Master of Science By Research

Permanent URI for this collectionhttp://192.248.9.226/handle/123/15770

Browse

Now showing 1 - 20 of 55

item: Thesis-Abstract
Self-supercised learning in gender classification using full-body images extracted from CCTV footage
(2023) Rajalingam, G; Ambegoda, T
Gender classification is regarded as one of the vital components of security systems, recommendation systems, data access authentication and surveillance. Facial features and supervised learning remain the predominant metrics to classify genders currently. But facial feature driven approach would falter in case of incomplete or unavailable details especially when analyzing masked faces or CCTV footage and supervised learning driven approach becomes tedious and time-consuming provided large volume of labelled data. Therefore, the need of analyzing full-body images is established instead of the sole focus of facial features driven analysis as well as the less dependency on supervised learning. The proposed approach establishes the implementation of convolutional neural network (CNN) based on self-supervised learning classification algorithm that needs fewer volumes of labelled data for fine-tuning. dBOT classifier, a state-of-the-art self-supervised image classification model, is used to perform transfer learning and the subsequential fine-tuning to facilitate the training on low-quality images. The proposed model on evaluation significantly outperforms SSL based methods for small, unclear full-body gender image classification techniques applied on CCTV footage extracts. Keywords: CNN, dBOT, Gender-Classification, CCTV
item: Thesis-Full-text
Developing a retrieval-based Tamil language chatbot for closed domain
(2023) Kugathasan, K; Thayasivam, U
Chatbots are conversational systems that interact with humans via natural language. Frequently, it is used to respond to user queries and provide them with the information they need. To build a highly functional chatbot, a good corpus and a variety of language-related resources are required. Since Tamil is a low-resource language those resources are not available for Tamil. Additionally, since Tamil is also a morphologically rich language, high inflexion and free word order pose key challenges to Tamil chatbots. Due to all the above reasons, it is evident that developing an effective End-to-End chat system is challenging even for a closed domain. This study introduces a novel method for building a chatbot in Tamil by leveraging a dataset extracted from Tamil banking website’s FAQ sections and extending it to encompass the language's morphological complexity and rich inflectional structure. Intent is assigned to each query, and a multiclass intent classifier is developed to classify user intent. The CNN-based classifier demonstrated the highest performance, achieving an accuracy of 98.72%. While previous works on short-text classification in Tamil focused only on a few classes and used a very large dataset, our method produced a superior accuracy of over 98% using a small number of per-class examples even when there are 56 classes and additional challenges like class imbalance problem in the data. This shows our approach is better than any other approach for short text classification in Tamil. The major contribution of this research is the generation of the first-ever chat dataset for Tamil. Our research is the first of its kind in Tamil to show how an efficient context-less chatbot can be built using short text classification. Although this project is done for the Tamil language and for the Banking domain, this approach can be applied to other low-resourced languages and domains as well.
item: Thesis-Abstract
Exploiting multilingual contextual embeddings for Sinhala text classification
(2022) Dhananjaya, GV; Ranathunga, S; Jayasena, S
Language models that produce contextual representations (or embeddings) for text have been commonly used in Natural Language Processing (NLP) applications. Partic ularly, Transformer based, large pre-trained models are popular among NLP practition ers. Nevertheless, the existing research and the inclusion of low-resource languages (languages that primarily lack publicly available datasets and curated corpora) in these modem NLP paradigms are meager. Their performance for downstream NLP tasks lags compared to that of high-resource languages such as English. Training a mono lingual Language model for a particular language is a straightforward approach in modem NLP but it is resource-consuming and could be unworkable for a low-resource language where even monolingual training data is insufficient. Multilingual models that can support an array of languages are an alternative to circumvent this issue. Yet, the representation of low-resource languages considerably lags in multilingual models as well. In this work, our first aim is on evaluating the performance of existing Multilingual Language Models (MMLM) that support low-resource Sinhala and some available monolingual Sinhala models for an array of different text classification tasks. We also train our own monolingual model for Sinhala. From those experiments, we identify that the multilingual XLM-R model yields better results in many instances. Based on those results we propose a novel technique based on an explicit cross-lingual alignment of sentiment words using an augmentation method to improve the sentiment classifica tion task. There, we improve the results of a multilingual XLM-R model for sentiment classification in Sinhala language. Along the way, we also test the aforementioned method on a few other lndic languages (Tamil, Bengali) to measure its robustness across languages. Keywords: Multilingual language models, Multilingual embeddings, Text classification, Sen timent analysis, Low-resource languages, Sinhala language
