Agus Buono

A TIME-DELAY CASCADING NEURAL NETWORK ARCHITECTURE FOR MODELING TIME-DEPENDENT PREDICTOR IN ONSET PREDICTION

The occurrence of rain before the real start of a r ainy season often mislead farmers into thinking tha t rainy season has started and suggesting them to start pla nting immediately. In reality, rainy season has not started yet, causing the already-planted rice seed to exper ience dehydration. Therefore, a model that can pred ict the onset of rainy season is required, so that draught disaster can be avoided. This study presents Time D elayCascading Neural Network (TD-CNN) which deals with situations where the response variable is determined by a number of time-dependent inter-rela ted predictors. The proposed model is used to predi ct the onset in Pacitan District Indonesia based on So uthern Oscillation Index (SOI). The Leave One Out (LOO) cross-validation with series data 1982-2012 a re used in order to compare the accuracy of the proposed model with the Back-Propagation Neural Network (BPNN) and Cascading Neural Network (CNN). The experiment shows that the accuracy of the proposed model is 0.74, slightly above than the t wo other models, BPNN and CNN which are 0.71 and 0.72, respectively.

Multiscale fractal dimension modelling on leaf venation topology pattern of Indonesian medicinal plants

This research proposed a new model to differentiate leaf venation topology patterns using Multiscale Fractal Dimension. Identification of medicinal plants is important considering wide range of biodiversity in Indonesia and significant role of medicinal plants in Indonesia. Plants identification can be performed with shape analysis using plant leaf venation as a feature. Multiscale Fractal Dimension is a shape analysis method that analyze shapes through its complexity. In this research three Indonesian medicinal plants species has their leaf venation topologies modelled with Multiscale Fractal Dimension. The result shows that while the difference is not remarkably clear, there are irregularities that can be made more evident with multiscale analysis. Future works can include Multiscale Fractal Dimension as one technique to identify plants.

A Bayesian network approach for image similarity

This paper proposed Bayesian Network approach for image similarity measurement based on color, shape and texture. Bayesian network model can determine dominant information of an image using occurrence probability of images characteristics. This probability is used to measure image similarity. Performance of the system is determined using recall and precision. Based on experiment, Bayesian network model can improve performance of image retrieval system. Experiment result showed that the average precision gain up of using Bayesian network model is about 8.28 %. The average precision of using Bayesian network model is better than using color, shape, or texture information individually.

A monsoon onset and offset prediction model using backpropagation and moron method: A case in drought region

First day (onset) and last day (offset) of monsoon are nature phenomena which are important elements at cultivation stages in agriculture. These 2 sets of time value influent harvest performance but it is difficult to predict onset and offset at drought region. One of technique that can be used to solve mentioned problem is prediction technique which is one of data mining task. In this research, Feed Forward Backpropagation (BPNN) were combined with Moron method to predict onset and offset at drought region. Data used were daily rainfall data from 1983 to 2013. This experiment used 2 kind of BPNN models and they used S different values for learning rate (alpha) from range 0.01 to 0.2. Root Mean Square Error (RMSE) is used to evaluate resulted prediction models along with correlation value and standard deviation of error for better understanding. For BPNN onset model, lowest RMSE value at alpha 0.15 is 32,0546 and lowest RMSE value for BPNN offset is 26,6977 at alpha 0.05. Developed model has been able to use for prediction, but the result was still not close enough to actual data. In order to achieve a better model with lower RMSE, it is neccesary to improve model architecture and to specify some methods to obtain certain number of input layer based on Southern Oscillation Index (SOI) data.

A comparison of backpropagation and LVQ: A case study of lung sound recognition

One way to evaluate the state of the lungs is by listening to breath sounds using stethoscope. This technique is known as auscultation. This technique is fairly simple and inexpensive, but it has some disadvantage. They are the results of subjective analysis, human hearing is less sensitive to low frequency, environmental noise and pattern of lung sounds that almost similar. Because of these factors, misdiagnosis can occur if procedure of auscultation is not done properly. In this research, will be made a model of lung sound recognition with neural network approach. Artificial neural network method used is Backpropagation (BP) and learning Vector Quantization (LVQ). Comparison of these two methods performed to determine and recommend algorithms which provide better recognition accuracy of speech recognition in the case of lung sounds. In addition to the above two methods, the method of Mel Frequency Cepstrum Coefficient (MFCC) is also used as method of feature extraction. The results show the accuracy of using Backpropagation is 93.17%, while the value of using the LVQ is 86.88%. It can be concluded that the introduction of lung sounds using Backpropagation method gives better performance compared to the LVQ method for speech recognition cases of lung sounds.

Forecasting the length of the rainy season using time delay neural network

Indonesia has abundant natural resources in agriculture. Good agricultural results can be obtained by determining a good growing season plan. One of important factors which determines the successful of crop is the length of the rainy season. The length of the rainy season is dynamic and difficult to be controlled. Therefore the planning of the growing season becomes inaccurate and cause crop failures. This research aims to develop a model to predict the length of the rainy season using time delay neural network (TDNN). Observational data used in this research is the length of rainy season from three weather and climate stations of the Pacitan region from 1982/1983 to 2011/2012. Predictor data used in this reserach is sea surface temperature (SST) derived from the region of Nino 1+2, Nino 3, Nino 4, and Nino 3.4 from 1982 to 2011. Model with the best accuracy was obtained by Pringkuku station with RMSE of 1.97 with parameters of delay [0 12 3], learning rate 0.1, 40 hidden neurons, and predictors of Nino 3 and R-squared of 0.82 with parameters of delay [0 1], learning rate 0.3, 5 hidden neurons, and predictors of Nino 3.

Clustering metagenome fragments using growing self organizing map

The microorganism samples taken directly from environment are not easy to assemble because they contains mixtures of microorganism. If sample complexity is very high and comes from highly diverse environment, the difficulty of assembling DNA sequences is increasing since the interspecies chimeras can happen. To avoid this problem, in this research, we proposed binning based on composition using unsupervised learning. We employed trinucleotide and tetranucleotide frequency as features and GSOM algorithm as clustering method. GSOM was implemented to map features into high dimension feature space. We tested our method using small microbial community dataset. The quality of cluster was evaluated based on the following parameters : topographic error, quantization error, and error percentage. The evaluation results show that the best cluster can be obtained using GSOM and tetranucleotide.