- 4. Department of Meteorology, University of Reading , Reading RG6 - PowerPoint PPT Presentation

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Practical machine learning based on cloud computing resources Kyriakos N. Agavanakis 1,a) , George. E. Karpetas 2,b) , Christos M. Michail 3,c) ,Michael Taylor 4,d) , LampriniKontopoulou 7,h) , Varvara Trachana 5,e), EvangeliaPappa 6,f) , John Filos 6,g) 1. Atrinno , Attica Research and Innovation PC 2.Department of Medical Physics, Faculty of Medicine, University of Thessaly 3. University of West Attica , Department of Biomedical Engineering Radiation Physics, Materials Technology and Biomedical Imaging Laboratory - 4. Department of Meteorology, University of Reading , Reading RG6 6BB, UK 5. Laboratory of Biology, Faculty of Medicine, University of Thessaly 6.Department of Public Administration School of Economy and Public Administration, Panteion University of Social and Political Sciences 7. General Department, University of Thessaly , 41110, Larissa, Greece

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Purpose Investigate the practical applications of machine learning (ML) algorithms in several scientific areas, and Utilize cloud resources to provide usable services not only within the scientific community, but to everybody!

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Case studies ➢ quality evaluation metrics for the tomographic image reconstruction of positron emission tomography (PET) images ➢ health implications of the vitamin D absorption function. Results showed that commercially available cloud resources are over sufficient to consolidate results from a variety of teams and applications and contribute to the built up of a valuable shared knowledge repository ➢ the investigations of the demographic determinants influencing the perception of corruption incidents within different industry sectors

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Achievements Using the suggested approach in the context of a widely available cloud service for feeding the training algorithms, will contribute to more accurate automation and successful operation of related activities in the application domains, breaking thus the knowledge silos and contributing to a more sustainable environment.

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon ➢ CASE STUDY Modulation Transfer Function calculation using cloud-based Machine Learning Services Evangelia Pappa John Filos

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Definitions Spatial resolution – the amount of geometric detail • How close can two points be before you can’t distinguish them

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Good Poor Imaging As spatial separation decreases, the “good” system maintains clear separation of point source images, while the “poor” system eventually can no longer distinguish them. MTF quantifies this phenomenon in terms of contrast between the center peak intensities versus intensity at their midpoint across a scale of separation distances. At large separations, even a poor system can completely resolve the two images. As separation decreases, only the good systems can still recognize separate sources.

Dr Georgios E. Karpetas Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Image Quality in Nuclear Imaging The response of the system to the incident signal amplitudes can be described bythe : M odulation T ransfer F unction( MTF ), which expresses the system’s response in the spatial frequency domain by taking the Fourier transform of the corresponding PSF from a reconstructed cross-sectional image. 8

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon 14:35:09 DIFFRACTION LIMIT jhimgrint WAVELENGTH WEIGHT T R 0.0 FIELD ( ) 0.00 O • MTF is a measure of 11500.0 NM 1 T R 0.7 FIELD ( ) -3.49 O 10000.0 NM 1 DIFFRACTION MTF T R 1.0 FIELD ( ) -4.79 O 9000.0 NM 1 intensity contrast 13-Mar-00 R -1.0 FIELD ( ) T 4.79 O 8000.0 NM 1 DEFOCUSING 0.00000 1.0 transfer per unit 0.9 resolution of an image Focal Length 3.94" 0.8 F#/1.64 or signal. 0.7 Pupil Diameter 2.4" M O 0.6 D • It is used in optics, U L 0.5 A electronics, and related T I O 0.4 N signal processing 0.3 applications. 0.2 0.1 R 1.0 8.0 15.0 22.0 29.0 36.0 43.0 50.0 57.0 64.0 71.0 T SPATIAL FREQUENCY (CYCLES/MM)

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon MTF curves obtained from iterative STIR reconstructed LSF SF images (the number of subsets was kept fi fixed and the number of iterations was increased with a step of 2) Simulationof the planesource for the MTF measurement Schematic representation of the line profile selection