item: Thesis-Abstract
High-performance 3D mapping of unknown environments using parallel computing for mobile robots
(2021) De Silva KTDS; Gamaga GD; Sooriyaarachchi SJ
Autonomous multi-robot systems are a popular research field in the 3D mapping of unknown environments. High fault tolerance, increased accuracy, and low latency in coverage are the main reasons why a multi-robot system is preferred over a single robot in an unpredictable field. Compared with 3D scene reconstruction which is a conceptually similar but resource-wise different technique, autonomous mobile robot 3D mapping techniques are missing a crucial element. Since most mobile robots run on low computationally powered processing units, the real-time registration of point clouds into high-resolution 3D occupancy grid maps is a challenge. Until recently, it was nearly impossible to perform parallel point cloud registration in mobile platforms. Serial processing of a large amount of high-frequency input data leads to buffer overflows and failure to include all information into the 3D map. With the introduction of Graphical Processing Units (GPUs) into commodity hardware, mobile robot 3D mapping now can achieve faster time performance, using the same algorithmic techniques as 3D scene reconstruction. However, parallelization of mobile robot 3D occupancy grid mapping process is a less frequently discussed topic. As a Central Processing Unit (CPU) is necessary to run conventional middleware, operating system, and hardware drivers, the system is developed as a CPU-GPU mixed pipeline. The precomputed free scan mask is used to accelerate the process of identifying free voxels in space. Point positional information is transformed into unsinged integer coordinates to cope with Morton codes, which is a linear representation of octree nodes instead of traditional spatial octrees. 64-bit M-codes and 32-bit RGBO-codes are stored in a hash table to reduce access time compared to a hierarchical octree. Point cloud transformation, ray tracing, mapping point coordinated into integer scale, Morton-coded voxel generation, RGBO-code generation are the processes that are performed inside the GPU. Retrieving point cloud information, map update using bitwise operations and map publish are executed within the CPU. Additionally, a multi-robot system is prototyped as a team of wheeled robots autonomously exploring an unknown, even-surfaced environment, while building and merging fast 3D occupancy grid maps and communicating using a multi-master communication protocol.
item: Thesis-Abstract
An Automated framework for precipitation-related decision making
(2021) Adikari AMHD; Bandara HMND; Chitraranjan C
With the effects of rapid urbanization and climate change, weather forecasting plays a vital role in disaster risk controlling and mitigation activities. Generating weather forecasts using numerical weather prediction methods is a tedious process as it requires multiple combinations of weather model workflows. Due to the weather’s chaotic nature, field experts frequently modify these static workflows to cater to their decisionmaking requirements. Moreover, infrequent weather events make these workflows more complicated and challenging to handle manually. There is a need for a decision support system (DSS) to build and update workflows dynamically with these circumstances. After studying the architectures of existing DSSs from different fields, we understand that they cannot handle all weather-related decision-making requirements. Therefore, we present a generic decision support system framework to create and control complex and dynamic weather model workflows. The proposed framework supports three types of decision-making conditions: accuracy-based, infrequent-event, and pump/gate control. The framework can terminate or dynamically update weather model workflow paths in accuracy-based decisions. In infrequentevent decisions, the framework identifies the weather events. In pump-control decisions, the framework attempts to find an optimized control strategy that minimizes the flood risk in the given catchment area. In addition to the above decisions, the system provides relevant workflow strategies for handling unexpected weather conditions. To demonstrate the utility of accuracy-based decision-making, we executed four workflow runs over six months (on randomly selected days for each month). We were able to achieve a 100% accuracy level with manual verification. Pump/gate control strategies were tested using the data from the 2010 Flood event in the Kelani basin area in Sri Lanka. Pump strategy decisions also had 100% accuracy on logic evaluation and model selection. The proposed framework is deployed in a Google cloud platform of the Center for Urban Water, Sri Lanka, for flood -related forecasts.
