Attentive Neighborhood Feature Augmentation for Semi-supervised Learning

Recent state-of-the-art semi-supervised learning(SSL)methods usually use data augmentations as core components.Such methods,however,are limited to simple transformations such as the augmentations under the instance’s naive representations or the augmentations under the instance’s semantic representations.To tackle this problem,we offer a unique insight into data augmentations and propose a novel data-augmentation-based semi-supervised learning method,called Attentive Neighborhood Feature Aug-mentation(ANFA).The motivation of our method lies in the observation that the relationship between the given feature and its neighborhood may contribute to constructing more reliable transformations for the data,and further facilitating the classifier to distinguish the ambiguous features from the low-dense regions.Specially,we first project the labeled and unlabeled data points into an embedding space and then construct a neighbor graph that serves as a similarity measure based on the similar representations in the embedding space.Then,we employ an attention mechanism to transform the target features into augmented ones based on the neighbor graph.Finally,we formulate a novel semi-supervised loss by encouraging the predictions of the interpolations of augmented features to be consistent with the corresponding interpolations of the predictions of the target features.We carried out exper-iments on SVHN and CIFAR-10 benchmark datasets and the experimental results demonstrate that our method outperforms the state-of-the-art methods when the number of labeled examples is limited.

Instance Reweighting Adversarial Training Based on Confused Label

Reweighting adversarial examples during training plays an essential role in improving the robustness of neural networks,which lies in the fact that examples closer to the decision boundaries are much more vulnerable to being attacked and should be given larger weights.The probability margin(PM)method is a promising approach to continuously and path-independently mea-suring such closeness between the example and decision boundary.However,the performance of PM is limited due to the fact that PM fails to effectively distinguish the examples having only one misclassified category and the ones with multiple misclassified categories,where the latter is closer to multi-classification decision boundaries and is supported to be more critical in our observation.To tackle this problem,this paper proposed an improved PM criterion,called confused-label-based PM(CL-PM),to measure the closeness mentioned above and reweight adversarial examples during training.Specifi-cally,a confused label(CL)is defined as the label whose prediction probability is greater than that of the ground truth label given a specific adversarial example.Instead of considering the discrepancy between the probability of the true label and the probability of the most misclassified label as the PM method does,we evaluate the closeness by accumulating the probability differences of all the CLs and ground truth label.CL-PM shares a negative correlation with data vulnerability:data with larger/smaller CL-PM is safer/riskier and should have a smaller/larger weight.Experiments demonstrated that CL-PM is more reliable in indicating the closeness regarding multiple misclassified categories,and reweighting adversarial training based on CL-PM outperformed state-of-the-art counterparts.

A novel CGBoost deep learning algorithm for coseismic landslide susceptibility prediction

The accurate prediction of landslide susceptibility shortly after a violent earthquake is quite vital to the emergency rescue in the 72-h‘‘golden window”.However,the limited quantity of interpreted landslides shortly after a massive earthquake makes landslide susceptibility prediction become a challenge.To address this gap,this work suggests an integrated method of Crossing Graph attention network and xgBoost(CGBoost).This method contains three branches,which extract the interrelations among pixels within a slope unit,the interrelations among various slope units,and the relevance between influencing factors and landslide probability,respectively,and obtain rich and discriminative features by an adaptive fusion mechanism.Thus,the difficulty of susceptibility modeling under a small number of coseismic landslides can be reduced.As a basic module of CGBoost,the proposed Crossing graph attention network(Crossgat)could characterize the spatial heterogeneity within and among slope units to reduce the false alarm in the susceptibility results.Moreover,the rainfall dynamic factors are utilized as prediction indices to improve the susceptibility performance,and the prediction index set is established by terrain,geology,human activity,environment,meteorology,and earthquake factors.CGBoost is applied to predict landslide susceptibility in the Gorkha meizoseismal area.3.43%of coseismic landslides are randomly selected,of which 70%are used for training,and the others for testing.In the testing set,the values of Overall Accuracy,Precision,Recall,F1-score,and Kappa coefficient of CGBoost attain 0.9800,0.9577,0.9999,0.9784,and 0.9598,respectively.Validated by all the coseismic landslides,CGBoost outperforms the current major landslide susceptibility assessment methods.The suggested CGBoost can be also applied to landslide susceptibility prediction in new earthquakes in the future.

Active High-Locality Landslides in Mao County: Early Identification and Deformational Rules

High-locality landslides are located on slopes at high elevations and are characterized by long sliding distances, large gravitational potential energy, high movement velocities, tremendous kinetic energy, and sudden onset. Thus, they often cause catastrophic damage to human lives and engineering facilities. It is of great significance to identify active high-locality landslides in their early deformational stages and to reveal their deformational rules for effective disaster mitigation. Due to alpinecanyon landforms, Mao County is a representative source of high-locality landslides. This work employs multisource data(geological, terrain, meteorological, ground sensor, and remote sensing data) and timeseries In SAR technology to recognize active high-locality landslides in Mao County and to reveal their laws of development. Some new viewpoints are suggested.(1) Nineteen active high-locality landslides are identified by the time-series In SAR technique, of which 7 are newly discovered in this work. All these high-locality landslides possessed good concealment during their early deformational stages. The newly discovered HL-16 landslide featured a large scale and a great slope height, posing a large threat to the surrounding buildings and residents.(2) The high-locality landslides in Mao County were mainly triggered by three factors: earthquakes, precipitation, and road construction.(3) Three typical high-locality landslides that were triggered by different factors are highlighted with their deformational rules under the functions of steep terrain, shattered rocks, fissure-water penetration, precipitation, and road construction. This work may provide clues to the prevention and control of high-locality landslides and can be applied to the determination of active high-locality landslides in other hard-hit areas.

Near real-time spatial prediction of earthquake-induced landslides:A novel interpretable self-supervised learning method

Near real-time spatial prediction of earthquake-induced landslides(EQILs)can rapidly forecast the occurrence position of widespread landslides just after a violent earthquake;thus,EQIL prediction is very crucial to the 72-hour‘golden window’for survivors.This work focuses on a series of earthquake events from 2008 to 2022 occurring in the Tibetan Plateau,a famous seismically-active zone,and proposes a novel interpretable self-supervised learning(ISeL)method for the near real-time spatial prediction of EQILs.This new method innovatively introduces swap noise at the unsupervised mechanism,which can improve the generalization performance and transferability of the model,and can effectively reduce false alarm and improve accuracy through supervisedfine-tuning.An interpretable module is built based on a self-attention mechanism to reveal the importance and contribution of various influencing factors to EQIL spatial distribution.Experimental results demonstrate that the ISeL model is superior to the excellent state-of-the-art machine learning and deep learning methods.Furthermore,according to the interpretable module in the ISeL method,the critical controlling and triggering factors are revealed.The ISeL method can also be applied in other earthquake-frequent regions worldwide because of its good generalization and transferability.