4.1. Linear Regression Coefficients as Indicators of Radiance Angular Persistency and Angular Minimum Detection Threshold
Figure 10 displays the slope coefficients derived from the linear regression models of near-nadir and off-nadir observations for each LUC class. When the slope coefficient equals one, both angular observations tend to have the same average radiance values. The closer to zero, the higher the off-nadir average radiances in comparison to near-nadir average radiances. The greater the number, the higher the near-nadir average radiances in comparison to off-nadir. Slope coefficients are ranked from lowest to highest, indicating the LUC’s degree of influence over the angular effect in areas where stable lights are found.
Regarding areas usually expressed by open spaces and lacking the frequent presence of vertical features (Soybean, Savanna formation, Grassland, and Pasture), their β1 coefficients agree with their general anisotropic characteristics. Considering the different anisotropic characteristics of each LUC class, it is possible to interpret the slope coefficient behavior by analyzing the reflectance from their respective LUC classes at specific angles. One should also take into account that the DNB sensor has a wide range of detection (0.4–0.9 µm), which requires a holistic analysis of the different bands within this interval. The most straightforward method to interpret the slope coefficients’ behavior among different LUC classes would be to consider the relation between the view zenith angle (VZA) and reflectance factor (RF). However, as the RF is an empirically derived variable that is not constrained to a single positional parameter (i.e., VZA), a few considerations should precede the analysis of the β1 coefficients.
Daily DNB observations inputted into the composition of the VNP46A4 series are already subjected to a procedure to compensate the lunar irradiance and the resulting airglow [
29]. Therefore, the following considerations take into account only the possibility of interaction between light sources on the ground, mainly lighting poles, and the surface itself. In the LUC context, VIIRS/DNB pixels were classified according to their predominantly LUC class, which means that only pixels consisting of urban areas and mining sites are likeable to display light sources that are arranged at lower zenith angles of irradiance. That means that, aside from urban areas and mining sites, all stably lit pixels are likely to be a result of oblique interactions between ground light sources and the surface itself, either because they are just partially present under a pixel’s footprint or due to the predominance of overglow in these areas.
Regarding the azimuth angles, although VIIRS observations do vary from approximately −90° to 120°, VNP46A4 series’ view azimuth angles (VAAs) are not sampled and reported based on specific angular ranges as with VZAs, probably due to a considerable difference between the available range of viewing azimuthal angles at different latitudes. Therefore, although these parameters are crucial to the interpretation of angular effects, the following discussion is based on the premise that both azimuthal angles of irradiance and VAAs are considered random due to the setting of this experiment in different stages of the analysis (composition of the VNP46A4 and sampling process prior to regressions); VZAs are determined by the type of composite (off-nadir or near-nadir); and irradiance zenith angles are generally oblique (>30°). The roles of back- and forward scattering are also unpredictable in this context, but, given the same established premises, a rough averaging should provide insights on how the β1 coefficient relates to the anisotropic characteristics of the different LUC classes. Under a sun zenith angle of 30°, soybean fields reach their peak of reflectance at a VZA of 30° on the principal plane, but decay rapidly when this angle is increased, shifting from approximately 80% to less than 10% higher at 45° if compared to their reflectance at 20°. Although this observation was limited to the red band (661 nm) and to an interval associated with backscattering (0–45°), the result is coherent with β1 = 0.95, a very small difference between off- and near-nadir radiance levels, considering that its peak reflectance is not contemplated by the interval of the angular composites [
45]. The remaining classes of this sort tend to have a smaller β1.
Grasslands have been documented to have a reflectance at 60° approx. 1.5 times higher than the reflectance at the 10–20° interval (i.e., 35% higher), therefore favoring the off-nadir radiance levels and a smaller β1 than soybean fields [
46]. Leaf area index (LAI) and the chlorophyll content also play a role in the bidirectional reflectance, but variations in reflectance values seem to affect both off- and near-nadir observations proportionally [
45,
46,
47,
48], which should have a much lesser impact on the β1 coefficients independently of changes in these variables.
Wetlands are subjected to a similar logic, although water seems to be the main driver of the angular specificities. The reflectance associated with this class in the panchromatic interval (0.51–0.73 nm) has a reflectance factor that does not differ much when considering different moist content, but if considered from different points of view, wetlands’ reflectance can vary from less than 0.05 (0–10°) to almost 0.5 (45° forwards), i.e., 10 times higher [
49]. In the BLA, wetlands are usually covered by non-forest formations, depending on which biome they occur in. The phytophysiognomy can vary from lowland species in the Amazon Forest to herbaceous and savanna arboreal species in Cerrado and shrubs and herbaceous species in Pantanal. All of them are partially or fully flooded over the year, share a significant presence of grassland, and are also found in the Brazilian Legal Amazon’s territory [
21]. In this case, the relative prominent off-nadir radiance levels seem to be not a result of the heavy presence of obstructing features, but due to the anisotropy effect of this class.
