The OMI instrument aboard the Aura satellite launched in 2004 is a push-broom UV–visible imaging grating spectrometer. It has a daytime equatorial crossing at ∼13:42 LT (local time). During normal global observation mode, the backscattered sunlight from the Earth is imaged by a telescope onto a rectangular entrance slit perpendicular to the flight direction. The light coming through the slit, which corresponds to an across-track angle of 115∘, or 2600 km on the ground, is dispersed by optical gratings and mapped on two 2D CCD detectors. Each detector image is aggregated across-track (along the length of the slit) into 60 spectra, corresponding to 60 across-track spatial pixels for the UV2 (307–383 nm) and visible (349–504 nm) bands, as shown by Fig. . Although the spatial response functions of OMI pixels are nonuniform , the OMI pixels are widely characterized as quadrilateral polygons defined by 75 % of the energy in the along-track field of view (FOV) and the halfway points of the across-track FOV the 75 FOV pixel edges from the OMPIXCOR product;. These OMI pixel polygons are close to rectangles, ranging from 14km×26km at nadir (or 13km×24km if assuming nonoverlapping pixels) to 28km×160km at the swath edges. Alternatively, OMI pixels can be represented as tiled polygons with no overlap between adjacent pixels. These tiled pixels produce a seamless swath image but are less accurate, especially in the along-track direction. OMI is a highly successful mission with long data records, and most of the successor missions follow a similar design . The oversampling technique demonstrated here can be readily adopted for a range of OMI products and OMI's successor missions, such as OMPS, TROPOMI, and TEMPO.

The IASI instrument is an FTS with an across-track scanning range of 2200 km (Fig. ). It has a daytime equatorial crossing time of ∼09:30 LT. The first IASI instrument (IASI-A) was launched aboard the MetOp-A satellite in 2006, with the launch of IASI-B following in 2012 and IASI-C in 2018. IASI scans across the track with 30 mirror positions, or fields of regard (FORs), and each FOR is composed of a 2×2 array of pixels, or FOV. Each FOV projected on ground is a 12 km diameter circular footprint at nadir and elongates to ellipses towards the swath edges . To simplify the ground pixel calculation, we represent each pixel as an ellipse with the major and minor axes and rotation angle interpolated from a lookup table based on latitude and FOR and FOV number.

We use the most recent neural network (NN) IASI NH3 retrieval based on calculation of a hyperspectral range index (HRI) and subsequent conversion to NH3 columns via a neural network . The IASI NH3 datasets are publicly available for both IASI-A and IASI-B, with the version 2  presenting significant improvements over version 1 , including the negative values that are crucial for observational error averaging near the detection limit.

The CrIS instrument, which is aboard the Suomi NPP satellite and the series of JPSS satellites, is a step-scan FTS with 2200 km across-track width (Fig. ). It has a daytime equatorial crossing time of ∼13:30 LT. It has the same number of FORs as IASI, but each FOR contains 9 FOVs (3×3 array), providing a better spatial coverage. Each CrIS FOV is 14 km at nadir, slightly larger than IASI. Due to the mounting angle of the scanning mirror, the FOR rotates differently at each scanning angle. Similar to IASI, each CrIS pixel is represented as a rotated ellipse.

The CrIS fast physical retrieval (CFPR) NH3 retrieval product is based on the TES optimal estimation approach that minimizes the differences between spectral radiances and a simulated fast forward line-by-line model .

