The Challenge of Surface Type Changes Over the Aral Sea for Satellite Remote Sensing of Precipitation

Frequent false signals of precipitation in satellite passive microwave retrievals over the Aral Sea have been identified as being caused by an outdated surface database. The database includes the surface type, elevation, and the percentage of the primary surface type in each grid. It was also found that the grid resolution of 1/6 degree (∼18 km at the equator) of the outdated surface data was too coarse to process global precipitation measurement mission microwave imager data, which have a field of view resolution of 31.68 km2 for channels at 89, 166.5, and 183.31 GHz). In this article, we generated a new surface database at a resolution of $0.05^\circ \times 0.05^\circ $ (∼5.6 km at the equator). Using the new surface data, the false precipitation problem above is addressed. At the same time, the retrieval accuracy for other parameters such as total precipitable water over the Aral sea is significantly improved as well. The resolution of the database is suitable for most microwave and infrared sounding measurements of atmospheric temperature and moisture profiles as well as precipitation. By comparing this new surface data against the outdated surface data, we also see the loss of permanent ice in Antarctica and a dramatic reduction of water surface over the Aral sea. Using a 40-year record of remote sensing data, we can observe the steady decrease in size of the Aral sea, as a result of regional water use policies and natural climate change.

was successfully launched on November 18, 2017. JPSS [2] will provide continuity of critical global observations about climate change, air quality, natural disasters such as tornados, hurricanes, and wildfires through 2039 and beyond. Remote sensing data [3] play an important role in support of weather forecasting and the monitoring of environmental change. The National Oceanic and Atmospheric Administration (NOAA) Microwave Integrated Retrieval System (MiRS) generates daily 10 satellite products or environmental data records (EDRs) including precipitation rate from passive microwave measurements, such as the SNPP and NOAA-20 Advanced Technology Microwave Sounder (ATMS) instruments [4]. The ten EDRs greatly support the monitoring of severe weather and climate studies. However, environmental variations over time, such as water/land surface changes can challenge remote sensing data applications. We recently became aware (from a regular user of MiRS retrieval products) of a false precipitation signal in MiRS operational retrievals produced on August 17, 2021 over the Aral sea. The issue was carefully investigated and it was found that changes to water/land surface coverage was the root cause. We then checked the MiRS precipitation product back to 2020 and found the false signals of precipitation over Aral sea were quite frequent, occurring at least several times per month. MiRS uses a surface dataset, including 24 surface types, at 1/6°(∼18 km at the equator) resolution. Because the emissivity characteristics of land and water surfaces are so different in the millimeter and microwave spectral region, MiRS uses water/land information from the dataset to choose the appropriate a prior constraints and retrieval state vector basis functions. In particular, in the MiRS variational algorithm surface emissivity empirical orthogonal functions (EOFs) and the surface emissivity mean (a prior spectrum) as well as the error covariance matrix are dependent on the surface type. Because microwave emissivity of water is much lower than that over land, and has a different dependence on frequency, incorrect surface-type information can lead to improper prior constraints. This can pose problems for the retrieval because the inversion is an ill-posed problem (i.e., multiple possible solutions can correspond to the same set of measurements). The surface area of the Aral sea has been steadily shrinking since 1960 as the former Soviet Union converted large acreages of pastures into irrigated farmlands and diverting water that originally fed the Aral sea (https://www.britannica.com/place/Aral-Sea). In addition, changes to the regional climate have accelerated this trend. Precipitation over Aral sea decreased from 9.4 km 3 in 1960 to This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 3.2 km 3 in 2009 and to 2.1 km 3 in 2010 [5]. During a similar period, the water surface area decreased from 68 000 km 2 in 1960 to less than 10 000 km 2 in 2017 [6]. This significant change of water surface area poses challenges for the MiRS precipitation retrieval which currently uses static and historical land/water surface data. This article is organized as follows: We introduce the global surface type in Section II; Aral Sea's land and water surface change is described in Section III; Section IV discusses MiRS precipitation remote sensing; and Section VI leads to the summary and discussions.