item: Thesis-Abstract
Computational model for glaucoma classification
(2023) Shyamalee KWT; Meedeniya DA
Glaucoma is a leading cause of blindness and affects millions of people worldwide. It is a chronic eye condition that damages the optic nerve and if left untreated, it can lead to vision loss and decreased quality of life. According to the World Health Organization, it affects approximately 65 million people worldwide. Thus, there is a requirement for an effective and reliable mechanism for the identification of Glaucoma. This study addresses a computational model for the Glaucoma identification process. The proposed system uses fundus images of the eye. The availability of computing resources and automated glaucoma diagnosis tools can now be supported due to recent developments in DL. Generic Convolutional Neural Networks (CNN) is still not frequently used in medical situations despite the advances made by deep learning in disease diagnosis using medical images. This is because of the limited trustworthiness of these models. Despite the rise in popularity of deep learning-based glaucoma classification in recent years, few studies have focused on the models’ explainability and interpretability, which boosts user confidence in such applications. To predict glaucoma conditions, this study uses state-of-the-art deep learning techniques to segment and classify fundus images. To make the results more understandable, visualization techniques are used to present the findings. Our forecasts are based on a modified InceptionV3 architecture and a U-Net with attention mechanisms. Additionally, us- ing Gradient-weighted Class Activation Mapping (Grad-CAM) and Grad-CAM++, we create heatmaps that show the areas that had an impact on the glaucoma diagnosis. With the RIM-ONE dataset, our findings demonstrate the best accuracy, sensitivity, and specificity values of 98.97%, 99.42%, and 95.59%, respectively. With the aid of fundus images, this model can be used to support automated glaucoma diagnosis.
item: Thesis-Full-text
Flexible and extensible infrastructure monitoring architecture for computing grids with infrastructure aware job matching
(2023) Wijethunga RMKD; Perera I; Amarasinghe G
Many research experiments with large data processing requirements rely on massive, distributed Computing Grids for their computational requirements. A Computing Grid is built by combining a large number of individual computing sites distributed globally. These Grid sites are maintained by different institutions across the world and contribute thousands of worker nodes possessing different capabilities and configurations. Developing software for Grid operations that works on all nodes while harnessing the maximum capabilities offered by any given Grid site is challenging without knowing what capabilities each site offers in advance. This research focuses on developing an architecture-independent Grid infrastructure monitoring design to monitor the infrastructure capabilities and configurations of worker nodes at sites across a Computing Grid without the need to contact local site administrators. The design presents a highly flexible and extensible architecture that offers infrastructure metric collection without local agent installations at Grid sites. The resulting design is used to implement a Grid infrastructure monitoring framework called “Site Sonar v2.0” that is currently being used to monitor the infrastructure of 7,000+ worker nodes across 60+ Grid sites in the ALICE Computing Grid. The proposed design is then used to introduce an improved Job matching architecture for Computing Grids that allows job matching based on any infrastructure property of the worker nodes. This dissertation introduces the proposed architecture for a highly flexible and extensible Grid infrastructure monitoring design and an improved job design for Computing Grids and the implementation of those designs to derive important findings about the infrastructure of ALICE Computing Grid while improving its job matching capabilities. This work provides a significant contribution to the development of distributed Computing Grids, particularly in terms of providing a more efficient and effective way to monitor infrastructure and match jobs to worker nodes.
item: Thesis-Abstract
An Adaptive software architectural framework for an interactive learning toolkit
(2023) Jayasiriwardene DPS; Meedeniya D
At present, a significant demand has emerged for online education tools that can used as a replacement for classroom education. Due to the ease of access and the highavailability of mobile devices, the preference of many users is focused on m-learningapplications. Thus, this study presents an adaptive software architectural framewfor an interactive learning toolkit. As a case study, the application is applied toprimary education sector in Sri Lanka, as there is a lack of learning tools that allteachers and students to interact effectively. Accordingly, a software architecturaframework was designed with the features of adaptivity, learning content authoring, learning content management, low resource utilization, and low power consumptiThe study includes an extensive literature review conducted to identify unique gapsexisting studies. Further, the study designs and develops an architecture with intended feature effectively embedded in it. Furthermore, an m-learning applicanamed “iLearn” is developed as a proof-of-concept by implementing the architecturdesign. Moreover, the prototype was evaluated for functional requirements successfully conducting unit tests and user interface tests. The non-functionarequirements of the application were evaluated by conducting a system usabilsurvey for 20 teachers and 20 students, which received a good usability score of 80.5%and 83.6%, respectively. Also, the performance of the application was tested received a good overall outlook on performance where it was found that the applicahas a below-average consumption of memory, CPU, and battery at peak performancThe application is concluded as a success, with the potential to enhance with cuttingedge technology.