Lit areas in the context of forest formations encompass both primary and secondary forests. Either way, as they are found in stably lit areas, some level of deforestation is expected. Higher deforestation levels lead to stronger angular effects, which result in differences in reflectance levels of deforested areas, varying from 35% to 75% higher at off-nadir angles, depending on the spectral interval (75% for visible, 35% for NIR, 45% for SWIR) [
47]. Since the proportion of differences between view angles is smaller than the ones found in wetlands, it is safe to assume that, although forest formations have a virtually equal β1, they are not subjected to the same drivers affecting wetlands. In this case, it is likely that near-nadir angles are the most affected by the forest canopy.
The ordering suggests that, although the β1 is higher in areas with lesser obstructing features (Soybean, Savanna formation, Grassland, and Pasture), the obstruction is not necessarily the only factor that interferes with the β1 levels. Classes that lie at both extremes of the ranking (Mining, Urban Area, and Mosaic Agriculture and Pasture) present more diverse settings of structures that can interact with the emitted radiance, which makes it difficult to trace all drivers of the angular effect without a further inspection of their characteristics.
In this particular experiment, urban classes are based on MapBiomas Version 6 data [
21]. These areas are mapped on a 30 m scale, meaning samples must represent the most frequent class under the VIIRS/DNB pixel’s footprint, either of urban areas or any other class. Samples with a lesser proportion of urban structures over their landscape are necessarily classified in another LUC context. Peripheral urban environments and small isolated settlements are subjected to a higher proportion of radiance emerging from overglowing areas, and a higher discrepancy between off- and near-nadir radiance levels is also expected. In the case of samples that are classified as urban areas, the proportion of urban environments with a variety of physical expressions results in well-balanced sampled average radiance values between both angular compositions, only because the region lacks highly verticalized areas.
Relational metrics can also provide further insights on this matter. Positive values of DIF relational metrics indicate that the radiance levels from mining sites also present an off-nadir predominance over near-nadir (
Figure 7 and
Figure 10). Specifically on extraction sites, the observation is just the opposite: near-nadir values tend to be higher (
Figure 7). In this sense, the β1 coefficient could be explained by a disproportionate sampling of supporting facilities to the detriment of extraction sites. However, extraction areas are proportionally more extensive. Consequentially, it is likely that supporting facilities have lighting structures that overlap the radiance levels found at extraction sites. These specific areas are not persistently illuminated, i.e., they usually lack the fixed lighting structures that are found on supporting facilities. As mining activity advances, extraction sites are subject to constant infrastructural changes to meet the current needs of the extraction activity, which include vehicle lights and other mobile light sources [
50]. Although this interpretation is able to provide an empiric insight on how mining areas have such particular low β1 levels when compared with urban areas, the reason why oblique observations tend to register such relatively higher radiance levels is still unclear.
The interpretation of β0 (intercept value) can be considered not only if it is an absolute value, but also if it is negative or positive. In most cases, intercept values are negative, indicating that if a certain spot emits or reflects light, it will only be detected by a near-nadir observation if the off-nadir radiance exceeds the absolute value of β0. Mining areas and wetlands are exceptions, for they present positive intercept values (see
Figure 4b,c). In this case, it is hard to set a scenario where night lights above a certain threshold are detected firstly near-nadir and only after it exceeds this limit, it will present an anisotropic behavior that favors the off-nadir viewing angle.
Mining areas comprise both artisanal mining spots (garimpos) and industrial mining sites, which are very different in terms of infrastructure [
21]. The potential variety of lighting levels and settings present at those sites might be the only plausible explanation aside from atmospheric-related processes and very specific anisotropic characteristics that could lead the class to present intercept values above zero and slope coefficients less than one. Wetlands, in turn, are more dynamic environments that are not defined by their inland cover, but instead, by the frequency that they stay dry versus covered by water during a year [
21]. Therefore, even though this class can be associated with areas of Forest Formation, Savanna Formation, Grassland Formation, and Pasture, it is expected to have relatively limited settings of occupation in these areas, if not represented by overglow areas in its totality.