The spatial interpolation methods generate continuous data fields from observations made at discrete locations. The main difference between interpolation and the point- and polygon-based oversampling approaches discussed in Sect.  and is that the values at grid cells that are not covered by satellite observations can be estimated. Therefore, the spatial interpolation methods are more commonly used for satellite datasets with significant spatial gaps or requiring additional smoothing. Common spatial interpolation methods include nearest neighbors, piecewise 2-D linear interpolation, spline interpolation, and various kriging methods. The moving window block kriging method has been proposed to generate global level 3 products for satellite observations of long-lived species, such as CH4 and CO2 . A comprehensive review of available spatial interpolation methods for environmental variables is provided by . There are relatively few applications of spatial interpolation methods to regional fine-grid oversampling, where each target grid cell usually receives a large number of overlapping satellite observations. proposed an interpolative gridding algorithm that reconstructs the trace gas distribution by a continuous parabolic spline surface, defined on the lattice of tiled satellite pixels. This approach produces smooth regional level 3 maps for the OMI NO2 products with specifically tuned smoothing parameters but has not been tested in non-tiled observations with significant numbers of missing values (e.g., IASI and CrIS).

The simple “drop-in-the-box” gridding method can be classified into this category, as each satellite observation is assumed to be a point on the surface. The value for each target grid cell is the average of all screened satellite observations with the center of the FOV falling inside the grid cell boundaries. A conventional oversampling approach has been developed based on the drop-in-the-box method; instead of only averaging “in the box”, it includes satellite observations within a certain radius (much larger than the grid size) from the center of each grid cell. This averaging radius is chosen to balance the smoothing and noise but is also somewhat arbitrary. For example, used a radius of 8 km to oversample the OMI NO2 tropospheric columns and a larger radius of 24 km to oversample the OMI SO2 total columns near the Canadian oil sands region; used 12 km to oversample the OMI SO2 total columns over the US; and used 24 km to oversample the HCHO total columns near Houston, TX. This oversampling approach is referred to as “point oversampling” hereafter, as the pixel geometry is not considered. The pixel-specific observational errors are also not taken into account.

Figure  reconstructs a point oversampling process for an arbitrary target grid point (red star) located near Denver, CO. OMI NO2 data  over the year 2005 are used in this demonstration. Pixels with a cloud fraction ≥30 % or a solar zenith angle ≥75∘ are screened out. Only across-track positions with relatively small pixel areas (6–55 out of 1–60) are included, a common practice to oversample OMI data. Adding pixels at the swath edges would induce more “false negative” cases, as shown below. The screened satellite pixel centers that fall within a 12 km radius (dashed circle) are plotted as black points and red triangles. The red triangles are “false positive” observations because the corresponding pixel quadrilaterals, provided by the OMPIXCOR product, do not cover the target grid point. The pixel geometry of an extreme false positive case is illustrated by the pixel quadrilateral, featuring the largest separation between its boundary and the target grid point. Likewise, the false negative observations are plotted as purple squares, whose pixel centers fall outside the averaging circle (and hence not averaged), but these pixels cover the target grid point. An extreme case of the false negatives is also illustrated. For this example, there are 243 pixels within the 12 km radius, of which 54 are false positives (22 %). There are 92 false negatives (38 %) not included in the point oversampling. Typically, false positives are pixels closer to nadir, whereas false negatives are pixels away from nadir. In combination, the oversampled value at this grid location has contributions from a much different set of satellite observations than what should be represented. A larger averaging radius will decrease the occurrence of false negative cases but increase that of false positive cases. Because the OMI pixel dimension is larger at the across-track direction, these sampling biases differ in direction; observations in the across-track direction of the target grid point are more likely to become false negatives, and observations in the along-track direction are more likely to become false positives.

In reality, the OMI ground pixel footprints are not as sharp as quadrilateral boundaries , so the false positive and negative cases are not as well defined as in Fig. . This will be discussed in Sect. .