item: Thesis-Full-text
Pre-training and fine-tuning multilingual sequence-to-sequence models for domain-specific low-resource neural machine translation
(2022) Thillainathan S; Jayasena,S; Ranathunga S
Limited parallel data is a major bottleneck for morphologically rich Low-Resource Languages (LRLs), resulting in Neural Machine Translation (NMT) systems of poor quality. Language representation learning in a self-supervised sequence-to-sequence fashion has become a new paradigm that utilizes the largely available monolingual data and alleviates the parallel data scarcity issue in NMT. The language pairs supported by the Self-supervised Multilingual Sequence-to-sequence Pre-trained (SMSP) model can be fine-tuned using this pre-trained model with a small amount of parallel data. This study shows the viability of fine-tuning such SMSP models for an extremely low-resource domain-specific NMT setting. We choose one such pre-trained model: mBART. We are the first to implement and demonstrate the viability of non-English centric complete fine-tuning on SMSP models. To demonstrate, we select Sinhala, Tamil and English languages in extremely lowresource settings in the domain of official government documents. This research explores the ways to extend SMSP models to adapt to new domains and improve the fine-tuning process of SMSP models to obtain a high-quality translation in an extremely lowresource setting. We propose two novel approaches: (1) Continual Pre-training of the SMSP model in a self-supervised manner with domain-specific monolingual data to incorporate new domains and (2) multistage fine-tuning of the SMSP model with in- and out-domain parallel data. Our experiments with Sinhala (Si), Tamil (Ta) and English (En) show that directly fine-tuning (single-step) the SMSP model mBART for LRLs significantly outperforms state-of-the-art Transformer based NMT models in all language pairs in all six bilingual directions. We gain a +7.17 BLEU score on Si→En translation and a +6.74 BLEU score for the Ta→En direction. Most importantly, for non-English centric Si-Ta fine-tuning, we surpassed the state-of-the-art Transformer based NMT model by gaining a +4.11 BLEU score on Ta→Si and a +2.78 BLEU score on Si→Ta. Moreover, our proposed approaches improved performance strongly by around a +1 BLEU score compared to the strong single-step direct mBART fine-tuning for all six directions. At last, we propose a multi-model ensemble that improved the performance in all the cases where we obtained the overall best model with a +2 BLEU score improvement.
item: Thesis-Full-text
Minimizing domain bias when adapting sentiment analysis techniques to the legal domain
(2022) Ratnayaka G; Perera AS; De Silva N
Sentiment Analysis can be considered as an integral part of Natural Language Processing with a wide variety of significant use cases related to different application domains. Analyzing sentiments of descriptions that are given in Legal Opinion Texts has the potential to be applied in several legal information extraction tasks such as predicting the judgement of a legal case, predicting the winning party of a legal case, and identifying contradictory opinions and statements. However, the lack of annotated datasets for legal sentiment analysis imposes a major challenge when developing automatic approaches for legal sentiment analysis using supervised learning. In this work, we demonstrate an effective approach to develop reliable sentiment annotators for legal domain while utilizing a minimum number of resources. In that regard, we made use of domain adaptation techniques based on transfer learning, where a dataset from a high resource source domain is adapted to the target domain (legal opinion text domain). In this work, we have come up with a novel approach based on domain specific word representations to minimize the drawbacks that can be caused due to the differences in language semantics between the source and target domains when adapting a dataset from a source domain to a target domain. This novel approach is based on the observations that were derived using several word representational and language modelling techniques that were trained using legal domain specific copora. In order to evaluate different word representational techniques in the legal domain, we have prepared a legal domain specific context based verb similarity dataset named LeCoVe . The experiments carried out within this research work demonstrate that our approach to develop sentiment annotators for legal domain in a low resource setting is successful with promising results and significant improvements over existing works.
