<p class="fluidplugincopy"><span class="contentpasted0">Landslides in remote or uninhabited regions can be undocumented, leaving gaps in landslide inventories </span><span class="contentpasted0">which are a key input for hazard and risk assessments</span><span class="contentpasted0">. This can lead to landslide events being missing from research studies, and contribute to a bias in the events used for training of machine learning models.</span></p> <p class="fluidplugincopy"><span class="contentpasted0">In this work we use satellite images, terrain information, and labelled examples of landslides to train a convolutional neural network (U-Net), for the purpose of adding previously undocumented and new landslides to inventories. This model segments the input images and highlights the pixels it labels as landslides.</span></p> <p class="fluidplugincopy"><span class="contentpasted0">Our work focusses on landslides with a range of types and triggers, so that the model is exposed to a variety of training data. We describe the key properties of the landslides in the training set, and discuss the implications for future uses of the trained model.</span></p>

<p>Image segmentation algorithms are a type of image classifier that assigns a label to each individual pixel in an image. U-Nets, initially developed for the analysis of biomedical images and now widely used in a variety of fields, are an example of such algorithms. It has been shown that U-Nets are specially interesting when working with small training datasets and combined with data augmentation techniques.</p> <p>In this study, we used satellite images with labelled landslide masks from known events to train a U-Net to identify areas of potential landslide. These landslide masks are time-consuming to create, resulting in a small initial training set. Even when working with U-Nets, the success of machine learning and AI tools depends on the availability and quality of training data, as well as the algorithm settings during the training process. Tuning machine learning models to achieve the best performance possible from limited amounts of data is important to generate trustworthy results that can be used to advance the knowledge of landslide events around the world.</p> <p>Here, we show the differences in algorithm performance as we use different types of data augmentation and model parameters. We also explore and assess the effects on performance of options such as including different satellite bands, terrain information and alternative colour band representations.</p>

<div> <p>Heterogeneities on the scale of the seismic wavelength in the Earth's crust and mantle cause complex wavefield fluctuations in time and amplitude which are known to affect velocity and source inversions, as well as other seismic characterisations. However, many seismic models ignore these heterogeneities for simplicity. As part of our longer-term goal to account for these, we attempt to rigorously and probabilistically characterise these lithospheric small-scale heterogeneities by combining a single-layer and a multi-layer energy flux models with a new Bayesian inference algorithm. The first technique characterizes energy losses to the ballistic arrival as intrinsic, diffusion and scattering quality factors, which allows us to compare the effects of these attenuation mechanisms on our data. With the second method, we can obtain synthetic coda envelopes for 1- and 2- layer models with different values of the correlation length and fractional velocity fluctuations in each layer. We then use the Metropolis-Hastings algorithm to sample the likelihood space and obtain the posterior probability distributions for each parameter and layer in the model. Our thorough testing of these methods reveals complicated trade-offs between the parameters and highly non-unique solutions, thus highlighting the importance of the Bayesian approach for scattering studies. Previous studies applying these methods used a more traditional grid search for their coda inversion, which may have affected their results. We applied this approach to a data set of over 300 events from three seismic arrays in Australia: Alice Springs array (ASAR), Warramunga Array (WRA) and Pilbara Seismic Array (PSA). The results from the single-layer energy flux model show that all quality factors take higher values for PSA than for the other two arrays, indicating that the structure beneath this array is less attenuating and heterogeneous than for the other arrays. Intrinsic and diffusion attenuation are strongest for ASAR, while scattering and total attenuation are similarly strong for ASAR and WRA. Our multi-layer model results show the crust is more heterogeneous than the lithospheric mantle for all arrays, with crustal values of the correlation length and velocity fluctuations being lower for PSA than for the other arrays, indicating the presence of weaker and smaller scale heterogeneity beneath this array. We attribute these differences and similarities in the attenuation and heterogeneity structure beneath the arrays to variations in the tectonic history of the areas they are located on. This new Bayesian approach to the multi-layer energy flux model, in combination with the single-layer model, not only allows us to determine and compare the different quality factors, but also gives us detailed information about the trade-offs and uncertainties in the determination of the scattering parameters, making it a useful tool for future scattering and small-scale structure studies. </p> </div>

No abstract

<div> <p><span>P waves are often used to calculate the yield of chemical or nuclear explosions in forensic seismology. These estimations often rely on amplitude measurements affected by seismic scattering and attenuation caused by the presence of heterogeneities on the scale of the seismic wavelength and seismic energy conversion into heat, both on the source and receiver side. It is therefore important to accurately characterize the effect of these phenomena on the recorded wavefields so that any source size (and type) obtained from them are not under or overestimated. </span> <br><span>In our previous study (González Alvarez et al., 2021), we combined single layer and multi-layer energy flux modeling with a Bayesian inference algorithm to characterize lithospheric small-scale heterogeneities beneath seismic stations or arrays by calculating the characteristic scale length and fractional velocity fluctuations of the crust and lithospheric mantle beneath them. Here, we take this approach further and remove the dependence on the less realistic, single layer energy flux model by including the intrinsic quality factor and its frequency dependence as free parameters into our Bayesian inference algorithm. We use the multi-layer energy flux model to produce synthetic envelopes for 2-layer models of the lithosphere for different values of the scattering and intrinsic attenuation parameters. We then use our improved Bayesian inference algorithm to sample the likelihood space by means of the Metropolis-Hastings algorithm and obtain posterior probability distributions for all parameters and layers in the model. To our knowledge, such an approach has not been attempted before. We thoroughly tested this inversion algorithm and its sensitivity to four different levels of crustal and lithospheric mantle intrinsic attenuation settings using 18 synthetic datasets. Our results from these tests, while showing complex trade-offs between the parameters, show that scattering parameters can be recovered accurately in most cases. Intrinsic attenuation shows higher variability and non-uniqueness in our inversions, but can generally be recovered for over half of the synthetic models. To further test the accuracy of the results obtained from this Bayesian algorithm, we applied this technique to the large, high-quality dataset from PSAR and IMS arrays ASAR and WRA used in our previous study and found excellent agreement between both approaches in all cases.</span> <br><span>Finally, we applied this technique to datasets of teleseismic earthquakes from several arrays part of the IMS (YKA, ILAR, TXAR, PDAR, BOSA and KURK) to characterize the lithospheric scattering and attenuation structure beneath them and relate our findings to the tectonic setting and history of the regions they are installed on. </span><span> </span></p> <p><span>González Álvarez, I.N., Rost, S., Nowacki, A. and Selby, N.D., 2021. Small-scale lithospheric heterogeneity characterization using Bayesian inference and energy flux models. <em>Geophysical Journal International</em>, <em>227</em>(3), pp.1682-1699.</span></p> </div>