As aforementioned, ice particles have relatively low emissivity and high single scatteringalbedo, which makes snowfalls are only retrievable at higher frequencies. Also, in mostcases, the variability on surface type, water vapor, cloud vapor, makes the inversion of snowfall rate from microwave brightness temperature a ill posed problem. The accuracyof real case physical retrieval are particularly depend on the quality to represent the realworld snowfall of the pre-deﬁned model space. The quality of the forward simulation isquite another issue, which is out of the scope of this project. To overcome the modelingproblem, snowfall on ocean surface was selected for this study. The ocean surface isrelatively well known and can be modeled with relatively high accuracy in comparison tothat of land.There are large amount of snowfalls near Wakasabay near Japan in the winter time.The primary mechanism of the intensive, large scale snow fall is the heating of cold airfrom the continent by the relatively warm ocean surface. Large sets of historical data areavailable for both validation and research purpose. A numerical weather prediction modelWRF was applied to generate 3D distribution of snowfall near Wakasabay in 12/19/2003.A Monte-Carlo method by Petty  was adopted to generate the 2D brightness temper-ature at 18.7, 36.5 and 89.0 GHz, which is comparable to that of Adanced MicrowaveScanning Radiometer (AMSR). Real data set for AMSR in 12/25/2005 near Wakasabay,as well as the radar retrieved snowfall rate which were supposed to be the
for researcher in meteorology ﬁeld, were used as testing case. The overall pattern of thetwo precipitation process do not diﬀer too much, which facilitate the retrieval process bytraining NN to the model. For this radiometric application, a feed forward neural net-work are chosen , which has a layer of
observables and a hidden layer of neurons withtan-sigmoid transfer functions and
hidden layer with a linear transfer functions and1 output neuron for the snowfall rate estimation. The neural network was setup by thestandard Matlab package. Instead of the standard back-propagation, for a faster training,the Levenberg-Marquardt method algorithm was selected for training stage. Results anddiscussion are shown in following sections.
3 Results and Discussion
Diﬀerent neural network conﬁgurations are selected to tested the validity of the NNmethod used in these study. The ﬁrst experiment is the selection of feature vectors.Since the brightness temperatures are the combination of diﬀerent thermal sources, thefrequencies that are most related to the snowfall need to be determined so as to gain abetter result. Once the proper feature space is decided, the conﬁguation of the NN, suchas the number of the hidden layer, has to be tested.The simulated brightness temperature ﬁeld are shown in Fig. 1, and snowfall ratethrough NN are shown in Fig. 2. The result of the numerical experiment is presentedin Tab.1. Following features can be grasped from these experiments. Generally, largernumber of hidden layer can have a better prediction of the snowfall rate, e.g with the samefeature vectors , the NN with larger number of hidden layer neurons perform better thansmall number of hidden layer neurons. From the experiments, it is apparent that, thesnowfall rate is most correlated to high frequencies, adding higher frequency can generate3