item: Thesis-Abstract
Automatic classification of multiple acoustic events using artificial neural networks
(2021) Egodage D; Sooriyaarachchi S
There are numerous scenarios where similar acoustic events occur multiple times. Acoustic monitoring of migratory birds is an ideal example. Birds make a type of call known as flight calls during migration. A flight call can be considered as an acoustic event because it is a short-term, intuitively distinct sound. It is challenging to identify multiple occurrences of extremely short-range acoustic events such as flight calls in real-world recordings using classification techniques that require more computational power. It is mainly due to background noise and complex acoustic environments. This research aims at developing a classification model that reduces the effect of background noise, extract ROIs from continuous recordings, extract suitable features of flight calls and detect multiple occurrences of flight calls. An improved algorithm that can extract features has been developed in this research—by combining a well known Maximally Stable Extremal Regions (MSER) technique with state of the art traditional techniques. Namely Spectral and Temporal Features(SATF) and a combination of SATF and Spectrogram-based Image Frequency Statistics(SIFS). We name this novel algorithm as Spectrogram-based Maximally Stable Extremal Regions (SMSER). Three distinct feature sets have formed such that Featureset-1 created using SATF. Featureset-2 is a blend of SATF and SIFS. Featureset-3 is a combination of SATF, SIFS, and SMSER. The kNN, RF, SVM, and DNN classification techniques evaluated a real-world dataset using the extracted feature sets. Research carried out several tests to find out the best performing combination of classification model and feature set. The results showed that the flight calls’ detection accuracy increased when the number of features increased, although high computational power requirement is a disadvantage. The performance of SMSER feature set was the best among almost every classification technique above. It should be because the SMSER Feature set has the highest number of features. Classification of the SMSER feature set from the DNN classifier showed the highest accuracy of 87.67%.
item: Thesis-Abstract
Optimization of RSSI based indoor localization and tracking using machine learning techniques
(2021) Aravinda SPP; Gamage CD; Sooriyaarchchi SJ; Kottege N
Localization and tracking of persons in industrial environment is critical in terms of safety, privacy and security, particularly when there are hazardous zones. In this research, RSSI of RF signals were used to localize, track and uniquely identify a person in a cluttered environment with a case study into a doorway from a safe zone to a hazardous zone in a cluttered warehouse. Vision based localization was impractical both due to visual obstruction by moving large objects and privacy issues. There were three approaches in RF based localization reviewed in this work.This research uses the approach in which RF receivers are fixed and the transmitter is worn by the target person. RSSI data in a doorway area of 420 cm × 450 cm was analysed both in simulation and in a real test bed and it was proved that DNN and RNN based location prediction was feasible with an accuracy of over 80% even though the environment had noise in the range of ±2 dB to ±15 dB and ±7 dB on average for RF signals. The experiments carried out with a test bed consisting of Raspberry Pi-3 as receivers and Kontakt-io Tough Beacon TB15-1 module as transmitter connected over POE module to a centralized server. The results show that a bounded type RF receiver arrangement to cover the whole area with at least few receivers mounted at a high elevation to capture line of sight signals was effective in accurately localizing the person. The density of positions at which the RSSI data is collected to train the DNN also considerably affected the localization accuracy. The body attenuation was found to be another critical factor affecting the localization accuracy. When the DNN was trained with data captured at one orientation of the person, this DNN was successful in localizing a person with the same orientation but not in localizing a person in completely different orientations. This behaviour was used to detect the body orientation of a person using multiple neural network. A straight path traversed by a walking person at an average speed of 25 𝑐𝑚/𝑠 was successfully tracked at a point-wise accuracy over 80% using time series RSSI data with a threshold of 25 cm. The threshold was reduced to half by averaging the data over three consecutive predicted positions in the form a centroid. Lastly, Timedomain based RSSI data were used to train RNNs. Deep-LSTM model showed around 95% path-wise localization accuracy for constructed walking paths. Also, RNNs were able to detect the walking direction in single RNN network compared to multiple DNN approach. Finally, this research was able to uniquely identify, localize, detect body orientation and track the walking path of a person and since the person is uniquely identified and RSSI data is MAC addressed this work can be extended to localization of multiple persons.
