Earth Science Informatics最新文献

Understanding the spatial distribution of forest properties can help improve our knowledge of carbon storage and the impacts of climate change. Despite the active use of remote sensing and machine learning (ML) methods in forest mapping, the associated uncertainty predictions are relatively uncommon. The objectives of this study were: (1) to evaluate the spatial resolution effect on growing stock volume (GSV) mapping using Sentinel-2A and Landsat 8 satellite images, (2) to identify the most key predictors, and (3) to quantify the uncertainty of GSV predictions. The study was conducted in heterogeneous landscapes, covering anthropogenic areas, logging, young plantings and mature trees. We employed an ML approach and evaluated our models by root mean squared error (RMSE) and coefficient of determination (R²) through a 10-fold cross-validation. Our results indicated that the Sentinel-2A provided the best prediction performances (RMSE = 56.6 m³/ha, R² = 0.53) in compare with Landsat 8 (RMSE = 71.2 m³/ha, R² = 0.23), where NDVI, LSWI and B08 band (near-infrared spectrum) were identified as key variables, with the highest contribution to the model. Moreover, the uncertainty of GSV predictions using the Sentinel-2A was much smaller compared with Landsat 8. The combined assessment of accuracy and uncertainty reinforces the suitability of Sentinel-2A for applications in heterogeneous landscapes. The higher accuracy and lower uncertainty observed with the Sentinel-2A underscores its effectiveness in providing more reliable and precise information for decision-makers. This research is important for further digital mapping endeavours with accompanying uncertainty, as uncertainty assessment plays a pivotal role in decision-making processes related to spatial assessment and forest management.

The underwater images (UWIs) are one of the most effective sources to collect information about the underwater environment. Due to the irregular optical properties of different water types, the captured UWIs suffer from color cast, low visibility and distortion. Moreover, each water type offers different optical absorption, scattering, and attenuation of red, green and blue bands, which makes restoration of UWIs a challenging task. The revised underwater image formation model (RUIFM) considers only the peak values of the corresponding attenuation coefficient of each water type to restore UWIs. The performance of RUIFM suffers due to the inter-class variations of UWIs in a water type. In this paper, we propose an improved version of RUIFM as the Diverse Underwater Image Formation Model (DUIFM). The DUIFM increases the diversity of RUIFM by deeply encountering the optical properties of each water type. We investigate the inter-class variations of Jerlov-based classes of UWIs in terms of light attenuation and statistical features and further classify each image into low, medium and high bands. Which, in turn, provides the precise inherent optical attenuation coefficient of water and increases the generality of the DUIFM in color restoration and enhancement. The qualitative and quantitative performance evaluation results on publicly available real-world underwater enhancement (RUIE), underwater image enhancement benchmark (UIEB) and enhanced underwater visual perception (EUVP) data sets demonstrate the effectiveness of our proposed DUIFM.

With the continuous development of power system and the growth of load demand, accurate short-term load forecasting (SLTF) provides reliable guidance for power system operation and scheduling. Therefore, this paper proposes a two-stage short-term load forecasting method. In the first stage, the original load is processed by improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN). The time series features of the load are extracted by temporal convolutional network (TCN), which is used as an input to realize the initial load prediction based on gated recurrent unit (GRU). At the same time, in order to overcome the problem that the prediction model established by the original subsequence has insufficient adaptability in the newly decomposed subsequence, the real-time decomposition strategy is adopted to improve the generalization ability of the model. To further improve the prediction accuracy, an error compensation strategy is constructed in the second stage. The strategy uses adaptive variational mode decomposition (AVMD) to reduce the unpredictability of the error sequence and corrects the initial prediction results based on the temporal convolutional network-gated recurrent unit (TCN-GRU) error compensator. The proposed two-stage forecasting method was evaluated using load data from Queensland, Australia. The analysis results show that the proposed method can better capture the nonlinearity and non-stationarity in the load data. The mean absolute percentage error of its prediction is 0.819%, which are lower than the other compared models, indicating its high applicability in SLTF.