II. GLOBAL SURFACE TYPE
Global surface-type classification and temporal changes to the surface classification provide critical information about climate change and human activities. In the 1990s, the International Geosphere-Biosphere Programme (IGBP) recommended IGBP surface-types using Advanced Very High-Resolution Radiometer (AVHRR) measurements [7]. The AVHRR red band at 0.615 µm and near infrared band at 0.912 µm [8] are good for calculating normalized difference vegetation index (NDVI) which strongly correlates with the greenness of surface vegetation. The NVDI [9] was defined as where R nir and R red are the reflectances or albedos of the AVHRR bands 2 and 1, respectively. Healthy vegetation (chlorophyll) reflects more near-infrared (NIR) and green light while it absorbs more red and blue light. Therefore, vegetation greenness is a measure of the healthiness of vegetation. NDVI has been used for many applications such as fraction of photosynthetically active radiation [10] and vegetation-climate interactions [11]. NDVI is the key parameter for determining global green vegetation fraction [12]. The NDVI is also an important parameter to the vegetation on land for IGBP surface-type classification [13], [14]. There are 17 IGBP surface types [15]. A total of 15 of 17 IGBP surface types depend on soil and vegetation.
In order to evaluate the existing surface-type datasets, we first received a dataset of 24 surface types in 1998 from National Polar-orbiting Operational Environmental Satellite System (NPOESS) Integrated program office. This dataset is called NPOESS surface data [16], however there appears to be no official documentation or publication which describes the origin of the NPOESS surface data. The dataset includes the surface elevation and the percentage of the primary surface type at a resolution of 1/6°(∼18 km at the equator). A static surface reflectance database for the 24 surface types and wavelengths between 0.2 µm and 15 µm is also included with the surface-type data. The two datasets had been used to simulate the visible infrared imaging radiometer suite (VIIRS) radiances for the algorithm development and validation such as the sea surface temperature algorithm and the surface net heat flux algorithm. The databases are apparently useful and are still used by the NOAA MiRS [17] and the community radiative transfer model (CRTM) [18]. Further details about the MiRS algorithm are provided in Section IV.
The CRTM model, developed at the Joint Center for Satellite Data Assimilation, is an operational radiative transfer model in support of weather forecast and the generation of many EDRs. The CRTM model is widely used by numerical prediction centers, satellite retrieval products, and research and education communities. The CRTM embeds the surface reflectance for NPOESS 24 surface types. However, changes in surface type that may have occurred over at least the last 20 years are not included in the database, and the 1/6°resolution is inadequate for use with microwave imagery data such as GMI data. In this article, we use the latest NOAA surface-type data [19] derived from the VIIRS radiance data in 2020. There are 17 IGBP surface types [14]. The NOAA Environmental Modeling Center then added three tundra surface-types resulting in a total of 20 surface types. This dataset is called NOAA surface-type data. The NOAA surface-type dataset has a high spatial resolution of 1/120 degrees, which is needed for studying detailed surface types by using visible imagery observations. However, microwave sounding data have a coarse resolution. For example, the spatial resolution at nadir for the Advanced Microwave Sounding Unit-A (AMSU-A) is 48 km. The ATMS has a sample resolution of 16 km at nadir. Global precipitation measurement (GPM) microwave imager (GMI) high frequency channels have a high spatial resolution of 5.6 km for channel central frequencies higher than 89 GHz. In this article, we downscale the database surface type and elevation from 1/120°to 1/20°resolution. It should be pointed out that Aral Sea land/water masks in this 1/120 high resolution haven't been updated for many years. In this article, we directly derived the Aral sea land/water fraction from VIIRS radiance (see Section III) and included the land/water fraction in our surface-type data for MiRS retrieval system, which is called NOAA-MiRS surface-types hereafter.