item: Thesis-Full-text
Using back-translation to improve domain-specific English-Sinhala neural machine translation
(2021) Epaliyana K; Ranathunga S; Jayasena S
Machine Translation (MT) is the automatic conversion of text in one language to other languages. Neural Machine Translation (NMT) is the state-of-the-art MT technique w builds an end-to-end neural model that generates an output sentence in a target language given a sentence in the source language as the input. NMT requires abundant parallel data to achieve good results. For low-resource settings such as Sinhala-English where parallel data is scarce, NMT tends to give sub-optimal results. This is severe when the translation is domain-specific. One solution for the data scarcity problem is data augmentation. To augment the parallel data for low-resource language pairs, commonly available large monolingual corpora can be used. A popular data augmentation technique is Back-Translation (BT). Over the years, there have been many techniques to improve vanilla BT. Prominent ones are Iterative BT, Filtering, Data Selection, and Tagged BT. Since these techniques have been rarely used on an inordinately low-resource language pair like Sinhala - English, we employ these techniques on this language pair for domain-specific translations in pursuance of improving the performance of Back-Translation. In particular, we move forward from previous research and show that by combining these different techniques, an even better result can be obtained. In addition to the aforementioned approaches, we also conducted an empirical evaluation of sentence embedding techniques (LASER, LaBSE, and FastText+VecMap) for the Sinhala-English language pair. Our best model provided a +3.24 BLEU score gain over the Baseline NMT model and a +2.17 BLEU score gain over the vanilla BT model for Sinhala → English translation. Furthermore, a +1.26 BLEU score gain over the Baseline NMT model and a +2.93 BLEU score gain over the vanilla BT model were observed for the best model for English → Sinhala translation.
item: Thesis-Abstract
Acoustic event detection in polyphonic environments using artificial neural networks
(2021) Mihiranga JPM; Sooriyaarachchi S
Our environment is a mixture of hundreds of sounds that are emitted by different sound sources. These sounds are overlapped in both time and frequency domains in an unstructured manner composing a polyphonic environment. Identification of acoustic events in a polyphonic environment has become an emerging topic with many applications such as surveillance, context-aware computing, automatic audio indexing, health care monitoring and bioacoustics monitoring. Polyphonic acoustic event detection is a challenging task aimed at detecting the presence of multiple sound events that are overlapped at a particular time instance and labeling. It requires a large amount of training data with a complex machine learning architecture thus making it a highly resource-consuming task. Hence, the accuracy of this research area is still not at a satisfactory level. This study presents a neural networks-based classifier architecture with data augmentation and post-processing methods to improve accuracy. Two neural network architectures as a multi-label and combined single label are implemented and compared in the study. Previous literature reveals that Mel frequency cepstral coefficients and log Mel-band energies are the widely used features in the state of the art research in the area. Different data augmentation methods were used to ensure that the neural networks are trained for even the slight variations of the environmental sounds. A novel binarization method based on the signal energy is proposed to calculate the threshold value for binarizing the source presence predictions. Finally, the median filter based post processing was implemented to smoothen the detection results. The experimental results show that the proposed binarizing method improved the detection accuracy and recorded a maximum of 62.5% combined with the data augmentation and post-processing.
item: Thesis-Abstract
Weather data integration and assimilation system
(2021) karunarathne HMGC; Bandara HMND
Numerical Weather Models (NWMs) utilize data collected via diverse sources such as automated weather stations, radars, air balloons, and satellite images. Before using such multimodal data in a NWM, it is necessary to transcode data into a format ingested by the NWM. Moreover, the data integration system’s response time needs to be relatively low to forecast and monitor timesensitive weather events like hurricanes, storms, and ﬂash ﬂoods that require rapid and frequent execution of NWMs. The resulting weather data also need to be accessed by many researchers and third-party applications such as logistic and agricultural insurance ﬁrms. Existing weather data integration systems are based on monolithic or client-server architectures; hence, unable to beneﬁt from novel computational models such as cloud computing and containerized applications. Moreover, most of these softwares are proprietary or closed-source, making it difﬁcult to customize them for an island like Sri Lanka with different weather seasons. Therefore, in this research, we propose Weather Data Integration and Assimilation System (WDIAS) that utilizes microservices to achieve scalability, high availability, and low-cost operation based on cloud computing. The use of stateless microservices also enables WDIAS to add new features on the ﬂy with rollover capabilities. Moreover, WDIAS provides a modular framework to integrate data from different sources, export into different formats, and add new functionality by adding extension modules. We demonstrate the utility of WDIAS using a cloud-based experimental setup and weather-related synthetic workloads.
