No description available.

No description available.

No description available.

No description available.

A relative, lie-value image normalisation (LVIN) procedure was investigated as a means of estimating surface reflectances from sequences of Landsat TM and ETM+ imagery, and standardising image data for change detection studies when there are uncertainties in sensor calibration and atmospheric parameters over time. The basis of the LVIX procedure is that for an A-date sequence, the digital numbers (DNs) of N-1 overpass images can be mapped to the reflectance values of a reference image for a set of pseudo- invariant targets (PITs) common to all images in the sequence. The robust M-estimator was employed to provide the transformation function that achieved the mapping. The investigation also showed that in some instances the LVIN procedure could incorporate the modelled Path DN-the modelled DN for a target of zero surface reflectance. A lack of surface validation data was a limitation in the investigation. However, a qualitative evaluation of the LVIN procedure was possible by examining the pre- and post-normalisation image histograms. In a comparison with the results of the 6S radiative transfer code, it war observed that when both overpass and reference images were acquired with the same sensor, the LVIK procedure appeared t o correct for atmospheric effects; and when overpass and reference images were with different sensors, the LVIN procedure also corrected for between-sensor differences. Moreover, it was demonstrated for the more "temporally-invariant" PITs that the procedure retrieved surface reflectances that were on average within ±0.02 reflectance units. / The ability of the LVIK procedure to standardise sequences of image data was further demonstrated in the study of vegetation change. The normalised difference vegetation index (NDVI) was calculated from LVIN estimates of surface reflectance for a selection of sites around the township of Mt. Barker, Western Australia. NDVI data had characteristics consistent with data that have been corrected for atmospheric effects. A modification to the LVIN procedure was also proposed based on an investigation of some empirically-derived vegetation reflectance relationships. Research into the robustness of the relationships for a greater range of vegetation types is recommended.

A vertical-style data structure and operations on data in that structure are explored and tested in the domain of sliding window average algorithms for geographical information systems (GIS) data. The approach allows working with data of arbitrary precision, which is centrally important for very large GIS data sets. The novel data structure can be constructed from existing multi-channel image data, and data in the structure can be converted back to image data. While in the new structure, operations such as addition, division, and bit-level shifting can be performed in a parallelized manner. It is shown that the computation of averages for sliding windows on this data structure can be performed faster than using traditional computation techniques, and the approach scales to larger sliding window sizes.

No description available.

No description available.

<p>Artificial Intelligence driven approaches, especially  based on deep learning algorithms, provided an alternative perspective in summarizing the common features in large-scale and complex datasets and aided the human professions in discovering novel features in cross-domain research. In this dissertation, the author proposed his research of developing AI-driven algorithms to reveal the real relation of complex medical data. The author started to identify the abnormal structures from the radiology images. When the abnormal structure was detected, the author built a model to explore the domain layers or cell phenotype of the specific tissues. Finally, the author evaluated cell-cell communication for the downstream tasks.</p> <p><br></p> <p>In his first research, the author applied IResNet, a two-stage prediction-interpretation Convolution Neural Network, to assist clinicians in the early diagnosis of Autism Spectrum Disorders (ASD). IresNet first predicted the input sMRI scan to one of the two categories: (1) ASD group or (2) Normal Control group, and interpret the prediction using a \textit{post-hoc} approach and visualized the abnormal structures on top of the raw inputs. The proposed method can be applied to other neural diseases such as Alzheimer's Disease. </p> <p><br></p> <p>When the abnormal structure was detected, the author proposed a method to reveal the latent relation at the tissue level. Thus the author proposed SiGra, an unsupervised learning paradigm to identify the domain layers and cellular phenotype in a particular tissue slide based on the corresponding gene expression matrix and the morphology representations. SiGra outperformed other benchmarking algorithms in three different tissue slides from three commercialized single-cell platforms.</p> <p><br></p> <p>At last, the author measured the potential interactions between two cells. The proposed spaCI, measured the correlation of a Ligand-Receptor interaction in the high-dimension latent space and predicted the interactive $L-R$ pair for downstream analysis. </p> <p><br></p> <p>In summary, the author presented three end-to-end AI-driven frameworks to facilitate clinicians and pathologists in better understanding the latent connections of complex diseases and tissues. </p>

Data collecting and preparing is generally considered a crucial process in data science projects. Especially for image data, adding semantic attributes when preparing image data provides much more insights for data scientists. In this project, we aim to implement a general-purpose central image data repository that allows image researchers to collect data with semantic properties as well as data query. One of our researchers has come up with the specific challenge of collecting images with weight data of infants in least developed countries with limited internet access. The rationale is to predict infant weights based on image data by applying Machine Learning techniques. To address the data collecting issue, I implemented a mobile application which features online and offline image and annotation upload and a web application which features image query functionality. This work is derived and partly decoupled from the previous project – ImageSfERe (Image Sharing for Epilepsy Research), which is a web-based platform to collect and share epilepsy patient imaging.