This approach assumes that each satellite observation footprint is a polygon on the surface, and calculates the areal proportions of grid cells inside each polygon. Because calculating these overlapping areas requires filling irregular satellite footprint polygons with rectangular grid cells, it is also known as the “tessellation” approach. The contribution of each satellite observation to a given grid cell is weighted by the overlapping area and inversely weighted by the total pixel polygon area and the observational uncertainty, as shown by the following equations modified from: C(j)=A(j)/B(j), where A(j)=∑iΩ(i)S(i,j)σ(i)p∑jS(i,j),B(j)=∑iS(i,j)σ(i)p∑jS(i,j). In the equations above, C(j) is the oversampled result for destination grid cell j; Ω(i) is the variable to be oversampled (e.g., total column) associated with the satellite pixel i; S(i,j) is the overlapping area between pixel i and grid cell j, and hence ∑jS(i,j) is the total area of pixel i, assuming that the grid extends beyond all pixel boundaries. When the destination grid is regular with constant grid cell area, it is convenient to normalize S(i,j) by the grid cell area, leading to overlapping fractions. We will follow this convention hereafter, and hence S(i,j) is always a dimensionless number. These equations take into account the extent of a pixel and give more weight to a nadir observation than to an observation at the edges of the satellite swath, where the information is more smeared out. The variable σ(i)p is the uncertainty term, and the power p has been assumed to be 1  or 2  by different studies. If we assume each observation Ω(i) is a measurement of a constant true value with Gaussian random error σ(i), p=2 yields the maximum likelihood estimate of the true value. However, the true measurement and sampling errors often show heavier tails than a Gaussian distribution. In this study we adopt p=1, following . The oversampled results are generally similar for both cases. Unlike the point oversampling discussed in Sect.  where C(j) is simply the average of Ω(i) within a circle, the tessellation approach fully utilizes the geometry and error information for each satellite observation. It has been adopted by many operational level 3 products and oversampling studies .

It is sometimes convenient to define D(j)=∑iS(i,j) to quantify the total number of overlapping pixel polygons used in averaging for grid cell j. Unlike the point oversampling, this number does not have to be an integer due to the consideration of partial overlaps. Because the location and size of these pixels vary day by day, averaging a large number of pixels reveals spatial patterns at scales finer than the satellite pixel scales, if these patterns are consistent through the averaging time period.

Figure  illustrates the tessellation process for OMI (a) and IASI (b) pixels, where the elliptical IASI pixel is represented by a 100-vertex polygon calculated from its minor/major axes and rotational angle lookup tables. The destination grid size is 5 km × 5 km, and the overlapping areas are normalized by the grid cell area (25 km2), as labeled in each grid cell.

The tessellation approach discussed in Sect.  inherently assumes that the satellite observation is uniformly sensitive to the scene inside the pixel polygon and has no sensitivity outside it. However, depending on target grid size and the spatial response function of specific satellite observations, this may be too strong of an assumption. For example, characterized the complex spatial response function of the AIRS instrument and used it to improve the comparison of radiances measured by AIRS and MODIS. and derived an in-flight spatial response function of OMI using collocated MODIS radiance. The operational Sentinel-5 Precursor, Sentinel-5, and Sentinel-4 cloud processors also rely on the spatial response functions of the imaging grating spectrometers to accurately calculate the cloud coverage within each FOV using collocated high-resolution cloud imagers .

For imaging grating spectrometers like OMI, the spatial response function depends on the diffraction of the fore optics, the instantaneous field of view (i.e., the instantaneous projection of the slit on the ground from the point of view of a native detector pixel), the numbers of across- and along-track bins, and the along-track movement of subsatellite point during the integration time. The satellite movement only affects the along-track direction, generally making the spatial response in the along-track direction smoother than that in the across-track direction. and fitted the OMI spatial response function using a 2-D super Gaussian function to parameterize the different smoothness in the along- and across-track directions. To standardize the representation of spatial response functions for diverse satellite sensors, we generalize the 2-D super Gaussian function as S(x,y)=exp⁡-xwxk1+ywyk2k3, where wx=FWHMxln⁡(2)1/(k1k3),wy=FWHMyln⁡(2)1/(k2k3). In these equations, x and y are distances to the center of ground FOV in orthogonal directions, usually transformed by geometric projections of the across- and along-track directions. FWHMx and FWHMy are full widths at half maximum of the spatial response function, S(x,y), in the directions of x and y. The three exponential terms, k1, k2, and k3, control the distribution of spatial response, as illustrated by Fig. . When k3=1 (Fig. a and c), Eq. () becomes the 2-D super Gaussian function used by and to characterize the OMI spatial response: S(x,y)=exp⁡-xwxk1-ywyk2. For OMI, k1∼4 and k2∼2 .

