Search

We consider synchronous data-parallel neural network training with a fixed large batch size. While the large batch size provides a high degree of parallelism, it degrades the generalization performance due to the low gradient noise scale. We propose a general learning rate adjustment framework and three critical heuristics that tackle the poor generalization issue. The key idea is to adjust the learning rate based on geometric information of loss landscape and encourage the model to converge into a flat minimum that is known to better generalize to the unknown data. Our empirical study demonstrates that the Hessian-aware learning rate schedule remarkably improves the generalization performance in large-batch training. For CIFAR-10 classification with ResNet20, our method achieves 92.31% accuracy using 16,384 batch size, which is close to 92.83% achieved using 128 batch size, at a negligible extra computational cost. • Hessian information allows to properly adjust noise scale in large-batch training. • Too early learning rate decay harms underlying margin distribution. • The minimum learning rate after the decay strongly affects the model sharpness. • The length of noise scale transition affects the final generalization performance. [ABSTRACT FROM AUTHOR]

Local Stochastic Gradient Descent (SGD) with periodic model averaging (FedAvg) is a foundational algorithm in Federated Learning. The algorithm independently runs SGD on multiple clients and periodically averages the model across all the clients. This periodic model averaging potentially causes a significant model discrepancy across the clients making the global loss converge slowly. While recent advanced optimization methods tackle the issue focused on non-IID settings, there still exists the model discrepancy issue due to the underlying periodic model averaging. We propose a partial model averaging framework that mitigates the model discrepancy issue in Federated Learning. The partial averaging encourages the local models to stay close to each other on parameter space, and it enables to more effectively minimize the global loss. We extensively evaluate the performance of the partial averaging strategy using CIFAR-10/100 and FEMNIST benchmarks. Given a fixed number of training iterations and a large number of clients (128), the partial averaging achieves up to 2.2% higher accuracy than the periodic full averaging. • A novel partial model averaging scheme that accelerates the federated optimization. • Frequent partial model synchronizations strongly suppress the model discrepancy across clients. • While maintaining the total data transfer size, the partial averaging accelerate the convergence of global loss. [ABSTRACT FROM AUTHOR]

Landslides are one of the commonly natural disasters triggered by rainfalls, earthquakes, and human activities, which cause fatalities, damage to properties, and economic losses in the world. Early warning method may predict landslide occurrence and reduce the risk. In this paper, we propose a self-developed site-specific landslide early warning system (LEWS), which has been progressively designed and accomplished over the past decade in Southwest China. The warning model for the prediction of slope instability is focused on multiple thresholds, including deformation rate, rate increment and tangential angle related to surface displacement measurement. The debris-flow initiation is determined by critical rainfall thresholds integrated with the topographic and geological conditions at a catchment scale. The recent Xingyi rockslide, Baige debris slide and debris flows in Zoumaling catchment are selected to explain the specific process of landslide early warning in a real-time operating system. The performance of the presented warning system is evaluated by comparison with the inverse-velocity model (INV) and gradient model (SLO) in the prediction of slope failure time. The Receiver Operator Characteristic (ROC) analysis is used to prove its feasibility and reliability in the debris-flow initiation prediction. The presented warning model and system can be applied to other regions in landslide early warning and to be useful to mitigate landslide losses and damages. • Provide a successfully real-time landslide early warning system. • Warning model and warning system can be applied in an emergency response to mitigate landslide losses. • Comparing with the SLO, INV model and using ROC analysis to evaluate the performance of the warning system. [ABSTRACT FROM AUTHOR]