The rapid and uncontrolled development in the urban environment leads to significant problems, negatively affecting the quality of life in many areas. Smart Sustainable City concept has emerged to solve these problems and enhance the quality of life for its citizens. A smart city integrates the physical, digital and social system in order to provide a sustainable and comfortable future by the help of the Information and Communication Technologies (ICT) and Spatial Data Infrastructures (SDI). However, the integrated management of urban data requires the inclusion of ICT enabled SDI that can be applied as a decision support element in different urban problems by giving a comprehensive understanding of city dynamics; an interoperable and integrative conceptual data modelling, essential for smart sustainable cities and successful management of big urban data. The main purpose of this study is to propose an integrated data management approach in accordance with international standards for sustainable management of smart cities. Thematic data model designed within the scope of quality of life, which is one of the main purposes of smart cities, offers an exemplary approach to overcome the problems arising from the inability to manage and analyse big and complex urban data for sustainability. In this aspect, it is aimed to provide a conceptual methodology for successful implementation of smart sustainable city applications within the international and national SDIs with environmental quality of life theme. With this object, firstly, the literature on smart sustainable cities was examined together with the scope of quality and sustainability of urban environment along with all related components. Secondly, the big data and its management was examined within the concept of the urban SDI. In this perspective, new trends and standards related to sensors, internet of things (IoT), real-time data, online services and application programming interfaces (API) were investigated. After, thematic conceptual models for the integrated management of sensor-based data were proposed and a real time Air Quality Index (AQI) dashboard was designed in Istanbul, Türkiye as the thematic case application of proposed models.

Soil salinity is one of the significant environmental issues that can reduce crop growth and productivity, ultimately leading to land degradation. Therefore, accurate monitoring and mapping of soil salinity are essential for implementing effective measures to combat increasing salinity. This study aims to estimate the spatial distribution of soil salinity using machine learning methods in Huludao City, located in northeastern China. By meticulously collecting data, soil salinity was measured in 310 soil samples. Subsequently, environmental parameters were calculated using remote sensing data. In the next step, soil salinity was modeled using machine learning methods, including random forest (RF), support vector machine (SVM), and artificial neural network (ANN). Additionally, to estimate uncertainty, the lower limit (5%) and upper limit (95%) prediction intervals were used. The results indicated that accurate maps for predicting soil salinity could be obtained using machine learning methods. By comparing the methods employed, it was determined that the RF model is the most accurate approach for estimating soil salinity (RMSE=0.03, AIC=-919, BIS=-891, and R²=0.84). Furthermore, the results from the prediction interval coverage probability (PICP) index, utilizing the uncertainty maps, demonstrated the high predictive accuracy of the methods employed in this study. Moreover, it was revealed that the environmental parameters, including NDVI, GNDVI, standh, and BI, are the main controllers of the spatial patterns of soil salinity in the study area. However, there remains a need to explore more precise methods for estimating soil salinity and identifying salinity patterns, as soil salinity has intensified with increased human activities, necessitating more detailed investigations.

The practice of categorizing the galaxies according to morphologies exists and offers crucial details on the creation and development of the universe. The conventional visual inspection techniques have been very subjective and time-consuming. However, it is now possible to classify galaxies with greater accuracy owing to advancements in deep learning techniques. Deep Learning has demonstrated considerable potential in the research of galaxy classification and offers fresh perspectives on the genesis and evolution of galaxies. The suggested methodology employs Residual Networks for variety in a transfer learning-based method. To improve the accuracy of ResNet, an attention mechanism has been included. In our investigation, we used two relatively shallow ResNet models, ResNet18 and ResNet50 by incorporating a soft attention mechanism into them. The presented approach is validated on the Galaxy Zoo dataset from Kaggle. The accuracy increases from 60.15% to 80.20% for ResNet18 and from 78.21% to 80.55% for ResNet50, thus, demonstrating that the proposed work is now on a level with the accuracy of the far more complex, ResNet152 model. We have found that the attention mechanism can successfully improve the accuracy of even shallow models, which has implications for future studies in image recognition tasks.