For FTS systems with stop-and-stare sampling, like IASI and CrIS, the spatial response function (also known as point spread function by the community) is more simply defined by the circular aperture and some diffraction around the edge. The nadir FOV is circular with no difference between across- and along-track directions, and hence the spatial response function can be characterized by a 1-D super Gaussian function rotating around the nadir point. This rotating super Gaussian function is another special case of the generalized 2-D super Gaussian (Eq. ) with k1=k2=2 and wx=wy: S(x,y)=exp⁡-Rw2k3, whereR=x2+y2 and w=wx=wy. The smoothness of the rotating super Gaussian is controlled by only one exponent, which equals to 2×k3. The elongated spatial response functions for off-nadir angles can be readily characterized by different values for wx and wy (Fig. a–b). The spatial response function of IASI is rather sharp at the edge with little variation at the top, close to a super Gaussian with an exponent of ∼18 . The spatial response function of CrIS is relatively smoother at the edge, best fit by a super Gaussian with an exponent of ∼8 . Details on the spatial response functions of IASI and CrIS can be found in Appendix A.

In the generalized 2-D super Gaussian function (Eq. ), k1×k3 and k2×k3 are the exponents in the x and y directions, respectively, and determine the sharpness of the spatial response in the corresponding direction. An exponent of 2 leads to a standard Gaussian function; the larger exponents produce a top-hat shape, converging to a boxcar shape when the exponent approaches infinity . Redistributing the contributions from k1/k2 and k3 makes hybrid spatial response functions that may have sharp edges in sensitivity but rounded corners in space, as in the case of OMPS (Glen Jaross, personal communication, 2017). The difference between this hybrid case and conventional 2-D super Gaussian is illustrated by Fig. c–d.

The projection of a rectangular FOV for imaging grating spectrometers like OMI on the surface at large viewing angles leads to distorted quadrilateral footprints, as shown by the polygon ABCD in Fig. a. To account for this effect, a geometric transformation function is determined by the OMI pixel corner points (ABCD in Fig. a) and the corresponding rectangle (A′B′C′D′ in Fig. b) defined by the distances between the middle points of opposing edges of the OMI pixel quadrilateral. The spatial response function is first calculated according to Eq. () with FWHMx =|A′D′| and FWHMy =|A′B′| as shown in Fig. b and then projected to match the OMI pixel corners ABCD (Fig. a) using the geometric transformation function. This algorithm is implemented using both the OpenCV library in Python and the Image Processing Toolbox in MATLAB.

The proposed oversampling approach represents each satellite observation as a sensitive distribution, instead of a point or a polygon. If the true satellite spatial response function is used as the sensitive distribution, this approach is the theoretically optimal solution to the oversampling problem, and is hence referred to as “physical oversampling” hereafter. It follows the same equations as the tessellation approach as in Eqs. ()–(), except that the fractional overlapping area S(i,j) is generalized to the integration of the spatial response function of satellite observation i, S(x,y|i), over the grid cell j: S(i,j)=∬gridjS(x,y|i)dxdy∬gridjdxdy, where the denominator is the grid cell area. Similar to the tessellation approach, S(i,j) is always a dimensionless number between 0 and 1. By normalizing the grid cell area, this accurate form of S(i,j) can be directly replaced by approximating values such as S(x,y|i) evaluated at the grid center. S(i,j)/∑jS(i,j) is just the normalized spatial response function for observation i so that its spatial integration is unity. If the spatial response is uniform inside the pixel polygon and zero outside the polygon, this integration of the spatial response function within the grid cell is equivalent to the fractional overlapping area used in the tessellation approach. As such, the tessellation is just the extreme case where the spatial response function is a perfect 2-D boxcar. This corresponds to k1×k3→∞ and k2×k3→∞ in Eq. ().

