Elvidge et al., 2010). They showed that discriminant analysis correctly classified all lamp types when based on their detailed spectra. The authors also succeeded to find a minimum set of broad bands ensuring sufficient classification quality: Under blue, green, red, and NIR bands (a slightly modified set represented on the Landsat Thematic Mapper), only 4.7% of the lamp types were classified incorrectly. In the second study by de Miguel with co-authors, it was demonstrated that main lamp types can often be separated using color-color diagrams with G/R and B/G ratios as the two coordinates (Sánchez de Miguel et al., 2019). At that, the RGB bands corresponded to Nikon D3s camera sensors, used by astronauts in the ISS. However, in both mentioned studies, the proposed lamp-type discriminating methods, although being based on spectral bands of existing satellites, were not tested on real imagery (Elvidge et al., 2010;Sánchez de Miguel et al., 2019).
To the best of our knowledge, the only explicit tests of such kind were performed in (Hale et al., 2013;Zheng et al., 2018). Thus, Hale with co-authors analysed a one-meter aerial image of the Birmingham city, UK, and a layer reporting location and type of lamp (Hale et al., 2013). They succeeded to classify four main lamp types with high accuracy (7.5% error), based on three focal statistics: B and G/R ratio for pixels up to 1 m from the lamp centre and the maximum averaged RGB level for pixels between 2 and 4 m from the lamp centre. At that, RGB bands corresponded to those of Nikon D2X digital camera. Zheng with co-authors, in the meantime, used RGB highresolution (0.92 m) commercial satellite JL1-3B image to discriminate between HPS and LED lampstwo most widespread lighting sources in the study area, represented by Hangzhou, China (Zheng et al., 2018). The authors used RGB levels of brightly lit pixels of the image as input data for ISODATA clustering procedure; While unlabelled classes, generated by clustering algorithm, together with morphological characteristics of bright pixels, were used as inputs in decision tree classification procedure to discriminate HPS from LED lamps. The authors matched each group of bright pixels in the satellite image with lighting source type, obtained from the field survey and report that upon 446 available observations, overall accuracy of classification reaches 83.86%.
These studies, however, benefit from high-resolution aerial images available only for limited sites and are typically costly.
In the present study, we test for the possibility to identify on-ground lamp types from freely available satellite imagery of relatively coarse spatial and spectral resolution. For the study area of Haifa, Israel, we superimpose two NTL data sources: (i) the radiometrically calibrated broadband RGB image provided by the ISS, and (ii) a set of in situ measurements with detailed spectral signatures conducted by ourselves. Since the ISS image is of ~20-meter resolution, each pixel likely represents a mixture of lamps. Thus, the detailed spectral signatures of the in situ measurements are first subjected to unmixing analysis, with the standard lamp types used as endmembers,to estimate the relative contribution of different lamp types in each measurement.
Afterward, we use the levels of RGB bands of the corresponding pixels in the ISS image and develop statistical models to predict the relative contribution of different lamp type (output) from the aggregated RGB data (input). Finally, we apply the successful models to all pixels from the ISS image to restore the maps of relative contribution of certain lamp types into outdoor lighting in Haifa area.

Data Sources
The ISS-produced NTL image of Haifa (ISS045-E-148262) was taken on November 29, 2015 with the Nikon D4 DSLR camera ("Search Photos," n.d.). The image was georeferenced and radiometrically calibrated by SAVESTARS Consulting SL ("Home -Savestars Consulting S.L.," n.d.) according to the procedure reported in (Sánchez de Miguel, 2021). In situ NTL measurements were performed in March 2015 with the Konica Minolta CL-500A spectrometer. Each of the 610 measurements reports spectral irradiance (w/m 2 ) at 1-nm pitch from 360 to 780 nm ("Illuminance Spectrophotometer CL-500A," n.d.). Fig. 1 reports the original and the radiometrically calibrated RGB images from the ISS, overlaid with the in situ measurements localities. liquid-fuel and pressurized-fuel lamps, were omitted due to the data unavailability.

Methods
The methodological scheme of the study is reported in Fig. 2 and described in detail in subsections below.
The paper as a non-peer reviewed EarthArXiv preprint 7 Figure 2: Methodological scheme of the study

Selection of the representative in situ measurements for the ground truth
Proceeding from the available data on relatively coarse spatial resolution of ISS image, with each pixel reporting emissions from multiple light sources, and simultaneously given the sporadic pointwise available in situ NTL measurements, we selected among 610 observations only those which in some sense coincided with corresponding pixels in the ISS image. Since each pixel might be characterized by RGB radiances only, we first simulated the radiances of synthetic RGB bands of in situ measurements as if they would appear on the ISS sensors of Nikon D4 DSLR camera, and then chose the observations with similar (to the corresponding ISS imagery pixel) RGB characteristics.
To simulate the radiance R (of either red, green, or blue band), we used the augmented equation   As a measure of similarity between RGB radiances, reported by pixels of calibrated ISS image, and corresponding in situ measurements, we used Euclidian distance in the coordinate system, represented by B/G and G/R ratios. Given the variance of such a distance (dmax = 2.00), we settled the threshold of d<0.2.