Hyperspectral Imaging (HSI) has revolutionized earth observation through advanced remote sensing technology, providing rich spectral and spatial information across multiple bands. However, this wealth of data introduces significant challenges for classification, including high spectral correlation, the curse of dimensionality due to limited labeled data, the need to model long-term dependencies, and the impact of sample input on deep learning performance. These challenges are further exacerbated by the costly and complex acquisition of HSI data, resulting in limited availability of labeled samples and class imbalances. To address these critical issues, our study proposes a novel approach for generating high-quality synthetic hyperspectral data cubes using an advanced Generative Adversarial Network (GAN) integrated with the Wasserstein loss and gradient penalty phenomenon (WGAN-GP). This approach aims to augment real-world data, mitigating the scarcity of labeled samples that has long been a bottleneck in hyperspectral image analysis and classification. To fully leverage both the synthetic and real data, we introduce a novel Convolutional LSTM classifier designed to process the intricate spatial and spectral correlations inherent in hyperspectral data. This classifier excels in modeling multi-dimensional relationships within the data, effectively capturing long-term dependencies and improving feature extraction and classification accuracy. The performance of our proposed model, termed 3D-ACWGAN-ConvLSTM, is rigorously validated using benchmark hyperspectral datasets, demonstrating its effectiveness in augmenting real-world data and enhancing classification performance. This research contributes to addressing the critical need for robust data augmentation techniques in hyperspectral imaging, potentially opening new avenues for applications in areas constrained by limited data availability and complex spectral-spatial relationships.

Geological domaining is an essential aspect of mineral resource evaluation. Various explicit and implicit modeling approaches have been developed for this purpose, but most of them are computationally expensive and complex, particularly when dealing with intricate mineralization systems and large datasets. Additionally, most of them require a time-consuming process for hyperparameter tuning. In this research, the application of the Learning Vector Quantization (LVQ) classification algorithm has been proposed to address these challenges. The LVQ algorithm exhibits lower complexity and computational costs compared to other machine learning algorithms. Various versions of LVQ, including LVQ1, LVQ2, and LVQ3, have been implemented for geological domaining in the Darehzar porphyry copper deposit in southeastern Iran. Their performance in geological domaining has been thoroughly investigated and compared with the Support Vector Machine (SVM), a widely accepted classification method in implicit domaining. The overall classification accuracy of LVQ1, LVQ2, LVQ3, and SVM is 90%, 90%, 91%, and 98%, respectively. Furthermore, the calculation time of these algorithms has been compared. Although the overall accuracy of the SVM method is ∼ 7% higher, its calculation time is ∼ 1000 times longer than LVQ methods. Therefore, LVQ emerges as a suitable alternative for geological domaining, especially when dealing with large datasets.

Accurate hourly streamflow prediction is crucial for managing water resources, particularly in smaller basins with short response times. This study evaluates six deep learning (DL) models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and their hybrids (CNN-LSTM, CNN-GRU, CNN-Recurrent Neural Network (RNN)), across two basins in Northwest Spain over a ten-year period. Findings reveal that GRU models excel, achieving Nash-Sutcliffe Efficiency (NSE) scores of approximately 0.96 and 0.98 for the Groba and Anllóns catchments, respectively, at 1-hour lead times. Hybrid models did not enhance performance, which declines at longer lead times due to basin-specific characteristics such as area and slope, particularly in smaller basins where NSE dropped from 0.969 to 0.24. The inclusion of future rainfall data in the input sequences has improved the results, especially for longer lead times from 0.24 to 0.70 in the Groba basin and from 0.81 to 0.92 in the Anllóns basin for a 12-hour lead time. This research provides a foundation for future exploration of DL in streamflow forecasting, in which other data sources and model structures can be utilized.

Floods, among the most destructive climate-induced natural disasters, necessitate effective prediction models for early warning systems. The proposed Multi-Attention Encoder-Decoder-based Temporal Convolutional Network (MA-TCN-ED) prediction model combines the strengths of the Temporal Convolutional Network (TCN), Multi-Attention (MA) mechanism, and Encoder-Decoder (ED) architecture along with filter-wrapper feature selection for optimal feature selection. This framework aims to improve flood prediction accuracy by effectively capturing temporal dependencies and intricate patterns in atmospheric and hydro-meteorological data. The proposed framework was pervasively assessed for predicting the real-world flood-related data of the river Meenachil, Kerala, and the results showed that MA-TCN-ED using a filter-wrapper feature selection approach achieved higher accuracy in flood prediction. Further the model was validated on the dataset of river Pamba, Kerala. The proposed model exhibits better performance with about 32% reduced MAE, 39% reduced RMSE, 12% increased NSE, 14% enhanced R², and 17% enhanced accuracy relative to the average performance of all the compared baseline models. The proposed work holds promise for enhancing early warning systems and mitigating the impact of floods and contributes to the broader understanding of leveraging deep learning models for effective climate-related risk mitigation.