Algorithms

Neural Network-based Disaggregation

Fig. 1. Application of DisaggNet using SHEELS/RTM-estimated microwave emissivities produced by the model at 0.8 km grid and aggregated to the 1.6 km and 12.8 km. DisaggNet then attempts to reconstruct the original 0.8 km soil moisture pattern (bottom image).
Soil moisture estimation using microwave remote sensing data is limited by several factors, with one of the most important being the large (> 25 x 25 km) footprints of microwave radiometers. This is adequate for climate modeling but inadequate for many other applications such as agriculture or initializing mesoscale weather models. We believe that there is much valuable ancillary information that can be used to disaggregate microwave brightness temperatures or soil moisture estimates. Toward this end we have developed a Neural Network-based model called DisaggNet to address the feasibility of disaggregating low-resolution satellite microwave remote sensing data to estimate soil moisture. The purpose of a disaggregation scheme is to produce the ‘correct’ high-resolution (sub-pixel) pattern of soil moisture from lower-resolution remotely-sensed observations. We designed DisaggNet to reconstruct the model (high-resolution) soil moisture pattern within each satellite footprint while preserving the mean remotely-sensed microwave emissivity ( ε ), which may differ significantly from the model mean over the footprint. To the extent that the emissivity-soil moisture relationship is linear, the neural network will also preserve the footprint-mean soil moisture. We approached the problem of disaggregation using a linear Artificial Neural Network (ANN). The ANN currently being tested is the simplest imaginable ANN, consisting of a single neuron. All of the inputs are weighted and then summed. The input to output mapping function is linear. Our approach was to train DisaggNet using soil moisture and emissivity output from the coupled SHEELS / Radiative Transfer Model. In so doing, the DisaggNet learns a ‘mapping’ from low (sensor) resolution ε to high (model) resolution soil moisture that is conservative in ε at the footprint scale and seeks to replicate the model patterns of soil moisture. We use ε instead of TB as input to eliminate the diurnal cycle caused by surface temperature variations. The accuracy of this relationship depends on how well SHEELS / RTM characterizes these sub-pixel scale patterns. Once DisaggNet is trained, this mapping can be applied to actual remotely-sensed observations. Because the mapping preserves the pixel-scale means, any large-scale errors in the model estimates will be ‘corrected’ via application of DisaggNet, based on our assumption that the remotely-sensed measurements are unbiased with respect to ground truth. DisaggNet has been trained with the following inputs:

  • Remotely sensed (low-resolution) emissivity with noise
  • Antecedent rainfall for several time intervals prior to current time: 0-1, 1-3, 3-6, etc.
  • Soil texture
  • Vegetation water content
  • Upstream contributing area (surface area draining into a grid cell)

Fig. 2. Application of DisaggNet using ESTAR microwave emissivities at 0.8 km grid and aggregated to the 1.6 km and 12.8 km. DisaggNet then attempts to reconstruct the original 0.8 km soil moisture pattern (bottom left panel). The SHEELS soil moisture pattern is shown for comparison; it is not used as model input, nor is it the benchmark.
Application of DisaggNet in the Little Washita River Watershed, OK, during the summer of 1997 is shown in figure 1. In this case, SHEELS / RTM emissivities at 0.8 km resolution were aggregated to 1.6 km and 12.8 km to simulate remotely-sensed microwave data. These aggregated emissivities, along with the other variables listed above, were used as DisaggNet inputs, and the resulting 0-5 cm soil moisture is shown below. The SHEELS fractional water content is shown as a benchmark. Figure 2 illustrates, for the same day, DisaggNet application using brightness temperature from ESTAR, an aircraft-borne sensor. In this case the benchmark is the high-resolution (0.8 km) ESTAR-estimated fractional water content. SHEELS soil moisture is shown as a reference. These results are very encouraging. Even when remote sensing data are input at 12.8 km resolution, the algorithm is capable of reconstructing a reasonably accurate soil moisture pattern at 0.8 km resolution.

Acknowledgements:
This research was supported by NASA through grant no. NCCW-0084 to Alabama A&M University , Center for Hydrology, Soil Climatology and Remote Sensing. This work was conducted in collaboration with Dr. Marius Schamschula, Alabama A&M University .

The following is a link to a book chapter that provides more details and application examples for DisaggNet:

Tsegaye, T.D., W.L. Crosson , C.A. Laymon, M.P. Schamschula and A.B. Johnson, 2003. Application of a neural network-based spatial disaggregation scheme for addressing scaling of soil moisture. In Scaling Methods in Soil Physics, Y. Pachepsky, D.E. Radcliffe and H.M. Selim, Eds., CRC Press, Boca Raton, FL, pp. 261-277.


Technical Contact: Dr. Bill Crosson (bill.crosson@msfc.nasa.gov)
Responsible Official: Dr. James L. Smoot (James.L.Smoot@nasa.gov)
Page Curator: Diane Samuelson (diane.samuelson@nasa.gov)