The problem with the interpretation of the equation’s intercept coefficient is that, in most cases, its value is lower than the minimum estimated detection limit of the DNB sensor, putting in check the association of the predominant LUC classes as main drivers for the presence of exclusive pixels (see
Figure 4c–i). Only classes that are more likely to present lighting infrastructure under the footprint of all their pixels have intercepts greater than 0.5 (mining and urban areas—
Figure 4a,b). As a result, while isolated exclusive near-nadir pixels still hold LUC and spatial distribution patterns that are worthy of investigation, most exclusive off-nadir pixels are probably a result of the higher sensibility of oblique viewing angles to the overglow effect. The occurrence of higher proportions of certain LUC classes on these exclusive off-nadir pixels, such as Forest Formation and Pasture, might be simply by chance since they are predominant in the study area and the estimated intercept coefficient is lower than 0.5 units.
4.2. Association between Man-Made Typologies and Relational Metrics
It is not reasonable to assume that every type of settlement or facility can be represented by the metrics derived from the relation between off-nadir and near-nadir radiance, but a couple of different types of structures should be more easily distinguishable because of their unique spatial features, types of lighting infrastructure, and landscape patterns in general. Considering the angular effect behavior reported in [
30,
31], areas with a rather high density of tall buildings should present a higher radiance towards the nadir, since off-nadir angle views might be blocked by buildings’ facades. Residential areas, on the other hand, will not be able to block the off-nadir radiance as much, allowing a clear differentiation between highly verticalized areas and any other urban typology.
If the slope were interpreted by itself, one could conclude that there is no preferential direction of propagation of anthropogenic lights in urban environments, which has been proven wrong multiple times [
30,
31,
32,
51]. Quite the opposite, it is likely that urban environments are complex enough across the BLA that the off-nadir and near-nadir preferential directions of propagation of the light can cancel each other out. This is an important feature of the relational metrics, for it shows that they can be used to set a baseline for analyzing a city’s structure by comparing the slope coefficients estimated on scales different from the BLA, such as city-wise. This possibility would only be valid if there was a clear relationship between different urban typologies and their relational metrics.
In the BLA, it is not as common to find cities with the same degree of verticalization as in other cities from different regions of the country and the world [
52,
53]. Even in some state capitals, the presence of tall buildings is rather scarce. This limits experimentation in certain ways, but it is still possible to define a range of urban typologies based on the presence of tall buildings. The structures of cities are also very different among capitals and smaller cities, for a large parcel of the BLA territory still has a very precarious urban system, lacking appropriate lighting systems and overall basic maintenance [
7,
8,
9].
In many cases, the inter-city transportation system is often dependent on inland waterways, which, without a proper navigation system, also results in a lack of means to provide the infrastructure that befits the local needs [
7,
8,
54]. Although there are many further causes of the current mismatch in resource provision in the BLA, often political, the following discussion is limited to the physical aspects that might be accountable for the angular effect.
In urbanized areas, there are a higher number of factors affecting the radiance values and the way radiation interacts with these surfaces. Even so, it is possible to reduce these factors to building height, vegetation, and building area (footprint) [
31]. These components might not be exactly as important in some human settlements of the BLA, for they manifest themselves in patterns that have not yet been studied from the perspective of nighttime lights, especially in small settlements far away from urban centers. Even though there is a significant overlap in the distribution of the relation metrics, their central tendency can most definitely be used as a parameter for classifying urban typologies in the BLA. However, these results are limited to a few representative samples. Therefore, relational metrics can be included as an extra dimension of any classification method, which would also benefit from other data derived from remote sensing images, such as the average size of building footprints, street and lot size, and variables that would help to differentiate wealthy areas and industrial sites as well [
9,
55].
A time series of these metrics would also help to understand how the urban class of a certain settlement is evolving. For instance, consider the variance of RES and REL in areas with an increasingly higher variety of RES and REL over the years. This could indicate that the area is developing into a more complex environment. At the same time, the analysis of the central tendency of the relational metrics in the same site over time would provide further information to decide whether it is a type of development that favors off-nadir radiance levels or near-nadir levels.
Concerning the relational metrics from mining spot samples illustrated in
Section 3.2, the behavior of these same metrics might be triggered by different processes. Given that variation in radiance values and the magnitude of the angular effect are both positively correlated with the radiance level [
31,
56], the narrower distributions of RES and DIF are likely due to low levels of radiance detected in extraction sites in comparison to supporting facilities. The higher scattering of the REL from these surfaces supports this observation, for it means that the radiance values found in areas for extracting minerals can present a less stable radiance value in between the sampled areas.
