As a manifestation of the contradiction between human and nature, endangering human survival and living environment, natural hazards happen all the times around the world. Few in sample points results from low-frequency in recuring same type of hazard in a region. Enlightened by the thought of swapping space for time, which can solve the larger issues of space-time development, hazards similarity can be used to provide more practiced knowledge for hazard management, for example, identifying potential solutions from the cities have suffered similar hazards. This study includes the following work: 1) Through computer or GIS approaches like natural language processing, reverse geocoding, spatial query to build a space-time hazard event database of China; 2) Mining characteristics among the data and produce multi-hazard groups in virtue of DBSCAN algorithm; 3) Analyzing spatial distribution of multi-hazard and its geographical attributions, as well as generating multi-hazard similarity indexes for cities; 4) Exploring the similar degree of multi-hazard among cities with similar geographical environment, accordingly obtaining the relations between multi-hazard and geographical environment. Our results can help search for more similar cities threatened by similar multi-hazard, and share resources and best practice on specific challenges, realizing efficient management of hazard risk.

The establishment of dense seismometer arrays in Japan approximately 20 years ago led to the discovery of deep low-frequency tremors, which were weak oscillations categorized as slow earthquakes. Currently, only digital data from the last 20 years are available for studying tremors. As many studies have been indicated the relations between tremors and large earthquakes, it is important to investigate tremors that occurred before establishing the dense seismometer arrays. Past seismometers used more than 50 years ago drew waveforms continuously on paper wrapped on a drum. The digitization of seismograph paper records by tracing the waveforms is effective for investigating large earthquakes because such waveforms are distinctive and consequently extractable given the low frequencies and large amplitudes. In contrast, tremors have much smaller amplitudes and higher frequencies than large earthquakes, so their digitization is much more difficult due to overlapping waveforms. Therefore, we use a convolutional neural network (CNN), a deep-learning model that exhibits high performance for image recognition, aiming to detect evidence of tremors from the seismograph paper records. Training a CNN from scratch with real data polluted by a variety of noises may hinder the model construction and hyperparameter tuning. Thus, we conducted numerical experiments to train a CNN with synthetic images generated according to seismograph paper records. We constructed the CNN based on the ResNet architecture. The results show that the trained model can learn tremor features and correctly determine the presence of tremors in the seismic waveforms. In addition, gradient-weighted class activation maps clearly indicate the tremor location on each image. These suggest that a CNN can be a promising alternative for effective tremor detection compared to individual waveform extraction through digitization. Based on the finding from the experiments, we will conduct CNN training with real data and apply the trained model to seismograph paper records.

Regional mass transfer in subduction-related metamorphic rocks remains information on fluid activities in subduction zones and provides a key to reveal fluid-related processes. However, quantitative metamorphic mass transfer analysis is limited in outcrop scales (i.e., 1–10m) due to large protolith heterogeneity inherited in regional scales. To analyze mass transfer in subduction zones (i.e., regional scales), a new analytical method considering protolith heterogeneity is needed. We have developed protolith reconstruction models from metamorphic rocks using machine learning to evaluate metamorphic mass transfer. The models were constructed by learning protolith basalt compositional dataset (Ocean Island basalt, Mid-ocean ridge basalt, and Island arc basalt), and designed to estimate trace element compositions from limited numbers of input trace element concentrations (i.e., 2–9 elements). Gradient Boosting Decision Tree was used as a machine learning algorithm, and models were evaluated by the Root Mean Squared Error in the log unit. Zr, Th, Ti, Nb, La, Ce, Nd, Yb, and Lu are chosen as input elements, and Rb, Ba, U, K, La, Ce, Pb, Sr, Nd, Y, Yb, Lu, Zr, Th, Ti, and Nb are chosen as output elements. To optimize the ideal numbers and combinations of input elements, models were constructed for each combination of input and output elements. As a result, it is revealed that basalt trace element compositions can be estimated from only 4 element inputs within errors of 10%–25% (i.e., <0.1 in log10 unit). Especially Th, Nb, Zr, and Ti, the most immobile elements in metamorphism (e.g., Ague 2017), were adopted as inputs for representative models. The models have applied to seafloor altered basalt with known protolith compositions to ensure model reliability. This developed method enables us to analyze mass transfer quantitatively on regional scale, and the obtained results provide insights into elemental budgets and fluid activities in subduction zones.

