Presentation on theme: "1 Backpropagation Neural Network for Soil Moisture Retrieval Using NAFE05 Data : A Comparison Of Different Training Algorithms Soo See Chai Department."— Presentation transcript:
1 Backpropagation Neural Network for Soil Moisture Retrieval Using NAFE05 Data : A Comparison Of Different Training Algorithms Soo See Chai Department of Spatial Sciences, Curtin University of Technology
2 CONTENT Neural Network and Soil Moisture Retrieval Backpropagation Neural Network Training of Neural Network Testing Results Q and A
3 Neural Network For Soil Moisture Retrieval Radiometric signatures of a vegetation- covered field reflect an integrated response of the soil and vegetation system to the observing microwave system Surface parameters and radiometric signatures :
5 Different Backpropagation Training Algorithms Several different training algorithms have a variety of different computation and storage requirements No one algorithm is best suited to all locations MATLAB : 11 different training algorithms Review : basic gradient descent and Levenberg-Marquardt(LM) algorithm How about the other algorithms ?
6 Data Preparation Roscommon area : 1/11, 8/11, 15/11 Determine the area coordinate : Roscommon : Top latitude : -32.15380 Bottom latitude : -32.18370 Left longitude : 150.120 Right longitude : 150.46900 MATLAB : cut the area, extract the fields in the PLMR file Copy the latitude, longitude, brightness temperature and altitude data into Excel Extract the aircraft altitude of medium resolution mapping which is around 1050m to 1270m ASL
9 Find the minimum and maximum of average Tb for each data set Next find the range (max-min) Find the width for each class ( 3 classes : training, validation and testing ) Range / 3 Find starting and ending point for each class A bit of Statistics …
10 We have now : Group 1 of date 1/11, 8/11 and 15/11 (combined : GRP1) Group 2 of date 1/11, 8/11, 15/11 (combined : GRP2) Group 3 of date 1/11, 8/11, 15/11(combined : GRP3) GRP 1 : randomly divide them into 3 groups : 60% for training, 30% for validation and 10% for testing Same with GRP2 and GRP3 All training data in one file, all validation data in one file, all testing data in one file
11 Training : K-Fold Cross Validation No. of data set is small, to get a better accuracy result, K-fold validation is used. Training data + validation data = 112 8-fold cross validation, each time 14 data will be used for validation, 98 data for training To make sure the data is random enough, each time the data will be randomized. Eg: First run : Second run : validationtraining validationtraining
12 Training :NN Parameters determination A series of experiments trial and error lowest RMSE If yes, then save the input weight, layer weight and bias for the NN to be used for the other training algorithms Fixed Layer : 3 layers ( 1 input, 1 hidden, 1 output ) Input : H polarized brightness temperature, TbH and physical soil temperature at 4cm Hidden : sigmoid function Output : linear function Soil moisture (%v/v)
13 Experiments carried out : Decision : Learning rate, lr = 0.005 Momentum, mc = 0.4 Input Weight, iw = W2.mat Layer weight, lw = LW2.mat Bias, b = B2.mat No. of hidden neuron = 4 No. of epochs = 200
15 Conclusions Different types of training algorithms of backpropagation NN is giving different but similar accuracy result The training data is representative of the testing data
16 Questions Is the NN architecture transferable ? Is number of data a factor contribute to the accuracy of the retrieval ? Adding ancillary data (beside soil temperature) : vegetation water content and land cover information help ? Adding V-polarized brightness temperature as an input ? Adding these data directly or let the NN account for these data ?