Unmixing of the detailed spectral signatures of the representative in situ measurements
The detailed spectral signatures of the representative in situ measurements (see subsection 3.1) were subjected to unmixing analysis. As the endmembers (i.e., spectra of pure 'materials'see (Shi and Wang, 2014)), we used the detailed spectral signatures of the standard lamps from LICA-UCM library. The endmembers' signatures are shown in Fig. 4. For unmixing analysis, we used the FNNLS algorithm (Bro and De Jong, 1997), implemented in MATLAB v.R2020b. The algorithm returned the percentages of all endmember lamps in each of the pre-selected in situ measurements. The obtained percentages were then aggregated within lamp types, and the sums were normalized to unit.

Statistical Models to predict the relative contribution of different lamp types from the ISS image
The percentage of each lamp type, obtained in the unmixing analysis (Subsection 3.2), served as the dependent variable in a set of statistical models. As the explanatory variables, we tried different characteristics of the pixels of the calibrated ISS image: (i) radiance in the red, green and blue bands per se; (ii) their ratios (G/R and B/G, or GG/RB ratio), and an additional derivative characteristic describing the pixel's 'proximity' to the lamp type in question. This distance was included in the models since we found that different lamps within each of the lamp type tend to lie along straight lines in the G/R, B/G coordinate plane (see Fig. 5). We tried several formalizations of such a distance: (i) the ratio between B/G and G/R of the pixel, as a measure of the line's slope; (ii) Euclidean distance from the pixel to the line representing the lamp type in question, and (iii) Mahalanobis distance from the pixel to the cloud representing the lamp type in question, which accounts for both the centre of mass and the direction of the cloud (Mahalanobis, 1936). Figure 5: Simulated synthetic bands of the lamps: B/G vs. G/R ratios.
Note: LPS lamp, reporting G/R=0.18 and B/G=0.02, is the only representative of the type, and is not depicted in the figure We examined statistical models of several classes: linear regression, neural network, and random forest. All models were run in ORANGE v.3.28 with the default settings. Specifically, linear regression was applied with intercept and without regularization. As a neural network, we used a multi-layer perceptron with backpropagation; the model parameters were the following: number of neurons in hidden layer is 100, number of hidden layers is 1, activation function is ReLu, solver for weight optimization is stochastic gradient-based optimizer, L2 penalty parameter is 0.0001, maximal number of iterations is 200. In random forest models, the number of trees was settled to 10, arbitrary set of attributes and limit depth of individual trees were left unchecked, and subsets smaller than 5 were required not to be split. The whole set of observations was split into the training (90%) and the testing (10%) subsets; the models were run ten times, and the average values were assigned to each model as its performance score. The input database used in the analysis is available from the authors upon request.

Restoring the relative contributions of different lamp types into light emissions in Haifa
Finally, the best-performing models (Subsection 3.3) were applied to the radiometrically calibrated ISS image (see Subsection 3.1), and estimates for different lamps' contribution into light emissions from the study area were obtained and depicted.

Results
Among all initially available in situ measurements, we chose the set of representative measurementsthose deviating from the corresponding pixels of the radiometrically calibrated ISS image by less than 0.2 in terms of the Euclidian distance in the G/R, B/G coordinate plane (see    Table S1).
Since models with the predictor sets 2 nd-4 th demonstrate similar performance, and proceeding from consideration of calculation simplicity, we run random forest models with G/R, B/G, and GG/RB ratios as predictors upon all pixels of the calibrated ISS image (see Fig.1(b)). Fig. 7 reports resulting maps for two the most frequent lamp types' (HPS and MH) prevalence in Haifa. As can be seen from the figure, relatively higher contribution of HPS lamps in the outdoor lighting in Haifa coincide with the spatial pattern of municipal roads (see Fig. 7 (a)), while the pattern for MH lamps is more local and site-specific (see Fig. 7

Discussion
In the present study, we tested for the possibility to identify on-ground lamp types from freely available satellite imagery of relatively coarse spatial and spectral resolution. To this end, we conducted a series of in situ NTL measurements in Haifa, Israel, and combined these data with a radiometrically calibrated NTL image of the city taken from the ISS. Since the ISS image is of ~20-meter resolution, each pixel likely represents light emission from a mixture of lamps. By applying unmixing analysis to the detailed spectral signatures, we estimated the relative contributions of different lamp types in the in situ measurements. Then, we tried to train statistical models to predict these contributions based exclusively on the ISS image, which represents broad spectral RGB bands.