This physical oversampling approach can also be considered as a spatial interpolation method as discussed in Sect.  because the spatial response function can be defined beyond the satellite pixel boundaries and theoretically on the entire 2-D space. Moreover, instead of the exact form of spatial response function, the satellite observations can be represented by similar (with the same FWHM) but smoother sensitivity distributions to enhance the quality of the oversampling results. This possibility will be demonstrated in Sect. .

The tessellation approach is perfect if the spatial response of satellite observation is a boxcar, but otherwise it will introduce some error in the oversampled results (referred to as “tessellation error” hereafter). When the satellite spatial response function is smooth (instead of a boxcar), the exact solution is to calculate S(i,j) as the integration of the spatial response of satellite observation i over the area covered by the target grid cell j (Eq. ). It is computationally demanding to numerically integrate the spatial response of all satellite pixels over each grid cell. To simplify it, one may discretize the spatial response function to the target oversampling grid and use the spatial response value at the grid center to approximate the integration. As such, the spatial response function only needs to be evaluated once per pixel per grid cell. To improve this simple discretization scheme, we calculate a weighted average of the spatial response values at the grid center and grid corners as proposed for MODIS by. Because the grid corners are shared by neighboring grid cells, this approach only doubles the spatial response calculation but significantly reduces the error induced by discretization (“discretization error” hereafter). Appendix B gives a detailed comparison of different discretization schemes.

The satellite sensors have very different spatial responses. The target grid size for level 3 data ranges from 0.25∘ (∼25 km) for many global operational products to 0.01∘ (∼1 km) for regional oversampling. The discretization error decreases as the size of the target grid cells becomes finer and the spatial response of satellite observations becomes better resolved. At any fixed target grid size, spatial response functions with smoother edges are better approximated by the discretization scheme. As such, it is essential to balance the tessellation and discretization errors based on the target grid cell size and the smoothness of the satellite spatial response so that the most accurate and efficient approximating method can be chosen.

Figure  compares the tessellation and discretization errors when oversampling synthetic OMI observations to a grid of 1 km (∼0.01∘). A checkerboard pattern is used as the “true” concentration distribution (alternating values of zeros and ones with a spatial period of 20 km × 20 km, as shown in Fig. a; it also shows OMI pixel polygons at across-track position no. 1 in red and across-track position no. 30 in cyan). Synthetic OMI observations are generated by sampling the checkerboard pattern using the OMI spatial response function, simplified using Eq. () with k1=4, k2=2 and discretized at a very fine grid (0.05 km, or ∼0.0005∘) so that the spatial response distribution is always fully resolved. The locations of OMI observations are from the real OMI NO2 products , filtered by cloud fraction <25 % and solar zenith angle <75∘ for 2005–2006. Instead of NO2 columns, the synthetic OMI observations at these locations are oversampled. The oversampled area is in the north midlatitude (∼40∘Ṅ). In Fig. b, the oversampling is conducted at a native grid size (0.05 km), and then the result is block-averaged to the 1 km target grid size to represent ideal OMI observations, as in Eq. (). One should note that this discretization at 0.05 km is used to get the true map of OMI observation where the discretization error is negligible. It is unnecessary to oversample at this fine grid in general. Figure c and e show the results for tessellation and discretization of the spatial response at 1 km grid, where S(i,j) is approximated by fractional overlapping area and the discretization scheme, respectively. They both reproduce the checkerboard pattern in general, but the tessellation method generates errors up to 40 % (Fig. d) relative to the peak-to-trough value of the ideal observation because the OMI spatial response is smooth (Fig. ) instead of boxcar. In contrast, the discretization error is much smaller (Fig. f) because of the small size of the target grid cells (1 km).

