Chih-Hua Chang

This study aimed to enhance wastewater treatment at the Yongkang Water Resource Recovery Center by integrating mechanistic modeling results into machine learning (ML). We combined IoT sensor data from treatment processes with the mechanistic modeling results to develop a hybrid model using ML methods for predicting effluent quality, specifically focusing on chemical oxygen demand. The developed hybrid model offers advantages in the evaluation of the Theil inequality coefficient, mean absolute error, and coefficient of variation and addresses the convergence issues encountered in ML models with IoT sensor data. Guidance provided by the model mitigates the challenges associated with poor-quality IoT data.

Regional water quality mapping is the key practical issue in environmental monitoring. Global regression models transform measured spectral image data to water quality information without the consideration of spatially varying functions. However, it is extremely difficult to find a unified mapping algorithm in multiple reservoirs and lakes. The local model of water quality mapping can estimate water quality parameters effectively in multiple reservoirs using spatial regression. Experiments indicate that both models provide fine water quality mapping in low chlorophyll-a (Chla) concentration water (study area 1; root mean square error, RMSE: 0.435 and 0.413 mg m⁻³ in the best global and local models), whereas the local model provides better goodness-of-fit between the observed and derived Chla concentrations, especially in high-variance Chla concentration water (study area 2; RMSE: 20.75 and 6.49 mg m⁻³ in the best global and local models). In-situ water quality samples are collected and correlated with water surface reflectance derived from Sentinel-2 images. The blue-green band ratio and Maximum Chlorophyll Index (MCI)/Fluorescence Line Height (FLH) are feasible for estimating the Chla concentration in these waterbodies. Considering spatially-varying functions, the local model offers a robust approach for estimating the spatial patterns of Chla concentration in multiple reservoirs. The local model of water quality mapping can greatly improve the estimation accuracy in high-variance Chla concentration waters in multiple reservoirs.

Exposure assessment is a key component in the risk assessment of engineered nanomaterials (ENMs). While direct and quantitative measurements of ENMs in complex environmental matrices remain challenging, environmental fate models (EFMs) can be used alternatively for estimating ENMs’ distributions in the environment. This review describes and assesses the development and capability of EFMs, focusing on surface waters. Our review finds that current engineered nanomaterial (ENM) exposure models can be largely classified into three types: material flow analysis models (MFAMs), multimedia compartmental models (MCMs), and spatial river/watershed models (SRWMs). MFAMs, which is already used to derive predicted environmental concentrations (PECs), can be used to estimate the releases of ENMs as inputs to EFMs. Both MCMs and SRWMs belong to EFMs. MCMs are spatially and/or temporally averaged models, which describe ENM fate processes as intermedia transfer of well-mixed environmental compartments. SRWMs are spatiotemporally resolved models, which consider the variability in watershed and/or stream hydrology, morphology, and sediment transport of river networks. As the foundation of EFMs, we also review the existing and emerging ENM fate processes and their inclusion in recent EFMs. We find that while ENM fate processes, such as heteroaggregation and dissolution, are commonly included in current EFMs, few models consider photoreaction and sulfidation, evaluation of the relative importance of fate processes, and the fate of weathered/transformed ENMs. We conclude the review by identifying the opportunities and challenges in using EFMs for ENMs.

This study investigates community-based landslide mitigation planning. The combination of a landslide susceptibility index (LSI) model, landslide inventory datasets, and field work is used to identify hazard-prone areas in Maolin District, Taiwan. Furthermore, to identify the challenges and opportunities affecting the sustainable development of mountain communities, a pilot survey was conducted in three such communities (Dona Village, Wanshan Village, and Maolin Village). The results reveal that there are two types of significant mass movement in such areas: debris avalanche and debris flow. The results also show that the LSI map and multi-temporal landslide inventory datasets correlate with landslide locations. Meander is identified as an important factor in landslide activity. The questionnaire results show that the residents of the study area lack awareness of and access to information related to landslide activity. Similarly, the local residents favor increased environmental protection, working within their community, and additional government spending in regard to managing geohazards. To increase the resilience of the community, an improved landslide susceptibility map is proposed based on the output of the results. Thus, this research improves upon the process of identifying, supporting, and bettering the management of communities prone to landslides.

Seagrasses are a crucial indicator species of coastal marine ecosystems that provide substratum, shelter, and food for epiphytic algae, invertebrates, and fishes. More accurate mapping of seagrasses is essential for their survival as a long-lasting natural resource. Before reflectance spectra could properly be used as remote sensing endmembers, factors that may obscure the detection of reflectance signals must be assessed. The objectives in this study are to determine the influence of (1) epiphytes, (2) water depth, and (3) seagrass genus on the detection of reflectance spectral signals. The results show that epiphytes significantly dampen bottom-type reflectance throughout most of the visible light spectrum, excluding 670–679 nm; the depth does influence reflectance, with the detection of deeper seagrasses being easier, and as the depth increases, only Heterozostera increase in the exact “red edge” wavelength at which there is a rapid change in the near-infrared (NIR) spectrum. These findings helped improve the detection of seagrass endmembers during remote sensing, thereby helping protect the natural resource of seagrasses.

Seagrasses are a vulnerable and declining coastal habitat, which provide shelter and substrate for aquatic microbiota, invertebrates, and fishes. More accurate mapping of seagrasses is imperative for their sustainability but is hindered by the lack of data on reflectance spectra representing the optical signatures of individual species. Objectives of this study are: (1) To determine distinct characteristics of spectral profiles for sand versus three temperate seagrasses (Posidonia, Amphibolis, and Heterozostera); (2) to evaluate the most efficient derivative analysis method of spectral reflectance profiles for determining benthic types; and to assess the influences of (3) site location and (4) the water column on spectral responses. Results show that 566:689 and 566:600 bandwidth ratios are useful in separating seagrasses from sand and from detritus and algae, respectively; first-derivative reflectance spectra generally is the most efficient method, especially with deconvolution analyses further helping to reveal and isolate 11 key wavelength dimensions; and differences between sites and water column composition, which can include suspended particulate matter, both have no effect on endmembers. These findings helped develop a spectral reflectance library that can be used as an endmember reference for remote sensing, thereby providing continued monitoring, assessment, and management of seagrasses.

This study describes the use of inventory-based landslide susceptibility index (LSI) models based on the selection of causative factors, functional relationships between factors and integration to a hazard-warning model. We first merged landslide inventory data of five typhoon-training events and obtained sets of data representing environmental conditions where landslides are likely to occur. These well-defined data sets are used to select five representative causative factors, i.e. the slope angle, rock strength, drainage, curvature and soil type. Four bivariate statistical model combinations were tested: the linear combination, geometric mean, and two other mixed combinations. As a result, the modulation effects between (i) rock strength and slope and (ii) drainage and curvature were intensified in mixed model 2 (MX2) through factor multiplication. The MX2 LSI was integrated with three multivariate landslide hazard-warning models and tested with four triggering factor rainfall parameter sets. Results lead to the conclusion that threshold models with terrain-influenced rainfall can better identify hazard-warning locations. Modulating factor combinations for the hazard warning can also mirror true environmental conditions, yielding more representative model results. This study improves the method for identifying LSI models for the application to rainfall-triggered landslide hazard models.