Tectonically stable far-field regions that are removed from the deformation zone of former Late Pleistocene ice sheets present an opportunity to understand global mean sea-level changes over the Holocene. However, the vertical spread in existing relative sea-level reconstructions from the Singapore region and temporal data gaps makes discerning the timing and amplitude of the mid-Holocene highstand and nature of late Holocene relative sea-level fall challenging. As different geological proxies each have their own advantages and limitations, the use of a multi-proxy approach can enhance the robustness of our relative sea-level reconstructions, particularly if multiple proxies converge on a similar story. Here, we present new data from in-situ fossil and living coral microatolls in Sentosa and St. John’s island, Singapore, which can hopefully complement existing data in the region.  Coral microatolls are precise sea-level indicators as they grow in shallow waters over a narrow vertical range, their upward growth restricted by subaerial exposure at the lowest of tides. Determination of relative sea-level change by comparison of the elevations of fossil to living microatolls at the same site relies on the crucial assumption that the indicative meaning of the microatolls (i.e. their vertical living range) has remained unchanged through time at that site. Comparison of the indicative meaning of the highest level of survival (HLS) of Porites microatolls at Sentosa and St. John’s may shed light on some of the processes controlling this upper limit of growth. We are also exploring the viability of using photogrammetry as a non-invasive means of reconstructing past sea levels using coral microatolls. Work is ongoing to date the fossil microatolls in Sentosa, which will be compared to published and as-yet-unpublished relative sea-level data in the region.

Coral microatolls are valuable proxies of relative sea-level (RSL) change. These corals grow up and outwards and are limited by surface water heights. Extended periods of low water result in desiccation and death of the exposed living surface. The highest level of survival (HLS) on a coral describes the upper limit to which the coral polyps can survive. Determining the uncertainties in RSL reconstructions commonly involves quantification of the variability in HLS, although the exact drivers of this variability remain unknown. Between February and July 2020, the microatolls off Sentosa and St. John’s island sustained a partial mortality event. The diedowns were larger at Sentosa, suggesting different factors controlling HLS at each site. In this study, we attempt to discriminate the drivers of HLS variability at these sites. We used photogrammetry as a non-invasive method to track HLS variability. We obtained imagery of one coral microatoll at Sentosa and three at St John’s Island in July and December 2020 respectively, during lowest tides and well-lit conditions. Using Agisoft Metashape, we generated 3D models and digital elevation models of each microatoll and used them to estimate the magnitude of the diedown along the coral’s circumference. Preliminary findings suggest an average diedown of ~7.5cm for the Sentosa microatoll and ~1.5±1cm for the St John’s microatolls. The values of the maximum HLS difference (and standard deviation) on each microatoll were as follows: ~6cm (1.36cm) at Sentosa, ~5cm (1.25cm), ~3cm (0.80cm) and ~3cm (0.68cm) at St John’s. Ongoing work seeks to reconcile our understanding on the importance of light and temperature factors affecting coral growth, in relation to water levels at each site. We are also undertaking efforts to improve and validate the vertical accuracy of the 3D models developed.

We examined the 2D resistivity method detection using synthetic model sets with cavities at six-different depths. Checkerboard tests were also used to assess the spatial resolution of the recovered models. We simulated the conceptual model to synthetically measure the apparent resistivity data for dipole-dipole, pole-dipole, Wenner-Schlumberger, and pole-pole arrays. The synthetically measured resistivity data were inverted to create the geoelectric models. The highest anomaly effect (1.46) and variance (24400 W.m) in resistivity data were recovered by dipole-dipole array, whereas the PP array obtained the lowest anomaly effect (0.60) and variance (2401 W.m) for the shallowest cavity set at 2.2 m depth. The anomaly effect and variance were showed direct dependency on the quality of the inverted models. The dipole-dipole array provided the highest model resolution that shows relatively distinct geometry of the cavity anomalies. On the contrary, the pole-dipole and Wenner-Schlumberger arrays recovered good model resolutions, but it is challenging in determining the correct anomaly geometries with them. The pole-pole array depicted the lowest model resolutions with less clear geometry of the cavity anomalies. The inverted models displayed a reduction in model resolutions, overestimation in anomaly sizes, and deviated from the anomaly positions at deeper depths, which can create ambiguity in resistivity model interpretation. Despite these uncertainties, our modelling specified that the 2D resistivity imaging is a potential technique to study subsurface cavity. Among the commonly used arrays, the dipole-dipole array is the most appropriate for cavity survey. The pole-dipole and Wenner-Schlumberger arrays are adequate while the pole-pole array is the least suitable for cavity study.