III. LAND & WATER CHANGE IN ARAL SEA
In this section, we focus on surface-type changes over the Aral sea. Aral sea was the fourth-largest freshwater lake in the world (https://en.wikipedia.org/wiki/Aral_Sea). In 1960, the water body had a surface area of 68 000 km 2 in 1960, larger than West Virginia of the United States. In this article, we used 40-year long (1981-2021) record of remote sensing data to estimate changes to the water surface area in the region. To do this we use red band reflectance and near-Infrared band reflectance from AVHRR, moderate resolution imaging spectroradiometer (MODIS), and VIIRS observations to compute the NDVI. The NDVI exploits the fact that over water, the red band reflectance is higher than near-infrared band reflectance, while the converse is true over land. The threshold NDVI value to distinguish land from water is a key parameter in the calculation of water surface area. The AVHRR does not have onboard calibrators for its red band and near-infrared band, and there are significant differences among the calibration of the various AVHRR instruments deployed on satellites [20]. However, rather than an absolute calibration, we only need to be able to identify the AVHRR NDVI features for each individual instrument. The AVHRR red band and near infrared band at a 4 km gridded data are used to calculate the NDVI. As shown in Fig. 1, the histogram of the weekly AVHRR  MODIS imagery red band (channel at 645 nm) and near infrared band (channel 2 at 858 nm) have a spatial resolution of 250 meters at nadir. In comparison to the moderate spatial resolution bands, the imagery bands have better spatial resolution and its wide bandwidth are more suitable for studying vegetation and surface types. Using the NDVI histogram features for AVHRR and the threshold value of -0.05 for MODIS and VIIRS NDVI, we can calculate the water and land surface over the Aral Sea. In Fig. 2, we show the NDVI-derived surface classification, with water surface shown in blue and land surface in green.  Fig. 2 that the water surface in the Aral Sea has decreased steadily and dramatically to a point where its size is a small fraction of its pre-1981 level. It has been shown in other studies that a minimum water surface area occurred in 2014 [21], [22], [23]. The size of water surface for 2021 in Fig. 2 is very similar to the minimum water surface area occurred in 2014. The 2021 derived surface type is also in very good agreement with the VIIRS true color image. Table I gives the water surface area in square kilometers and the water area percentage relative to the water surface area in 1960.

IV. REMOTELY SENSED PRECIPITATION
MiRS has been run operationally at NOAA since 2007. MiRS has been generating m ultiple EDRs (i.e., satellite retrieval products) for NOAA satellites, European satellites Meteorological Operational Satellites Metop-B and Metop-C, GPM satellite and other satellites (https://www.star.nesdis.noaa.gov/mirs/ highresolutionv.php). The satellite retrieval products include atmospheric profiles of temperature and water vapor, cloud liquid water, ice water content, rainfall rate, snow cover and snow water equivalent, snow fall rate, surface temperature and microwave emissivity, and sea ice concentration.
The retrieval is an ill-posed problem and the information content of the satellite observations is not sufficient to uniquely determine the solution for all atmospheric and surface parameters. Prior information such as the surface type greatly helps the retrieval system by imposing constraints on the solution including use of optimal EOF basis functions for surface emissivity and the emissivity error covariance matrix, since these constraints are surface-type dependent.
MiRS precipitation retrieval is based on the measured emission and scattering information from clouds. Rain drops are much larger than the particle size for non-precipitation clouds. The microwave scattering signatures increase as the cloud particle sizes increase. The surface type is also important in the retrieval process because it determines how the surface emissivity will be calculated and used in the radiative transfer model for simulating the brightness temperatures from state variables. The MiRS retrieval starts from nonprecipitation or an emission only mode. If no profile can be found to meet a convergence within a defined number of iterations, the MiRS will start the second attempt with a small precipitation in a scattering mode. The convergence can be achieved when the difference between satellite measured and simulated brightness temperatures is smaller than a threshold that is typically comparable to the measurement uncertainty. Typical iteration number is 3 with an upper limit value of 7. The MiRS system works well. The global retrieval convergence level is about 95%.
