Kyaw Kyaw Htike

Rainfall forecasting models using focused time-delay neural networks

Rainfall forecasting is vital for making important decisions and performing strategic planning in agriculture-dependent countries. Despite its importance, statistical rainfall forecasting, especially for long-term, has been proven to be a great challenge due to the dynamic nature of climate phenomena and random fluctuations involved in the process. Artificial Neural Networks (ANNs) have recently become very popular and they are one of the most widely used forecasting models that have enjoyed fruitful applications for forecasting purposes in many domains of engineering and computer science. The main contribution of this research is in the design, implementation and comparison of rainfall forecasting models using Focused Time-Delay Neural Networks (FTDNN). The optimal parameters of the neural network architectures were obtained from experiments while networks were trained to perform one-step-ahead predictions. The daily rainfall dataset, obtained from Malaysia Meteorological Department (MMD), was converted to monthly, biannually, quarterly and monthly datasets. Training and testing were performed on each of the datasets and corresponding accuracies of the forecasts were measured using Mean Absolute Percent Error. For testing data, results indicate that yearly rainfall dataset gives the most accurate forecasts (94.25%). As future work, more parameters such as temperature, humidity and sunshine data can be incorporated into the neural network for superior forecasting performance.

Comparison of supervised and unsupervised learning classifiers for human posture recognition

Human posture recognition is gaining increasing attention in the fields of artificial intelligence and computer vision due to its promising applications in the areas of personal health care, environmental awareness, human-computer-interaction and surveillance systems. Human posture recognition in video sequences is a challenging task which is part of the more comprehensive problem of video sequence interpretation. In this paper, an intelligent human posture recognition system in video sequences is proposed. Firstly, the system was trained and evaluated to classify five different human postures using both supervised and unsupervised learning classifiers. The supervised classifier used was Multilayer Perceptron Feedforward Neural Networks (MLP) whilst for unsupervised learning classifiers, Self Organizing Maps (SOM), Fuzzy C Means (FCM) and K Means have been employed. Results indicate that MLP performs (96% accuracy) much better than SOMs, FCM and K Means which give accuracies of 86%, 33% and 31% respectively. Secondly, all the classifiers were then trained and evaluated again to classify two postures. With only 2 postures, the accuracies of all the classifiers have increased dramatically, especially for unsupervised classifiers. This shows that supervised learning classifiers are superior to unsupervised ones for the task of human posture recognition and that the unsupervised classifiers do not learn very well for cases where a lot of postures have to be learnt as compared to the supervised learning classifier which gives high accuracy in either case.

A non-contact capacitance type level transducer for liquid characterization

The liquid properties such as buoyancy, pressure at a depth, relative electrical permittivity, electrical conductivity, thermal conductivity, absorption of radiation, liquid surface reflection of sound or light waves, are used to design the different types of liquid level transducers for liquid level measurement in any process industry. The contact-type level-sensing transducers have the disadvantage that their characteristic properties may change due to physical or chemical reaction between the liquid and the probing material, and hence may affect accuracy besides their life time. The non-contact-type level-sensing probes may have longer life period, but they are comparatively costly and require various environmental and experimental precautionary measures when being used. In this paper, a low-cost non-contact capacitance type liquid level measuring technique has been designed and used for liquid identification objectives. The results obtained are very much consistent with the theoretical derivations.

Human posture recognition and classification

Human posture recognition is gaining increasing attention in the field of computer vision due to its promising applications in the areas of personal health care, environmental awareness, human-computer-interaction and surveillance systems. Human posture recognition in video sequences is a challenging task which is part of the more general problem of video sequence interpretation. This paper presents a novel an intelligent human posture recognition system for video surveillance using a single static camera. The training and testing were performed using four different classifiers. The recognition rates (accuracies) of those classifiers were then compared and results indicate that MLP gives the highest recognition rate. Moreover, results show that supervised learning classifiers tend to perform better than unsupervised classifiers for the case of human posture recognition. Furthermore, for each individual classifier, the recognition rate has been found to be proportional to the number of postures trained and evaluated. Performance comparisons between the proposed systems and existing systems were also carried out.

Forests of unstable hierarchical clusters for pattern classification

Classification of patterns is a key ability shared by intelligent systems. One of the crucial components of a pattern classification pipeline is the classifier. There have been many classifiers that have been proposed in literature, and it has been shown recently that ensembles of decisions trees tend to perform and generalize well to unseen test data. In this paper, we propose a novel ensemble classifier that consists of a diverse group of hierarchical clusterings on data. The proposed algorithm is fast to train, fully automatic and outperforms existing decision tree ensemble techniques and other state-of-the-art classifiers. We empirically show the effectiveness of the algorithm by evaluating on four publicly available datasets.