The analysis for Fig.  is repeated for a range of target grid sizes (1–50 km, or about 0.01–0.5∘) and different smoothness of the spatial response functions using the same OMI observation locations. The spatial response function is assumed to be 2-D super Gaussian (Eq. ). The exponent in the along-track direction (k2) is tuned from 2 to 64, whereas the exponent in the across-track direction (k1) is set to be 2×k2. Figure a shows, for satellite observations with a quadrilateral FOV, the contour of the ratio between the discretization error and the tessellation error, calculated as the root-mean-squares of the differences between the ideal observation and the simplifications using tessellation and spatial response discretization, respectively. The contour line of unity divides the regimes where tessellation and discretization errors are dominant: discretization of the spatial response is more accurate for fine-grid oversampling of satellite observations with smooth spatial responses (small k1 and k2); tessellation is more accurate for coarser target grids and sharper spatial responses. Tessellation is perfect if k1 and k2 both approach infinity. The case of OMI (k1=4, k2=2) lies at the left edge (red vertical dashed line in Fig. a), and its intersect with the unity contour line is located at the target grid size of ∼16 km. In other words, it is beneficial to explicitly consider the spatial response of OMI observation for target oversampling grids finer than ∼16 km (about 0.15∘).

Similarly, Fig. b shows the ratios between discretization and tessellation errors for satellite observations with circular FOVs. The pixel dimensions and locations of IASI observations for 2015–2016 are used with standard data screening, and the spatial response function is assumed to be a rotating super Gaussian (Eq. ). The exponential term (equal to 2×k3) varies from 2 to 64. When characterizing the IASI spatial response as a rotating super Gaussian function, the exponent is about 18, intersecting the unity contour line at the target grid size of ∼2 km. If the IASI instrument had the same spatial response as CrIS (the exponent is about 8), the intersect would be at the target grid size of ∼4 km. The results would be very similar when using the CrIS observation locations instead of IASI because the exact locations of any observations are averaged out and the IASI and CrIS pixel sizes are similar.

As shown by Fig. , the balance between tessellation and discretization errors depends on both the target grid size and the deviation of satellite spatial response function from an ideal 2-D boxcar shape. The uncertainty in the knowledge of the spatial response functions is not considered here, but the spatial response function can be characterized prelaunch and validated on orbit . For all three cases, the tessellation error significantly outweighs the discretization error at 1 km oversampling grid size by a factor of 4 for IASI and over 200 for OMI. Therefore, we recommend discretization of the spatial response function at a 1 km (or 0.01∘) grid for regional scale oversampling of OMI, IASI, and CrIS data and then co-adding to coarser grids if necessary. The threshold grid size where tessellation and discretization errors balance also depends on the ground size of satellite FOV. For the OMI successor missions with significantly smaller pixels (e.g., TROPOMI, TEMPO), the threshold grid size is expected to be finer.

The difference between resolution and sampling density for 1-D spectral data has been thoroughly discussed in the literature e.g.,. However, for 2-D, spatially resolved data, it is common to refer to both the sizes of the level 2 pixels and the size of the level 3 grid as the spatial “resolution” of the data. To avoid confusion, it is emphasized here that the true spatial resolution is limited by the sizes of level 2 pixels. The size of level 3 grid only determines the density of spatial sampling, which does little to enhance the true resolving power of the data after reaching a certain point. For example, the oversampling results using synthetic OMI data at 1 vs. 0.05 km grids are very similar (Fig. ). Nonetheless, it is still beneficial to oversample, i.e., make level 3 grid size significantly smaller than level 2 pixel sizes, as demonstrated by Fig. . As the ground truth, an array of 2-D Gaussian functions are generated with FWHM ranging from 1 to 16 km (the second column of Fig. ) and peak height of unity, and this true field of concentration is measured by an imaginary sensor whose spatial response function is a 2-D super Gaussian (Eq. ) with FWHM = 10 km and k1=k2=8 (the first column and the white boxes inserted in the third column). The third column shows the oversampling results using 10 000 randomly located observations. The fine structures in the ground truth are clearly smoothed, limited by the spatial resolution that is inherent to the level 2 pixel sizes (10 km). However, by oversampling at a fine grid (0.2 km for the first row vs. 5 km for the second row), the spatial gradients are better recovered, and spatial features finer than individual level 2 pixels can be identified. Additionally, the details in the spatial response function is better resolved with a finer target grid, which is particularly beneficial when collocating with higher resolution measurements (e.g., a cloud imager). As such, although the spatial resolving power is ultimately determined by the spatial extent of satellite pixels, the physical oversampling approach helps in enhancing the visualization of spatial gradient and the identification of emission sources.