In this sense, the combination of relational metrics could benefit different purposes of classification. Considering a single mining complex, the DIF metric has the potential to further classify mining sites based on areas used for administration and support, and areas used for the extraction of minerals. Since the relative difference has a higher variance among the relational metrics, other experiments should consider its use to differentiate at least the method used for extracting minerals. If considering that the BRDF of different types of bare soils is also specific [
45], one should expect a different behavior of the relational metrics from different types of mining activities, specifically from the REL metric.
4.3. Suitability of the Annual Composites in Comparison to Daily DNB Images
Although a recent letter published by Kyba et al. (2022) has discussed some applications that might benefit from the use of angular observations of nighttime lights [
51], the angular effect is, so far, generally treated as a noisy aspect of nighttime observations [
31,
32]. In studies aiming to detect significant changes in the time series of radiance levels at night, relying on daily images is usually the best option since they are sampled on a frequency that matches the studied issue. In these cases, the angular effect should be considered a component of variation in radiance levels, such as in power blackout detection, monitoring the recovery of compromised structures affected by natural disasters, vehicle traffic at night, light pollution, etc. [
57,
58,
59,
60].
Daily images are, however, susceptible to a range of issues that do not play a major role in annual composites. In the urban context, for example, human activities that take place at night are usually paced by local nighttime habits, which may vary both in space and time [
56,
61]. In an experiment performed in two neighboring cities of considerable different sizes in northern Spain, Bará et al. observed that lights coming from residential areas and vehicles can account for up to approximately 10% of light emissions at 1:00 am local time, but they decrease steadily until they reach a minimum close to 0% at 4:00 a.m. in both cases [
61]. Similar results were found by Li et al. [
62]. Since this roughly corresponds to the timespan of the SNPP/VIIRS overpass, variations are expected daily, simply due to the satellite’s overpass time.
When considering pinned locations, fluctuations in nighttime radiance levels have been shown to be high enough to make analyses based on single pixels problematic. Moreover, variations of nearby pixels are often correlated, which indicates that conclusions drawn from methods that do not analyze pixels based on their statistical terms might be more susceptible to outliers and transient sources of light that are hardly accountable. In this context, it is fair to assume that studies focused on images with coarser spatial resolution will manifest a more stable radiance over time since differences are eventually canceled by each other once they are grouped by some spatial criteria [
56].
At the city scale, intra-annual patterns of light emissions can be correlated to a diverse range of sociodemographic profiles and cultural patterns that, ultimately, make use of lighting differently throughout the year [
63]. Therefore, spatially aggregating pixels in hopes of minimizing the fluctuation of radiance levels will only be a good decision if one is not interested in investigating intracity radiance fluctuations, or any other spatial reference that does not match the spatial distribution of the object of study. Even so, nighttime daily images will also be affected by trends resulting from different human routines across the week and their cultural patterns across the year. Basically, if one is interested in studying lighting systems at a daily scale, such as power outage detection, energy consumption, or natural disaster monitoring, the ideal scenario would consist of modeling the human activities for one specific study area with common settings and also considering their behavior on a weekly and annual basis.
By reducing daily images to annual composites of nighttime lights, unpredictable sources of variation in radiance levels can be coerced to a minimum, so the differences in angular composites can better represent the “true” radiance detected at different angles. Of course, this represents a limitation for studies in which a higher frequency of data is desirable, but it may also provide a better baseline for detecting spatial patterns of human settlements based on the expected behavior of the angular effect.
Recent studies have shown that at least three main types of angular effects are present in urban environments: positive, negative, and U-shaped. These denominations illustrate the graphical scattering of radiance values of nighttime lights when a fixed location is sensed at progressively higher VZAs. A positive or negative effect reflects the monotonic behavior of the radiance as VZA varies and is closely influenced by the different types of buildings on the sensor’s line of sight. If positive, it means that the radiance level increases as the VZA also increases. This behavior is usually found in residential areas with few obstructing features. By contrast, a negative angular effect is often found in areas with a higher level of verticalization, such as commercial areas with taller buildings. The U-shaped angular effect is interpreted as areas where there is a mixture between the positive and negative angular effects, usually associated with buildings with larger footprints, such as in industrial areas [
31,
32].
In the annual composites, the angular effect cannot be exactly reproduced, for the process of composition relies on the averaging of the radiance values in only two different VZA ranges: near-nadir (0–20 degrees) and off-nadir (40–60 degrees) [
29]. Hence, assessing parameters of urban infrastructure and morphology from annual angular composites requires a different approach, such as using relation metrics (
Section 3.2.1). On top of that, cities that were considered “representatives” in other studies might not represent the typology of human settlements in the BLA. This reinforces the idea that methods relying on nighttime images, angular or not, should consider a more regional approach and also focus on how scale affects the behavior of relational metrics and similar parameters.