The surface type is very important to MiRS because the surface emissivity between oceans and lands are very different. However, it is known that the surface type in the Aral Sea region has changed dramatically as the water body has shrunk in size (see Fig. 2). The Aral Sea land/water mask in the NPOESS surface data is very similar to land/water fraction for 1981 in Fig. 2. For determining surface type in the NOAA-MiRS dataset, the primary surface-type percentage is calculated from the VIIRS surface data within each 0.05°grid cell. Fig. 3 shows the MiRS derived surface type and MiRS retrieved surface precipitation over Aral Sea based on NOAA-19 AMSU-A and MHS data, corresponding to the use of two different surface databases are used: the original NPOESS surface data in Fig. 3(c), and the updated NOAA-MiRS dataset in Fig. 3(d). The surface types shown were used to determine a prior constraint in the retrieval as described above.
Most significantly, for the operational case of using the original NPOESS database, the MiRS precipitation in the upper-left map in Fig. 3 was confirmed as spurious based on surface observations and satellite true color images, which indicated mainly clear conditions. The false or spurious precipitation is directly related to the use of the outdated surface types. The outdated surface-type data shows a large portion of water surface in the center of the map. The new surface type does not show any water surface in the center region because the water coverage within the FOVs is less than 50% when the varying size of the AMSU-A FOVs is considered.
The second important factor is the footprint size of satellite measurements. The footprint size for most microwave sounders is larger than 15 km at nadir and increases with the zenith angles. The MiRS retrieval system only takes one surface type at the center point of each field of view. It is quite possible that the center point is water, for example a river or small lake, but the water area percentage is insignificant within the field of view. As one can see from the low right image in Fig. 4, there are many improper water surfaces to represent the satellite data footprints even we use the new surface with a high spatial resolution. Once we use the new surface-type data and consider the field of view size of satellite measurements, the improper water surfaces are removed (see lower left image in Fig. 4).
The false signals of the MiRS retrieved precipitation over Aral Sea were quite frequent. This article found that using the new surface-type data with a high spatial resolution, the false precipitations in the MiRS NOAA-19 products over Aral Sea on August 17, 2021 are removed. Both Global Precipitation Climatology Project (GPCP) daily precipitation analysis (https://www.ncei.noaa.gov/products/climate-data-records/ precipitation-gpcp-daily) and the MiRS rain product by using the new surface-type data reported zero precipitation over Aral sea on August 17, 2021. The GPCP analysis is quite accuracy and can be used as a reference. However, the MiRS rain product by using the old surface-type data reported a mean rain rate of 7.83 and 6.61 millimeters for the same day for NOAA-19 ascending and descending orbits, respectively. The mean value is the averaged MiRS retrieved rain rate over a selected area  Using the new surface-type data in the MiRS improved the retrieval accuracy for temperature and water vapor profiles as well as total precipitable water (TPW). Over the selected Aral Sea area above, the standard deviation error in MiRS retrieved TPW was reduced from 8.15 to 2.47 mm (see Fig. 5) by using the new surface-type data. The differences between European centre for medium-range weather forecasts (ECMWF) analysis and the MiRS retrieval for temperature and water vapor profiles are significantly reduced as well (see Fig. 6).

V. DISCUSSION
The Aral sea was once the fourth-largest freshwater lake in the world. However, water use policies in the former Soviet Union in combination with regional climate changes resulted in reduced water inputs to the Aral SEA. Recent decreases in precipitation related to climate change has accelerated the reduction in size of the Aral sea. The surface-type change can also affect remote sensing applications in which surface type is an important source of a prior information. In this article, examination of a 40-year record of remote sensing data further confirms the significant reduction of Aral Sea water surface from 1981 to 2021. Our study found that this surface-type change related in part to climate change can impact remote sensing data applications. MiRS false precipitation retrievals over the Aral sea are an indirect consequence of climate change as the static ancillary surface database was not updated to account for local and temporal changes. To mitigate the risk, it is necessary to update prior static information frequently. We have generated surface data using 2020 VIIRS surface-type data and NOAA-20 VIIRS radiances over Aral sea. As shown in Fig. 2, for the Aral sea region, the NOAA-MiRS surface-type data have very good agreement with the VIIRS true color image. In addition, when implementing the higher spatial resolution database, it is critical that satellite FOV size is taken into account so that the database is sampled at the approximate spatial resolution of the satellite measurements.