Landslide Susceptibility Modeling: An Integrated Novel Method Based on Machine Learning Feature Transformation

Predicting aqueous sorption of organic pollutants on microplastics with machine learning

Microplastics (MPs) are ubiquitously distributed in freshwater systems and they can determine the environmental fate of organic pollutants (OPs) via sorption interaction. However, the diverse physicochemical properties of MPs and the wide range of OP species make a deeper understanding of sorption mechanisms challenging. Traditional isotherm-based sorption models are limited in their universality since they normally only consider the nature and characteristics of either sorbents or sorbates individually. Therefore, only specific equilibrium concentrations or specific sorption isotherms can be used to predict sorption. To systematically evaluate and predict OP sorption under the influence of both MPs and OPs properties, we collected 475 sorption data from peer-reviewed publications and developed a poly-parameter-linear-free-energy-relationship-embedded machine learning method to analyze the collected sorption datasets. Models of different algorithms were compared, and the genetic algorithm and support vector machine hybrid model displayed the best prediction performance (R2 of 0.93 and root-mean-square-error of 0.07). Finally, comparison results of three feature importance analysis tools (forward step wise method, Shapley method, and global sensitivity analysis) showed that chemical properties of MPs, excess molar refraction, and hydrogen-bonding interaction of OPs contribute the most to sorption, reflecting the dominant sorption mechanisms of hydrophobic partitioning, hydrogen bond formation, and π-π interaction, respectively. This study presents a novel sorbate-sorbent-based ML model with a wide applicability to expand our capacity in understanding the complicated process and mechanism of OP sorption on MPs.

Application of Tree-Based Ensemble Models to Landslide Susceptibility Mapping: A Comparative Study

Ensemble machine learning methods have been widely used for modeling landslide susceptibility, but there has been no uniform ensemble method for this problem. The main objective of this study is to compare popular ensemble machine learning-based models and apply them to landslides susceptibility mapping. The selected models include the random forest (RF), which is a typical bagging ensemble model, and three advanced boosting models, namely, adaptive boosting (AB), gradient boosting decision trees (GBDT), and extreme gradient boosting (XGBoost). This study considers 94 landslide points and 12 affecting factors. The data are divided into a training dataset consisting of 70% of the overall data, and a validation dataset, containing the remaining 30% of the data. The models are evaluated using the area under the receiver operating characteristic curve (AUC) and three common performance metrics: sensitivity, specificity, and accuracy. The results indicate that the four ensemble models have an AUC of more than 0.8, suggesting that they can appropriately and accurately predict landslide susceptibility maps. In particular, the XGBoost model achieves the best performance among all models, having a sensitivity of 92.86, specificity of 90.00, and accuracy of 91.38. Furthermore, the bagging model has a sensitivity of 89.29, specificity of 86.67, and accuracy of 87.93, and it is superior to the GBDT, which achieves a sensitivity of 86.21, specificity of 86.21, and accuracy of 86.21, and the AB, reaching a sensitivity of 82.14, specificity of 80.00, and accuracy of 81.03. The results presented in this study indicate that the advanced ensemble model, the XGBoost model, could be a promising tool for the selection of ensemble models for predicting landslide susceptibility mapping.

A Meta-Learning Approach of Optimisation for Spatial Prediction of Landslides

Optimisation plays a key role in the application of machine learning in the spatial prediction of landslides. The common practice in optimising landslide prediction models is to search for optimal/suboptimal hyperparameter values in a number of predetermined hyperparameter configurations based on an objective function, i.e., k-fold cross-validation accuracy. However, the overhead of hyperparameter optimisation can be prohibitive, especially for computationally expensive algorithms. This paper introduces an optimisation approach based on meta-learning for the spatial prediction of landslides. The proposed approach is tested in a dense tropical forested area of Cameron Highlands, Malaysia. Instead of optimising prediction models with a large number of hyperparameter configurations, the proposed approach begins with promising configurations based on several basic and statistical meta-features. The proposed meta-learning approach was tested based on Bayesian optimisation as a hyperparameter tuning algorithm and random forest (RF) as a prediction model. The spatial database was established with a total of 63 historical landslides and 15 conditioning factors. Three RF models were constructed based on (1) default parameters as suggested by the sklearn library, (2) parameters suggested by the Bayesian optimisation (BO), and (3) parameters suggested by the proposed meta-learning approach (BO-ML). Based on five-fold cross-validation accuracy, the Bayesian method achieved the best performance for both the training (0.810) and test (0.802) datasets. The meta-learning approach achieved slightly lower accuracies than the Bayesian method for the training (0.769) and test (0.800) datasets. Similarly, based on F1-score and area under the receiving operating characteristic curves (AUROC), the models with optimised parameters either by the Bayesian or meta-learning methods produced more accurate landslide susceptibility assessment than the model with the default parameters. In the present approach, instead of learning from scratch, the meta-learning would begin with hyperparameter configurations optimal for the most similar previous datasets, which can be considerably helpful and time-saving for landslide modelings.

Autonomous Learning Interactive Features for Hyperspectral Remotely Sensed Data

In the field of remote sensing, most of the feature indexes are obtained based on expert knowledge or domain analysis. With the rapid development of machine learning and artificial intelligence, this method is time-consuming and lacks flexibility, and the indexes obtained cannot be applied to all areas. In order to not rely on expert knowledge and find the effective feature index with regard to a certain material automatically, this paper proposes a data-driven method to learn interactive features for hyperspectral remotely sensed data based on a sparse multiclass logistic regression model. The key point explicitly expresses the interaction relationship between original features as new features by multiplication or division operation in the logistic regression. Through the strong constraint of the L1 norm, the learned features are sparse. The coefficient value of the corresponding features after sparse represents the basis for judging the importance of the features, and the optimal interactive features among the original features. This expression is inspired by the phenomenon that usually the famous indexes we used in remote sensing, like NDVI, NDWI, are the ratio between different spectral bands, and also in statistical regression, the relationship between features is captured by feature value multiplication. Experiments were conducted on three hyperspectral data sets of Pavia Center, Washington DC Mall, and Pavia University. The results for binary classification show that the method can extract the NDVI and NDWI autonomously, and a new type of metal index is proposed in the Pavia University data set. This framework is more flexible and creative than the traditional method based on laboratory research to obtain the key feature and feature interaction index for hyperspectral remotely sensed data.