Typhoon Morakot brought up to 3 meters of rainfall within a few days in August 2009, leading to numerous landslides in Southern Taiwan. The subsequent erosional processes result in a reduction of normal load at depth. Previous studies have suggested that such transient stress change can trigger small to moderate-size earthquakes. In this study, we examine spatio-temporal changes of microseismicity around the 2009 Typhoon Morakot to better understand the possible triggering relationship between extreme weather events, landslides and sediment removal, and tectonic activities. Because many microearthquakes were missing from the standard earthquake catalogs, we apply matched filter and deep-learning based methods to systematically detect microearthquakes half a year before to one year after Typhoon Morakot. We use 71 stations in the Central Weather Bureau (CWB) and 31508 CWB catalog events as the templates to scan through the continuous seismic data. We detect 7 times as many events as in the standard CWB catalog. After relocation with the double-difference technique, there are 4 times as many events as in the standard CWB catalog. In addition, we train a deep learning model to measure the P-wave first motion polarities for events. Based on these polarities, we calculate the focal mechanisms for 2487 events. We observe a transient decrease of microseismicity during and in the following few weeks following Typhoon Morakot, likely due to high noise levels and station outage caused by the typhoon and associated rainfalls/landslides. This drop of seismicity rate is also observed in deep-learning based earthquake catalogs. We find a possible seismicity rate decrease of the normal-fault events after Morakot in the heavy–rainfall region. However, such changes are not clear for strike-slip and thrust-fault events. In summary, our results do not show a clear change in microseismicity before/after Typhoon Morakot in the analyzed time period.

Polar-orbit satellites such as Landsat, Sentinel, and MODIS have been widely used to estimate regional gross primary production (GPP). However, polar-orbit satellite products have limitations such as many data gaps due to low revisit frequency, and cannot observe the diurnal variation of GPP. Geostationary satellites provide new opportunities to estimate GPP with fine temporal intervals that have been not explored in-depth. Here, we use NIRvP that combines canopy structure (NIRv) and photosynthetically active radiation (PAR) information to estimate GPP. First, we retrieved GK-2A/AMI Nadir BRDF Adjusted reflectance (NBAR) using the 6S atmospheric radiative transfer model and semi-empirical BRDF model to compute vegetation indices. Second, we retrieved GK-2A/AMI PAR (photosynthetically active radiation, PAR) using an atmospheric radiative transfer model (FLiES) which is combined with an artificial neural network (ANN) to compute SW, PAR, and diffuse PAR. Last, we calculated GK-2A/AMI NIRvP with 10-min interval and then integrated it into 30-min, daily, and monthly time scales. Our results showed that the number of observations of GK-2A/AMI were 9 times more at a rice paddy site, 13 times more at an evergreen needleleaf forest site, and 11 times more at a deciduous broadleaf forest site than MODIS NBAR in 2020. In addition, we found that GK-2A NIRvP could track in-situ GPP well from diurnal to seasonal time scale in the three sites. At the regional level, GK-2A/AMI NIRvP had a strong linear relationship with Breathing Earth System Simulator (BESS) GPP over the whole Korean Peninsula at daily to seasonal scales. Our findings highlight that the geostationary satellite help to understand diurnal to seasonal photosynthetic dynamics on a regional scale. Furthermore, geostationary satellites could allow us to monitor the rapid vegetation response to environments which complement the limitation of polar-orbit satellites.

MF-OWHM is software by the U.S. Geological Survey specifically to analyze conjunctive-use of surface water and groundwater. California’s Sustainable Groundwater Management Act of 2014 (SGMA) provides a framework to manage groundwater and empowers local agencies to assess hydrologic issues contributing to water-level declines, storage-loss and subsidence. MF-OWHM’s subsidence package (SUB) calculates four-components of aquifer-compaction and the corresponding release of water from storage. The components can be split into two sets of elastic (recoverable) an inelastic (permanent) compaction that are either instantaneous (occurring-within-a-simulated-timestep) or delayed. SUB delayed-compaction occurs at time intervals longer-than-a-timestep (possibly hundreds of years in the future). In addition to the traditional storage loss from groundwater level declines, the groundwater budget contains the change-in-storage term for the compressible fine-grained interbeds within the aquifers, which can determine if and where land subsidence is occurring.
The California’s Central Valley, about 50,000 km², is one of the most productive agricultural regions in the world. Agricultural demand for irrigation is reliant on surface water and groundwater. In parts of the valley, groundwater pumping has caused severe groundwater level declines, resulting in large storage losses and land subsidence. From this long history of overdraft, many basins in the Central Valley that must comply with SGMA’s “undesirable results”. The application of MF-OWHM to the Central Valley Hydrologic Model (CVHM) helps managers to understand water budgets and determine management strategies that mitigate adverse-impacts and optimize water availability. MF-OWHM CVHM can estimate groundwater budgets and groundwater levels to address the SGMA’s “undesirable results”. Hydraulic properties are estimated using a texture analysis of over 10,000 lithology logs. Observed groundwater level, streamflow, extensometer, continuous GPS, and InSAR data were used to calibrate hydraulic parameters. CVHM can help understand historical changes in groundwater-level, storage, and subsidence and forecast future conditions under different climatic, land-use, and conjunctive surface-water and groundwater-management strategies.