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon Spatial Data Frequency format… Subset s 1 1 1 1 1 1 3 3 Iterations 1 2 6 8 14 20 2 6 1,00000 1,00000 1,00000 0,99533 1,00000 1,00000 1,00000 1,00000 0,000000 X 0 0 0 4 0 0 0 0 0,92441 0,99517 0,99930 0,99941 0,99875 0,99931 0,001151888 6 2 0,9987 1 5 7 1 5 0,73024 0,98082 0,99480 0,99815 0,99722 0,99766 0,99501 0,99726 Subs Spatial 0,002303776 6 7 9 4 4 9 1 2 ets Frequency Iterations MTF 0,49295 0,95737 0,98835 0,99376 0,99476 0,98881 1 1 1 0,000000 1,000000 0,003455664 4 7 90,98708 5 4 10,99385 0,28436 0,92548 0,97939 0,98488 0,98894 2 1 0,99071 2 0,98019 0,98909 0,000000 1,000000 6 5 8 8 3 1 5 2 0,004607552 3 1 6 0,000000 1,000000 0,14017 0,88603 0,96799 0,97939 0,98277 0,98552 0,96922 0,98300 . 0,00575944 9 8 6 9 8 7 7 9 . 0,05905 0,95424 0,96272 0,97529 0,97922 0,95598 0,97562 . 0,006911328 1 0,8401 1 2 4 7 9 5 . 8 0,000000 0,995334 324 Data conversion 0 21 20 0,000000 1,000000

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon ➢ CASE STUDY Bio-uv products Neural network calculation of Vitamin-D and DNA-damage doses from spectral UV irradiance using cloud-based services Michael Taylor, Lamprini Kontopoulou, Surftemp Satellite Remote Sensing Group Varvara Trachana 5,

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon SATELLITE UV DOSE • Satellites like SCIAMACHY and GOME-2 have operational processing algorithms that retrieve erythemal UV dose (kJ m -2 ) from space: Used to calculate the UV index in Greece and across Europe http://meteo.gr/u v.cfm Van Geffen, J., Van Weele, M., Allaart, M. and Van der A, R.: 2017, TEMIS UV index and UV dose operational data products: http://www.temis.nl/uvradiation/UVarchiv e.html

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon BIOLOGICAL UV PRODUCTS • Interestingly, you can use the satellite UV together with window functions (“action spectra”) to calculate important biological UV products across the Earth’s surface: 1) Vitami n-D dose 2) DNA-damage dose Zempila, M. M., van Geffen, J. H., Taylor, M., Fountoulakis, I., Koukouli, M. E., van Weele, M., Bais, A., Meleti, C., Balis, D. (2017). TEMIS UV product validation using NILU-UV ground-based measurements in Thessaloniki, Greece. Atmospheric Chemistry and Physics , 17 (11), 7157-7174.

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon • Using the viewing potential of satellites, this means we can generate maps of these UV products for most of the Earth surface - but only once a day : Van Geffen, J., Van Weele, M., Allaart, M. and Van der A, R.: 2017, TEMIS UV index and UV dose operational data products: http://www.temis.nl/uvra diat ion/U Va rchive.ht ml

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon • As well as being sensitive to cloud, the UV reaching ground is also sensitive to absorbing aerosol (e.g. desert dust) – the combination of these 2 factors is a challenge for neural network models: Van Geffen, J., Van Weele, M., Allaart, M. and Van der A, R.: 2017, TEMIS UV index and UV dose operational data products: http://www.temis.nl/uvra diat ion/U Va rchive.ht ml

Practical machine learning based on cloud computing resources TMREES’19, Beirut, Libanon BPNN MODEL • A high frequency (1 minute interval) back-propagation neural network (BPNN) model has recently been developed to calculate these biological products from UV irradiances at 5 wavelengths plus the solar zenith angle (SZA) as inputs: Ir (3 0 5 n m ) Ir (3 1 2 n m) Ir (3 2 0 n m) E rythe ma l UV dose Ir (3 4 0 n m ) Vita m in-D dose Ir (3 8 0 n m ) SZA DNA-da m a ge dose Zempila, M. M., van Geffen, J. H., Taylor, M., Fountoulakis, I., Koukouli, M. E., van Weele, M., Bais, A., Meleti, C., Balis, D. (2017). TEMIS UV product validation using NILU -UV ground-based measurements in Thessaloniki, Greece. Atmospheric Chemistry and Physics , 17 (11), 7157-7174.