Spatial attitudes are widely used in geological field investigations, to describe the geometry of geological structural surfaces. Smartphones have a variety of sensors that can be used to obtain these measurements. This paper examines several methods for measuring attitudes using smartphones and analyzes the principles and calculation formulas of these methods. We programmed an Android application to test and analyze the accuracy of different calculation methods. We propose the use of an orientation angle to quantitatively describe the position of the smartphone during measurement. We designed and manufactured a test platform that can change three angles separately, thereby greatly increasing the convenience of the test. The results show that the trigonometric function method can deliver an accurate measurement regardless of the orientation of the phone, and this method is convenient and quick to use. • Introduced a variety of complete calculation methods of measuring the spatial attitude. • A trigonometric function method using the orientation sensor is proposed, and the results are relatively accurate. • Proposed the orientation angle to quantitatively describe the posture of the mobile phone during measurement. • Designed and manufactured a test bench to simulate multiple angle combinations. [ABSTRACT FROM AUTHOR]

Early warning of landslides is crucial for risk management and reduction, attracting a lot of attention from both scientists and stakeholders. However, it is challenging due to the complex nature of landslide behaviors and failure mechanisms. Here, we present a recent case of successful early warning and timely evacuation in advance of a large rockslide that occurred on 17 February 2019, in Guizhou Province, China. The rockslide was initially triggered in 2014 due to the excavation of slope toe for road expansion. Since then, the rockslide had become a potential threat to the local residents, pedestrians, and traffic. To ensure their safety, a wireless monitoring network combining on-site sensors and the geodetic method by Global Navigation Satellite System (GNSS) was installed to continuously monitor the surface displacement of the rockslide. Field monitoring data measured by crack gauges, rain gauges, and tiltmeter were transmitted to a real-time early warning system developed using the new artificial intelligence by the authors' institute. Since the deformation of the rock mass was found increasing, nearby residents were evacuated immediately. Using predefined early warning thresholds for rockslides in the system, the rockslide was successfully forecasted 53 min in advance. Prompt action taken by scientists and local authorities averted human and economic losses completely. In this study, we introduce the real-time early warning system, its concept, the method for determining warning threshold, and performance, followed by the emergency mitigation measures performed for this particular rockslide. It is the 5th time our early warning system successfully forecasted a landslide since its implementation in 2017, and hence we discuss the key characteristics of the system in order to make it applicable for other cases globally. [ABSTRACT FROM AUTHOR]

We release two datasets that track the enhanced landsliding induced by the 2008 Mw 7.9 Wenchuan earthquake over a portion of the Longmen Mountains, at the eastern margin of the Tibetan Plateau (Sichuan, China). The first dataset is a geo-referenced multi-temporal polygon-based inventory of pre- and coseismic landslides, post-seismic remobilisations of coseismic landslide debris and post-seismic landslides (new failures). It covers 471 km 2 in the earthquake's epicentral area, from 2005 to 2018. The second dataset records the debris flows that occurred from 2008 to 2017 in a larger area (∼17000 km 2), together with information on their triggering rainfall as recorded by a network of rain gauges. For some well-monitored events, we provide more detailed data on rainfall, discharge, flow depth and density. The datasets can be used to analyse, on various scales, the patterns of landsliding caused by the earthquake. They can be compared to inventories of landslides triggered by past or new earthquakes or by other triggers to reveal common or distinctive controlling factors. To our knowledge, no other inventories that track the temporal evolution of earthquake-induced mass wasting have been made freely available thus far. Our datasets can be accessed from https://doi.org/10.5281/zenodo.1405489. We also encourage other researchers to share their datasets to facilitate research on post-seismic geological hazards. [ABSTRACT FROM AUTHOR]

In mountainous areas of Southwestern China, debris flows occur every rainy season in recent years and have become one of the main threats to human life and properties. After a giant debris-flow event on August 13, 2010 in Sichuan province, the need to deploy a real-time monitoring network and early warning system (EWS) in this region with high potential risk became of paramount importance. This paper, therefore, presents a proposal for a 3D WebGIS-based platform combined with wireless sensor network (WSN) for debris-flow monitoring, analysis of rainfall thresholds for potential debris-flow gullies, and preliminary criteria for automatic and continuous monitoring and alert of debris flows. The choice of a WSN environment simplifies data collection, especially for the combination of various monitoring data provided by different sensor terminals. Warning thresholds stored in the database can easily be modified to improve the prediction of debris-flow occurrence, and there is a user-friendly 3D graphical interface which allows an almost real-time solution for decision-makers. The automatic computer-based EWS applied in the case study proves to be useful for debris-flow early warning, and improvements expected from continuation of this study could ultimately reduce the losses caused by such type of natural hazard. [ABSTRACT FROM AUTHOR]

