Effective porosity estimation using multiattribute analysis

In this study, an attempt is made to predict effective porosity in one of the oil fields in the Persian Gulf by designing a probablistic neural network (PNN) and simultanusely making use of seismic attributes and effective porosity logs in the reservoir window. This was done by deriving a multiattribute transformation between an optimum subset of seismic attributes and the effective porosity logs.
The geophysical data used in this study consist of 3D seismic pre-stack time migrated (PSTM) data with 12.5*12.5 m grid size and a 4 ms sampling rate. The length of the seismic traces are two seconds. Well logs of five vertical wells in the study area, including Sonic (DT), Density (RHOB), Effective Porosity (PHIE) and Seismic Well Velocity Surveys (Check Shots), were used. The reservoir layer is a Mishrif member of the Sarvak formation with Cretaceous age, which is common in oil reservoirs in the Persian Gulf. The top of the Mishrif is adjusted with the Middle Turonian Unconformity and covered with shaley Laffan formation. The Mishrif Reservoir in study area contains two reservoir zones. The lower zone with higher clay content is separate from the upper zone. The upper zone consists of clean limesone with better reservoir properties. Seismic traces close to the well locations were used to generate seismic attributes. Effective porosity logs at the reservoir area were the target logs in this study.
The designed neural network consists of one input layer, one hidden layer with four processing units (neuron), and one output layer with one neuron. In order to prepare training samples for the neural network, PHIE logs were converted to time domain using a time-depth relationship calculated from the DT logs and check shot curves for each well location. Subsequently, these logs were filtered (using a Hanning filter with 4 ms length) and resampled with  seismic sampling rate (4 ms). Finally, a set of seismic attributes, including sixteen sample-based seismic attributes, were generated using HRS software. Training samples in this study consisted of 57 samples (selected seismic attributes and their related effective porosity from PHIE logs in the time domain). For training the network, the samples were divided into three data sets: the training samples, cross validation samples and testing samples. The training data were used for adjusting the weights of the network; the cross validation data were used to prevent overtraining the neural network; and the testing data were used to ensure generalizabillity of the network output.
A forward stepwise regression process was used to determine an optimum subset of attributes for use in the training of the neural networks. The optimum subset of attributes in this study consists of the Dominant Frequency, Amplitude Weighted Frequency, Integrated Absolute Amplitude and Filter 45-60 Hz.
After the network was trained using training and cross validation data sets, it was used to predict the testing data. The results show a good correlation between real and predicted data, with 92% correlation. Finally, in order to attain a better generalization of the network, testing data sets were inserted to trained data and the network was trained again. This network was then used to predict effective porosity in well locations which increased the correlation coefficient to 95%. This study shows the ability of the PNN networks to predict effective porosity even with a paucity of training examplares.