item: Thesis-Full-text
Forecasting agricultural crop yield variations using big data and supervised machine learning
(2020) Lakmal KJTD; Nanayakkara V
The government of Sri Lanka is struggling to make appropriate policy decisions regarding paddy cultivation due to absence of accurate and timely data to estimate the paddy yield, land usage for paddy cultivation and area affected by various paddy diseases. Remote sensing data based machine learning implementations can be identified as a potential solution for the above issue, as remote sensing data can be used for accurate and timely estimations. However, the traditional remote sensing data resources have failed to generate accurate estimates regarding cultivated paddy extent estimations. In this study, novel optical remote sensing data resources and a hybrid approach are employed to mitigate previously reported issues. Furthermore, a multi-temporal approach is used instead of traditional mono-temporal approach by leveraging deep neural networks. This study also consists of a comprehensive comparison on novel optical remote sensing data resources and the evaluations of the capability of using deep neural networks for temporal remote sensing analysis. Outcomes of the study shows quite impressive results over 97% of accuracy in terms of cultivated paddy area detection using optical remote sensing imagery. Moreover, the research was extended to identify cultivated paddy areas using synthetic aperture radar (SAR) imagery. It also outputs a promising result over 96% of accuracy in terms of detecting cultivated paddy regions. The study then extends to detect Brown Planthopper attacks in cultivated paddy fields. Brown Planthopper is considered as the most destructive insect in paddy cultivation. There are no previous studies for identifying Brown Planthopper attacks using satellite remote sensing data under field conditions. In this study, ratio and standard difference indices derived from optical imagery are fed into a Support Vector Machine model to identify the regions affected by Brown Planthopper attacks. Using the results of cultivated paddy fields detection model as a filter, SVM model results are improved. The combined approach shows accuracy over 96% for detecting Brown Planthopper attacks.
item: Thesis-Full-text
Developing a trip distribution model for identified mobility groups using big data
(2020) Rathnayaka BA; Chitraranjan C; Perera AS; Kumarage AS
The need for frequent transportation planning has become a key factor since people started becoming more mobile making urban traffic patterns more complex. The primary source for analysing such travel behavior is through manual surveys. These surveys are expensive, time consuming and often are outdated by the time the survey is completed for analysis. To overcome these issues, Mobile Network Big Data (MNBD) which concerns large data sets can be used over such traditional data collection processes. Call Detail Records (CDR) which is a subset of MNBD is readily available as most of the telecommunication service providers maintain CDR. Thus, analyzing CDR leads to an efficient identification of human behavior and location. However, many researches on CDRs have been done focusing to identify travel patterns in order to understand human mobility behavior. Relatively high percentage of sparse data and other scenarios like the Load Sharing Effect (LSE) causes difficulties in identifying precise location of the user when using CDR data. Existing approaches for identifying precise user location patterns have certain constraints. Past researches utilizing CDRs have used primary approaches in recognizing load sharing effects and have given minimum consideration to the transmission power of the respective cell towers when localizing the users. Furthermore, these studies have neglected the differences in mobility behavior of different segment of users and taken the entire community of users as a single cluster. In this research, a novel methodology to overcome these limitations is introduced for locating users from CDRs by dividing the users into distinct clusters for identifying the model parameters and through enhanced identification of load sharing effects by taking the transmission power into consideration. Further, this study contributes to the transport sector by identifying secondary activities from CDR data, without limiting to the primary activity recognition. This research uses approximately 4 billion CDR data points, voluntarily collected mobile data and manually collected travel survey data to find techniques to overcome the existing limitations and validate the results. Proposed dynamic filtering algorithm for load shared records identification showed a significant improvement on accuracy over previous predefined speed based filtering methods. Further, we found that, IO-HMM outperforms standard HMM results on activity recognition.