As our analysis indicates, contributions of two the most widespread lamp types in the region, HPS
and MH lamps, could be successfully predicted (with adjusted R 2 reaching 0.882 for the training and 0.848 for the testing sets) by random forest models. Using them, we restored HPS and MH lamps' contribution into outdoor lighting in all Haifa area. The obtained HPS map demonstrated high concordance with the network of municipal roads, while MH map showed notable coincidence with industrial facilities and airport (see Fig. 7).
In the developed models, we used three explanatory variables. Two of them, G/R and B/G ratios, are similar to those previously used by Sánchez de Miguel (2019). An additional informative predictor described a 'proximity' of the analysed pixel of the ISS image to the lamp type in question. Interestingly, it turned out that lamps of the same type form clear-cut line segmentshaped loci in the G/R, B/G coordinate plane (see Fig. 5). With this respect, we tried three alternative formalizations for the above-mentioned proximity: GG/RB ratio of the pixel, its Euclidean distance to the line, and, and its Mahalanobis distance to the locus. Without this additional explanatory variable, the models' performance was somewhat worse (with adjusted R 2 <0.861 for the training and R 2 <0.828 for the testing sets).
However, their results cannot be directly compared with those presented here since we solved regression rather than classification problem, implying continuous rather that binary dependent variable. Yet, it seems reasonable to expect better performance of our models if they were based on an image of better spectral (like in (Elvidge et al., 2010)) and/or spatial (like in (Hale et al., 2013;Zheng et al., 2018)) resolution. Again, as mentioned above, we used the ISS image intentionallygiven its free availability for many geographical sites. We think that our results argue for the principal possibility to assess the lamp type composition of outdoor lighting from the color satellite imagery.
Some limitations and perspectives of the study should be mentioned. First, proceeding from the available data, we did not succeed to obtain a reasonable spatial pattern for LEDs' contribution in the outdoor lighting in Haifa. A trivial reason may be insufficiency of the used data (small sample size, low prevalence of LEDs in the studied region in 2015). It also seems possible that the ISS imagery does not allow discriminating LEDs due to their high spectral variability and, therefore, overlapping with some other lamp types (which becomes even more pronounced after considering the reflectance of the ground), such as MH lamps, in the B/G, G/R space (Sánchez de Miguel et al., 2019). It should be mentioned that Elvidge et al. (2010) did succeed to discriminate LED lamps from other types since they used almost non-overlapping red, green, blue, and NIR bands.
Nowadays, LEDs' popularity grows rapidly (Alamús et al., 2017;Elvidge et al., 2010;Sánchez de Miguel et al., 2019;Schubert and Kim, 2005) mainly due to their versatility and energy saving potential, and some precedents of total LED-based street lighting already exist (Kyba et al., 2020;Sánchez de Miguel et al., 2019). At the same time, LEDs' primary emission peak, ~450-460 nm (Elvidge et al., 2010), extremely distorts circadian rhythms and suppresses melatonin production in humans, contributing to sleep disorders (Czeisler, 2013), obesity (McFadden et al., 2014), hormone-dependent cancers (Haim and Portnov, 2013), and other diseases. Thus, further analysis is needed to explore the possibility to identify LED lamps from the ISS-provided imagery. It seems promising to exploit, in addition to ISS imagery, VIIRS-provided data, which covers also NIR diapason on night-time light.
Second, acquisition time of the used ISS-provided image and of the in situ NTL measurements do not coincide perfectly, which may cause some inaccuracy in the herein obtained estimates.
However, weather conditions in Israel are rather similar in March (when the in situ measurements were conducted) and in November (when the ISS image was taken). Additionally, our sample did not include observations from residential areas, which brightness may vary during the night (Bará et al., 2019); instead, it included observations along major roads, entertainment areas, hospitals, and high-tech enterprisesthat is, represented by streetlights of nearly-constant brightness during the night.

Conclusions
Numerous medical and environmental studies have shown that night-time light of different spectra differs in its impact on human and ecosystem health. With this respect, the development of smart and precise policies for diminishing the adverse effects of light pollution requires fine-tuned analysis of multispectral satellite imagery which would enable remote recognition of different onground light sources. In the present study, we introduce a new approach to discriminate between lamp types proceeding from the freely available night-time ISS imagery and test it over the Haifa region, Israel. At the first stage, we applied unmixing analysis to the detailed spectral signatures of in situ measurementsto characterize each measurement by a set of percentages estimating the relative contributions of different lamp types to the overall light emission. Afterward, we matched the in situ measurements with the corresponding pixels in the ISS image and used the RGB characteristics (that is, G/R, B/G ratios, and the ratio between the two) of these pixels as explanatory variables in a set of machine learning techniquesto predict the earlier obtained percentages of different lamp types. For two lamp types, HPS and MH lamps, the predictions appeared fairly accurate. The restored maps of the relative contribution of these lamp types demonstrated high spatial concordance with specific on-ground activities, such as the network of municipal roads (for HPS lamps) and industrial facilities and airport (for MH lamps). These two lighting sources are the most widespread in the study area. The third popular lighting source, LED lamps, appeared hard to predict. A possible way to discriminate LED lamps may be combining ISS imagery with VIIRS data since the latter cover also NIR diapason. Such a fusion is a promising direction for future investigations.

Funding:
This work was supported by the Council for Higher Education of Israel and Cities at Night Project.