Figure  compares the drop-in-the-box method, point oversampling, tessellation, and physical oversampling using OMI NO2 tropospheric vertical column density (TVCD) within a 200 km × 200 km square centered around a power plant in Arizona. The first column shows the simple drop-in-the-box method on a 10 km grid. The second column averages OMI observations within a 12 km radius of each grid center. These two approaches assume OMI observations as points without consideration of pixel geometry and retrieval uncertainties. The third column shows results using the tessellation approach, and the fourth column shows the physical oversampling using the OMI spatial response functions as a 2-D super Gaussian function with k1=4 and k2=2. The target grid size is 1 km for the last three approaches. The first and third rows show the oversampled results (C(j) in Eq. ) using 5 days (1–5 July 2005) and 5 months (May–September 2005) of data, respectively. The second and fourth rows show the corresponding numbers of pixels included in the averaging for each grid cell (D(j) in Eq. ). For the drop-in-the-box approach, the total number of satellite observations included for each grid cell is much smaller and shown with a different color scale for the 5-month averaging.

The drop-in-the-box approach shows significant data gaps (5-day averaging) and high level of noise (5-month averaging), even when its target grid is 10 times coarser than the other oversampling approaches. There are two gaps where no observation is available for point oversampling over the 5 days (column 2, rows 1–2 in Fig. ), which is an example of false negatives as these gaps are actually covered by OMI pixels (column 3, rows 1–2 in Fig. ). The physical oversampling in the fourth column consistently shows the smoothest results with clear identification of the point source at the center of the domain, because the spatial response function of OMI is properly incorporated. The oversampled NO2 TVCD is biased high for the point oversampling approach because all observations within the averaging radius are averaged equally, but larger observation values generally are associated with larger uncertainties. The results from tessellation become increasingly similar to those from physical oversampling for longer averaging times, because the tessellation error is randomly distributed and will eventually be averaged out. The physical oversampling also does not require more computational resources than point oversampling and tessellation, making it suitable for a wide range of spatial scales and target grids.

Although the physical oversampling using the true satellite spatial response functions produces the optimal estimation, the result is sometimes noisy and even unphysical, especially when the observations are noisy and sparse. In these cases, some spatial interpolation or smoothing methods are often needed. In addition to the specialized interpolation and smoothing methods discussed in Sect. , some smoothing can be applied within the oversampling framework. For example, the level of smoothing can be adjusted by the averaging radius in the point oversampling approach. used a Gaussian filter to smooth tessellation results for OMI HCHO and CHOCHO products. When using the generalized 2-D super Gaussian function to characterize the satellite spatial response function (Eq. ), it is also simple to tune the exponents (k3 in the cases of circular FOVs such as IASI and CrIS and k1 and k2 in the cases of quadrilateral FOVs such as OMI) so that the assumed satellite spatial sensitivity distribution is smoother than the true spatial response function. This often leads to better visualization and identification of local hot spots, especially for products with a high noise level or sparse spatial sampling. The advantage of this approach is that the smoothing is applied at the satellite pixel level (level 2) instead of grid level (level 3), so the geometry and error information for each satellite observation are preserved.

