In recent years, remote sensing researchers have investigated the use of different modalities (or combinations of modalities) for classification tasks. Such modalities can be extracted via a diverse range of sensors and images. Currently, there are no (or only a few) studies that have been done to increase the land cover classification accuracy via unmanned aerial vehicle (UAV)–digital surface model (DSM) fused datasets. Therefore, this study looks at improving the accuracy of these datasets by exploiting convolutional neural networks (CNNs). In this work, we focus on the fusion of DSM and UAV images for land use/land cover mapping via classification into seven classes: bare land, buildings, dense vegetation/trees, grassland, paved roads, shadows, and water bodies. Specifically, we investigated the effectiveness of the two datasets with the aim of inspecting whether the fused DSM yields remarkable outcomes for land cover classification. The datasets were: (i) only orthomosaic image data (Red, Green and Blue channel data), and (ii) a fusion of the orthomosaic image and DSM data, where the final classification was performed using a CNN. CNN, as a classification method, is promising due to hierarchical learning structure, regulating and weight sharing with respect to training data, generalization, optimization and parameters reduction, automatic feature extraction and robust discrimination ability with high performance. The experimental results show that a CNN trained on the fused dataset obtains better results with Kappa index of ~0.98, an average accuracy of 0.97 and final overall accuracy of 0.98. Comparing accuracies between the CNN with DSM result and the CNN without DSM result for the overall accuracy, average accuracy and Kappa index revealed an improvement of 1.2%, 1.8% and 1.5%, respectively. Accordingly, adding the heights of features such as buildings and trees improved the differentiation between vegetation specifically where plants were dense.

Assessment of the most appropriate groundwater conditioning factors (GCFs) is essential when performing analyses for groundwater potential mapping. For this reason, in this work, we look at three statistical factor analysis methods—Variance Inflation Factor (VIF), Chi-Square Factor Optimization, and Gini Importance—to measure the significance of GCFs. From a total of 15 frequently used GCFs, 11 most effective ones (i.e., altitude, slope angle, plan curvature, profile curvature, topographic wetness index, distance from river, distance from fault, river density, fault density, land use, and lithology) were finally selected. In addition, 917 spring locations were identified and used to train and test three machine learning algorithms, namely Mixture Discriminant Analysis (MDA), Linear Discriminant Analysis (LDA) and Random Forest (RF). The resultant trained models were then applied for groundwater potential prediction and mapping in the Haraz basin of Mazandaran province, Iran. MDA has been successfully applied for soil erosion and landslide mapping, but has not yet been fully explored for groundwater potential mapping (GPM). Although other discriminant methods, such as LDA, exist, MDA is worth exploring due to its capability to model multivariate nonlinear relationships between variables; it also undertakes a mixture of unobserved subclasses with regularization of non-linear decision boundaries, which could potentially provide more accurate classification. For the validation, areas under Receiver Operating Characteristics (ROC) curves (AUC) were calculated for the three algorithms. RF performed better with AUC value of 84.4%, while MDA and LDA yielded 75.2% and 74.9%, respectively. Although MDA performance is lower than RF, the result is satisfactory, because it is within the acceptable standard of environmental modeling. The outcome of factor analysis and groundwater maps emphasizes on optimization of multicolinearity factors for faster spatial modeling and provides valuable information for government agencies and private sectors to effectively manage groundwater in the region.

Landslide susceptibility mapping has significantly progressed with improvements in machine learning techniques. However, the inventory/data imbalance (DI) problem remains one of the challenges in this domain. This problem exists as a good quality landslide inventory map, including a complete record of historical data, is difficult or expensive to collect. As such, this can considerably affect one’s ability to obtain a sufficient inventory or representative samples. This research developed a new approach based on generative adversarial networks (GAN) to correct imbalanced landslide datasets. The proposed method was tested at Chukha Dzongkhag, Bhutan, one of the most frequent landslide prone areas in the Himalayan region. The proposed approach was then compared with the standard methods such as the synthetic minority oversampling technique (SMOTE), dense imbalanced sampling, and sparse sampling (i.e., producing non-landslide samples as many as landslide samples). The comparisons were based on five machine learning models, including artificial neural networks (ANN), random forests (RF), decision trees (DT), k-nearest neighbours (kNN), and the support vector machine (SVM). The model evaluation was carried out based on overall accuracy (OA), Kappa Index, F1-score, and area under receiver operating characteristic curves (AUROC). The spatial database was established with a total of 269 landslides and 10 conditioning factors, including altitude, slope, aspect, total curvature, slope length, lithology, distance from the road, distance from the stream, topographic wetness index (TWI), and sediment transport index (STI). The findings of this study have shown that both GAN and SMOTE data balancing approaches have helped to improve the accuracy of machine learning models. According to AUROC, the GAN method was able to boost the models by reaching the maximum accuracy of ANN (0.918), RF (0.933), DT (0.927), kNN (0.878), and SVM (0.907) when default parameters used. With the optimum parameters, all models performed best with GAN at their highest accuracy of ANN (0.927), RF (0.943), DT (0.923) and kNN (0.889), except SVM obtained the highest accuracy of (0.906) with SMOTE. Our finding suggests that RF balanced with GAN can provide the most reasonable criterion for landslide prediction. This research indicates that landslide data balancing may substantially affect the predictive capabilities of machine learning models. Therefore, the issue of DI in the spatial prediction of landslides should not be ignored. Future studies could explore other generative models for landslide data balancing. By using state-of-the-art GAN, the proposed model can be considered in the areas where the data are limited or imbalanced.