Surface enhancing recharge with rice fields is one of a potential measures to increase the groundwater reserve and reduce the subsidence hazard. For planning the surface enhancing recharge, it is crucial for ones to locate the potential area for surface artificial recharge. In the study we utilize the transient electromagnetic method (TEM) at over 60 sites, in order to quickly delineate the area with less near-surface clay layers in the upper fan of the Choushuichi alluvial fan in central Taiwan. Small coil with a diameters of 2-m and center-loop configuration are used for the TEM surveys. To verify the TEM results, we also conduce the electrical resistivity method with Wenner-Schlumberger configurations at several locations with the TEM surveys. Our electrode interval are 1-m and the total length of each surveys is about 100-m in the electrical resistivity surveys. Borehole logs from the observation in the Choushuichi alluvial fan provides ground trues to further check the results from our interpretations. A preliminary results that mapping the area with less conductive clay layers in the gravel upper fan should provide valuable information for locating the potential surface recharge areas in the Choushuichi alluvial fan.

With the progress of industrial development, the problem of water scarcity has gradually been magnified. How to use water resources more efficiently has become an important topic for many research institutions. Although Taiwan is not like some areas lacking rainfall, due to the uneven distribution of rainfall in time and space, water sources in specific areas and seasons are also severely insufficient. Smart water distribution is particularly important. This study constructs the two-stage stochastic programming model to optimize water resources management. Then the concept of risk preference is also incorporated with the stochastic programming model. 

In this study, dimensionality reduction is applied to analysis spatial and temporal climate data in the Pacific Islands. When working with high-dimensional climate data, it is important to find out meaningful features that could represent the underlying structure of the data. That is, the process of dimensionality reduction. Classical techniques, such as principal component analysis (PCA) or multidimensional scaling (MDS), are simple to implement and efficiently computable for linear dimensionality reduction. However, both methods might show poor performance when dealing with datasets containing nonlinear structure. Therefore, an approach based on manifold learning, called isometric feature mapping (Isomap), is introduced. In Isomap method, distances between all pairs of data points are measured by their geodesic distance instead of Euclidean distance, which represents the true low-dimensional geometry of complex datasets. The first step of Isomap algorithm is constructing the neighberhood graph of data points. Next, shortest paths computed by using Dijkstra’s algorithm is considered as geodesic distances. And last, the geodesic distance matrix is imported into multidimensional scaling (MDS) to create a two-dimension map representing their intrinsic characteristics.

Land subsidence is a phenomenon that currently attracted great notice around the world, especially in populated and large building areas. Studies and predictions of land subsidence rate and understanding the factors that cause soil compaction are widely studied and understood. However, the uncertainty in the geological model, as well as the parameters, are less considered in the literature. This study used GMS (Groundwater Modeling System) to operate the groundwater flow and land subsidence simulations to quantify the geological model uncertainty. A synthetic geological model was constructed as the original model. Different numbers of borehole data extracted from the original model and a combination of the data from resistivity tomography and borehole data were conducted to build several new 3D geological models. 3D transient groundwater flow model (MODFLOW) and subsidence model (SUB) are developed based on the constructed geological models. The geological model uncertainty was assessed by calculating the root mean square error (RMSE) and coefficient of determination (R-Square) using the reconstructed models' results compared to those of the original model. The results showed that using resistivity tomography data with the correction of boreholes data showed a better result than only using borehole data. The results revealed an acceptable RMSEs (7mm) and R-square value (0.8) in land subsidence simulation. The results also indicated that the prescribed boundary condition dramatically affects the results of subsidence simulation, as well as the value of hydraulic conductivity significantly affects the delay behavior in compaction calculation. Based on these results, geologists can assess the appropriate number of boreholes and combine the geophysical results to mitigate the geological model and parameter uncertainties to a specific study area.

Hydraulic tomography (HT) is a new hydraulic test method for characterizing heterogeneous aquifer properties. For the HT test, we need to perform sequential pumping/injection tests, and a group of wells is employed to measure the groundwater pressure variations for every event. The pressure variations are then transformed into aquifer properties using the geostatistics approach. For the traditional HT test, the conventional wells for collecting pressure data often open screens at the single target depth. In other words, a single well can only measure the pressure data in one depth range. On the contrary, a multilevel well has several open-screen at the target depths, and thus the amount of pressure data from the multilevel well can be several times more than that of a traditional well. Accordingly, this study employed a new multilevel well system, developed by fiber Bragg grating technology, to conduct HT at a contaminated site. We then compare the estimate with those from depth-average pressure data, which represents the traditional well. The results show that the parameter estimate from FBG multilevel well is better than that of a conventional well. Results also reveal that HT with a multilevel well system can be the next-generation hydraulic test.