Figure  shows similar oversampling results as Fig. , but using IASI NH3 total column density data  for 2015 in eastern Colorado, centered around a large cattle feedlot. The drop-in-the-box approach is not shown for IASI. The results from point oversampling, tessellation, and physical oversampling to a 1 km grid are presented in the first three columns. The true IASI spatial response functions have rather sharp edges (see Appendix A), so the physical oversampling shown in the third column of Fig.  is very similar to tessellation shown in the second column. Although this is the optimal estimation based on the physics of IASI observation, the spatial gradients are hard to identify for 5-day averaging and noisy for 5-month averaging. Instead of applying smoothing after the oversampling process, the fourth column uses a smooth spatial sensitivity distribution of a 2-D standard Gaussian function (2k3=2, rather than the true IASI spatial response function with 2k3∼18). As illustrated by the first row in Fig. , the physical oversampling using smoother spatial sensitivity distributions provides the best results by clearly identifying the central point source using only sparse (5-day) data. The third row in Fig.  demonstrates that with 5 months of averaging, the local NH3 gradients are well resolved. The point oversampling using a 12 km radius overly smooths the results, making the central hot spot artificially larger, whereas the general spatial gradients are still noisy (column 1, row 3). The overall number of IASI observations used in point oversampling is also significantly higher than tessellation and physical oversampling, as shown by the fourth row. This is because the 12 km averaging circle is much larger than most IASI footprints, and hence many IASI observations are double counted as false positives. The smoothing based on physical oversampling is much more effective in suppressing the noise, and the spatial gradients are adequately preserved (column 4, row 3). This is because each satellite FOV keeps the same FWHM and overall weight, and only the distribution of sensitivity becomes more spread out.

Oversampling based on Eqs. ()–() also provides a flexible way to categorize the results according to environmental and temporal variables. The conventional way is to save the averaging weights for each level 2 observation (i.e., the level 2G product, where level 2 pixels are assigned to points of the latitude and longitude grid), but the averaging weights can only be defined for a specific grid. When representing each level 2 observation as a spatial sensitivity distribution (the actual instrument spatial response function or a smoother version of it), A(j) and B(j) can be calculated at fine spatial and temporal grids and then aggregated spatially and/or temporally. The level 3 map C(j) is just the grid-by-grid ratio of the aggregated A(j) and B(j). Similarly, A(j) and B(j) can be calculated according to environmental variables such as wind and temperature at fine intervals and binned to coarser categories as needed. Figure  shows the physical oversampling of NH3 total column under southerly winds (meridional wind component >0, panels a and c) and northerly winds (meridional wind component <0, panels b and d) and high PBL temperature (>15 ∘C, panels a and b) and low PBL temperature (<15 ∘C, panels c and d). Here the PBL temperature is the average air temperature from the surface to the top of the PBL, weighted by pressure. The average wind speed and wind direction under each category are labeled in the corresponding panels. IASI-A daytime data from 2008 to 2017 over northeastern Colorado are included in the oversampling, and a 2-D standard Gaussian is used as the spatial sensitivity distribution to smooth the results. The 3-D wind field, atmospheric temperature, surface pressure, and PBL height are interpolated from the North American Regional Reanalysis NARR; from their native resolutions of 32 km and 3 h to the IASI pixel locations and overpass time. Using the concentrated animal feeding operation (CAFO) locations (colored dots; data courtesy of Daniel Bon, Colorado Department of Public Health and Environment) as a spatial reference, the downwind dispersion of the total NH3 column under different wind directions is clearly seen. The close match between large cattle CAFOs and the NH3 hot spots seen from space confirms that they are the dominant source of atmospheric NH3 in this region. The overall abundance of NH3 is significantly higher at warmer temperatures, in agreement with the previous in situ quantification of CAFO NH3 emissions in the same region .