Real-time monitoring and intelligent early warning system are crucial and significant to take mitigation measures and reduce casualties and property losses related to landslides. It is difficult to obtain entire monitoring data in the accelerated deformation phase in a landslide event, and hard to issue early warning information using a traditional monitoring approach with fixed and low sampling frequency. Displacement increments of loess landslides induced by agriculture irrigation on the Heifangtai terrace could be sudden and extremely rapid. Typical landslide types include loess flowslides and loess falls. It is of practical significance to develop a self-adaptive data acquisition monitoring technique and establish a real-time landslide early warning system (LEWS) to meet the needs for risk mitigation of rapid sliding slopes on the Heifangtai terrace. The monitoring technique can wirelessly transmit displacement data and the LEWS was devised using the new artificial intelligence. The LEWS could automatically release the warning information in advance of the event once the early warning parameters exceed default thresholds. In this study, the early warning procedures, real-time monitoring approach, intelligent LEWS, a multiple criteria warning model, warning release and emergency mitigation measures, and performance are introduced in detail. Six loess landslides at Heifangtai and eight landslides in other regions of China have been successfully warned since its implementation in 2012. This study proposed an effective and practical solution for the early warning of loess landslides at Heifangtai. Two typical loess landslides that had successful early warnings at Heifangtai were presented. The successful implementation could serve as a reference for global rapid slope failure cases, considering the complex nature of landslide behaviors and failure mechanisms. • A self-adaptive data acquisition technique was successfully developed to monitor sudden failure of loess landslides. • 3D WebGIS-based platform developed by SKLGP becomes more intelligent and accurate. • Multiple alert criteria model including deformation rate, the rate increment, and tangential angle, is established. • The LEWS with multiple alert criteria model has successfully forecasted six loess landslides since its implementation in 2017. [ABSTRACT FROM AUTHOR]

The upper Minjiang catchment is frequently affected by a significant number of large to giant landslides. In one such massive older landslide, approximately 2 million m3 of sediment was reactivated, threatening the safety of a pylon on the landslide and the safety of the only route that allows high-speed travel from Wenchuan to Maoxian. To trace the deformation evolution history and analyze the current deformation of the new landslide, we examined SAR data since 2007, carried out geological and UAV surveys, and designed and installed real-time monitoring equipment. This paper presents the main spatial and temporal characteristics of the landslide. We found that the landslide has experienced creep deformation in the past five years and its sensitivity to rainfall. The development process of the landslide from the formation of tensile crack at the trailing edge to the gradual pushing forward of the crack propagation was revealed. The innovation of this paper is to reveal the whole process of the spatial-temporal evolution from its beginning to large deformation of a real case of creep landslide, which can provide a very good reference for the analysis and monitoring of such creeping landslide evolution process. This representative case also shows how to combine remote sensing and in situ monitoring methods to carry out continuous monitoring and the establishment of a warning system for potential landslide sites, so as to trace the evolution history, derive activity maps and analyze development trend, which can significantly reduce landslide risk. Additional mitigation procedures can be put into place if continuous monitoring of the Minjiang River Canyon occurs, resulting in a reduction of direct losses. In particular, modeling of the slope deformation and failure, damming process, and prospective flooding can be done in advance. Therefore, it is expected that this report can strengthen the case for the deployment of systematic monitoring of potential landslide sites by integrating earth observation methods and in situ monitoring systems as an approach to reduce risk. • The whole process of the spatial-temporal evolution of a real case of creep landslide is revealed. • This representative case shows how to combine remote sensing and in situ monitoring methods to carry out continuous monitoring. • Ground-based real-time monitoring provides accurate landslide early warning. [ABSTRACT FROM AUTHOR]