Inland Open Water-Body Detection Using Multipolarized L-band SMAP SAR

The Soil Moisture Active Passive (SMAP) satellite provided global measurements of normalized radar cross section (NRCS) for the period of May until July 2015. The dual-polarized (HH and VV) NRCS is used to identify inland water-bodies globally to support water and carbon science. Wetland characterization is important because the variations in surface water area explain more than 50% of the variance of methane emissions. Water surface fraction within a SMAP's radiometer footprint is one of the largest sources of soil moisture retrieval errors, and thus the open-water map also assists soil moisture retrieval. The proposed algorithm to identify open water-bodies is to combine a threshold to the HH/VV ratio at -3 dB and a HH threshold at -25 dB. The accuracy target is 5 to 10% for open water surface classification. The target is comparable with the passive-microwave performance in the literature and the need to support soil moisture retrieval. The algorithm was found robust with the detection error smaller than 10 % in the presence of varying incidence angle, instrument noise level, and scene diversity using airborne SAR and spaceborne scatterometer data (Kim et al. 2015).

Temporal speckle reduction allows 1-km resolution global mapping at ~ 15 day interval. A global waterbody map was derived using the preliminary SAR data from SMAP for a one-month period of May 2015. Transient features such as a flooded dry lake are correctly identified. Comparison with the 500m-resolution static database (MOD44W derived with mainly Shuttle Radar data of 2000) was made, because MOD44W is currently used to correct for the effect of waterbody extent in the SMAP's soil moisture algorithm. Over a 500km x 500km area in Ethiopia in Africa, the two maps agree over an area of 2050 km2. SMAP reports a smaller water area by 732 km2, which appears to correspond to the shrunk extents of lakes. Soil moisture retrieval also supports this attribution. SMAP produces additional 252-km2 area of water, which may as well be the recent inundation as well as false alarms. The false alarms are associated with the radar speckle and the intrinsic errors in the classification algorithm.

By treating MOD44W as a reference in this region, the probability of detection (POD) is 0.73, the false alarm rate is 0.11, the Jaccard distance is 0.32 (away from the perfect similarity of zero), and the Brier score is 0.004 (0 is the best match; the near-zero value may increase when the radar speckle error is incorporated into SMAP's detection probability). These statistics indicate that the two databases differ, partly because MOD44W is outdated as evidenced by the errors in the soil moisture retrieval. The categorical statistics were derived over about 40 waterbody targets in the world.

On going investigations are as follows. The speckle effects will be further reduced through spatial-domain analysis. The mixed-pixel cases of partial water and land will be identified using initial guess of NRCS of the two surfaces. Cross-validation with high-resolution optical data such as the Landsat8 data will help confirm the SMAP's classification. We will explore L-band capability to penetrate vegetation and clouds to monitor the underlying surface.

Kim, S.B., J. Oullette, J. J. van Zyl, and J. T. Johnson, Dual-copolarized approach to detect surface water extent using L-land radar, IEEE Trans. Geosci. Remote Sens, revision, 2015.