item: Thesis-Full-text
Analyzing and modelling web server based systems
(2020) Tennage PN; Jayasena S; Jayasinghe M
Server based systems are widely used in modern computer systems. Understanding the performance of web server based systems, under different conditions is important. This requires a step by step approach that includes modelling, designing, implementing, performance testing and analyzing of results. In this research, we aim at characterizing the web s erver s ystems under d ifferent c onfigurations. We p resent a s ummary o f p revalent server architectures, provide a systematic approach for performance testing, and present a novel open source Python library f or latency analysis. We experiment on existing server architectures, and propose eight new server architectures. Our analysis shows that under different conditions the new architectures outperform the existing architectures. Moreover we do an extensive tail latency analysis of Java microservices.
item: Thesis-Full-text
Automatic generation of elementary level mathematical question
(2020) Keerthisrini PLVS; Ranathunga S
Mathematical Word Problems (MWPs) play a vital role in mathematics education. An MWP is a combination of not only the numerical quantities, units, and variables, but also textual content. Therefore, in order to understand a particular MWP, a student requires knowledge in mathematics as well as in literacy. This makes it difficult to solve MWPs when compared with other types of mathematics problems. Therefore, students require a large number of similar questions to practice. On the other hand, the composition of numerical quantities, units, and mathematical operations impel the problems to possess specific constraints. Therefore, due to the inherent nature of MWPs, tutors find it difficult to produce a lot of similar yet creative questions. Therefore, there is a timely requirement of a platform that can automatically generate accurate and constraint-wise satisfied MWPs. Due to the template-based nature of existing approaches for automatically generating MWPs, they tend to limit the creativity and novelty of the generated MWPs. Regarding the generation of MWPs in multiple languages, language-specific morphological and syntactic features paves way for extra constraints. Existing template-oriented techniques for MWP generation cannot identify constraints that are language-dependant, especially in morphologically rich yet low resource languages such as Sinhala and Tamil. Utilizing deep neural language generation mechanisms, we deliver a solution for the aforementioned restrictions. This thesis elaborates an approach by which a Long Short Term Memory (LSTM) network which can generate simple MWPs while fulfilling above-mentioned constraints. The methodology inputs a blend of character embeddings, word embeddings, and Part of Speech (POS) tag embeddings to the LSTM network and the attention is produced for units and numerical values. We used our model to generate MWPS in three languages, English, Sinhala, and Tamil. Irrespective of the language, the model was capable of generating single and multi sentenced MWPs with an average BLEU score of more than 20%.
item: Thesis-Abstract
Use of machine learning for the prediction of diabetes from photoplethysmography (PPG) measurements & physiological characteristics
(2020) Hettiarachchi CY; Chitraranjan C
Type 2 Diabetes (T2D) is a chronic disease affecting millions of people worldwide. It is a result of impaired glucose regulation, leading to abnormally high levels of glucose causing microvascular and macrovascular problems. The failure to timely identify and treat, results in complications such as limb amputations, blindness and heart disease. Busy unhealthy lifestyles are a root cause and not much effort undertaken to obtain regular health checkups for early T2D detection. Photoplethysmography (PPG) is a non-invasive, optic technique mostly used towards disease estimation in clinical environments. Recent technological advancements have integrated PPG sensors within smartphones and wearables. However, these signals suffer from various noise components, which is intensified in signals acquired in routine everyday environments. The research analysed the feasibility of short (~2.1s) PPG segments in order to address these limitations and identify biomarkers related to T2D. The identified biomarkers mainly relate to the vascular system of the body. Several classification algorithms were evaluated using cross validation to estimate T2D, focussing on a public PPG dataset. Linear Discriminant Analysis (LDA) achieved the highest area under the ROC curve of 79% for the estimation of T2D in a setting where healthy individuals, T2D only, T2D subjects with hypertension and prehypertension were present. It is important to identify relationships between standard medical measures such as Fasting Blood Glucose (FBG) and PPG features, for better understanding T2D estimation. FBG measurements were collected, and several regression algorithms evaluated using leaveone-out cross validation to assess the suitability of predicting FBG using PPG features. The results were examined using the Clarke’s Error Grid, where 75% & 22.5% of predictions were distributed in regions A & B respectively for both ElasticNet and Lasso Regression. The results were comparable with long PPG signal based approaches. The suitability of the method in practical environments was evaluated using simulated PPG signals with noise and motion artifacts. The ElasticNet Regression achieved 70% and 27.5% in regions A & B respectively. The analysis of short PPG segments shows promise towards the development of an early T2D estimation system in a routine everyday environment.