We have recently introduced a novel methodology for the noninvasive analysis of the structure and composition of human skin in vivo. The approach combines pulsed photothermal radiometry (PPTR), involving time-resolved measurements of mid-infrared emission after irradiation with a millisecond light pulse, and diffuse reflectance spectroscopy (DRS) in the visible part of the spectrum. Simultaneous fitting of both data sets with respective predictions from a numerical model of light transport in human skin enables the assessment of the contents of skin chromophores (melanin, oxy-, and deoxy-hemoglobin), as well as scattering properties and thicknesses of the epidermis and dermis. However, the involved iterative optimization of 14 skin model parameters using a numerical forward model (i.e., inverse Monte Carlo - IMC) is computationally very expensive. In order to overcome this drawback, we have constructed a very fast predictive model (PM) based on machine learning. The PM involves random forests, trained on approx. 9,000 examples computed using our forward MC model. We show that the performance of such a PM is very satisfying, both in objective testing using cross-validation and in direct comparisons with the IMC procedure. We also present a hybrid approach (HA), which combines the speed of the PM with versatility of the IMC procedure. Compared with the latter, the HA improves both the accuracy and robustness of the inverse analysis, while significantly reducing the computation times.
F.01 Acquisition of new practical knowledge, information and skills
COBISS.SI-ID: 33232423Surrogate models approximate the predictions of other models. The motivation for learning surrogate models can come from computational concerns, when the predictions of the original model are computationally expensive to obtain. In contrast, the surrogate models are computationally efficient. In this paper, we propose a framework for machine learning of surrogate models, which operate on the same input and output spaces as their original models. Instead of learning direct mappings from the input to the output space (and vice versa), we first assess the intrinsic dimensionality of the input and output spaces and reduce it appropriately, by using PCA and autoencoders. Predictive models are learned on the reduced spaces by the use of neural networks and their predictions are mapped to the original spaces. We apply the framework to learn a surrogate model for a complex radiative transfer model RemoTeC, designed and built at SRON in the Netherlands. The original model predicts shortwave infrared (SWIR) spectra, for a given state vector of atmospheric parameters, representative of any geo-location that the Sentinel 5P satellite may encounter. The results indicate a low dimensionality of both the input and the output space and are accurate in both the forward and reverse direction.
F.01 Acquisition of new practical knowledge, information and skills
COBISS.SI-ID: 52269827We organized DS 2019, the 22nd International Conference on Discovery Science in Split, Croatia. The conference took place in the period October 28-30, 2019. Content-wise, this series of conferences is closest to the topic of the present SESAME project. The conference attracted around 80 participants from around the world. The program included three keynote talks on methods for discovering knowledge from complex data and their use in environmental and life science-medicine. It also included 26 regular papers (and 19 short ones) selected by peer review from a set of 76 submissions. Professor Džeroski, the project leader, was the general chair of the conference. He was co-editor of the proceedings, regularly published by Springer in the Lecture Notes in Computer Science series. Professor Džeroski is currently the chair of the steering committee for this conference series (in the period 2018-2023), supervised the organization of the 2020 edition, and is currently overseeing the organization of the 2021 edition.
B.01 Organiser of a scientific meeting
COBISS.SI-ID: 32805927