Landslide susceptibility modeling, an essential approach to mitigate natural disasters, has witnessed considerable improvement following advances in machine learning (ML) techniques. However, in most of the previous studies, the distribution of input data was assumed as being, and treated, as normal or Gaussian; this assumption is not always valid as ML is heavily dependent on the quality of the input data. Therefore, we examine the effectiveness of six feature transformations (minimax normalization (Std-X), logarithmic functions (Log-X), reciprocal function (Rec-X), power functions (Power-X), optimal features (Opt-X), and one-hot encoding (Ohe-X) over the 11conditioning factors (i.e., altitude, slope, aspect, curvature, distance to road, distance to lineament, distance to stream, terrain roughness index (TRI), normalized difference vegetation index (NDVI), land use, and vegetation density). We selected the frequent landslide-prone area in the Cameron Highlands in Malaysia as a case study to test this novel approach. These transformations were then assessed by three benchmark ML methods, namely extreme gradient boosting (XGB), logistic regression (LR), and artificial neural networks (ANN). The 10-fold cross-validation method was used for model evaluations. Our results suggest that using Ohe-X transformation over the ANN model considerably improved performance from 52.244 to 89.398 (37.154% improvement).

In recent years, remote-sensing (RS) technologies have been used together with image processing and traditional techniques in various disaster-related works. Among these is detecting building damage from orthophoto imagery that was inflicted by earthquakes. Automatic and visual techniques are considered as typical methods to produce building damage maps using RS images. The visual technique, however, is time-consuming due to manual sampling. The automatic method is able to detect the damaged building by extracting the defect features. However, various design methods and widely changing real-world conditions, such as shadow and light changes, cause challenges to the extensive appointing of automatic methods. As a potential solution for such challenges, this research proposes the adaption of deep learning (DL), specifically convolutional neural networks (CNN), which has a high ability to learn features automatically, to identify damaged buildings from pre- and post-event RS imageries. Since RS data revolves around imagery, CNNs can arguably be most effective at automatically discovering relevant features, avoiding the need for feature engineering based on expert knowledge. In this work, we focus on RS imageries from orthophoto imageries for damaged-building detection, specifically for (i) background, (ii) no damage, (iii) minor damage, and (iv) debris classifications. The gist is to uncover the CNN architecture that will work best for this purpose. To this end, three CNN models, namely the twin model, fusion model, and composite model, are applied to the pre- and post-orthophoto imageries collected from the 2016 Kumamoto earthquake, Japan. The robustness of the models was evaluated using four evaluation metrics, namely overall accuracy (OA), producer accuracy ($\mathrm{PA}$), user accuracy ($\mathrm{UA}$), and F1 score. According to the obtained results, the twin model achieved higher accuracy (OA = 76.86%; F1 score = 0.761) compare to the fusion model (OA = 72.27%; F1 score = 0.714) and composite (OA = 69.24%; F1 score = 0.682) models.

Zero-shot learning (ZSL) is an approach to classify objects unseen during the training phase and shown to be useful for real-world applications, especially when there is a lack of sufficient training data. Only a limited amount of works has been carried out on ZSL, especially in the field of remote sensing. This research investigates the use of a convolutional neural network (CNN) as a feature extraction and classification method for land cover mapping using high-resolution orthophotos. In the feature extraction phase, we used a CNN model with a single convolutional layer to extract discriminative features. In the second phase, we used class attributes learned from the Word2Vec model (pre-trained by Google News) to train a second CNN model that performed class signature prediction by using both the features extracted by the first CNN and class attributes during training and only the features during prediction. We trained and tested our models on datasets collected over two subareas in the Cameron Highlands (training dataset, first test dataset) and Ipoh (second test dataset) in Malaysia. Several experiments have been conducted on the feature extraction and classification models regarding the main parameters, such as the network’s layers and depth, number of filters, and the impact of Gaussian noise. As a result, the best models were selected using various accuracy metrics such as top-k categorical accuracy for k = [1,2,3], Recall, Precision, and F1-score. The best model for feature extraction achieved 0.953 F1-score, 0.941 precision, 0.882 recall for the training dataset and 0.904 F1-score, 0.869 precision, 0.949 recall for the first test dataset, and 0.898 F1-score, 0.870 precision, 0.838 recall for the second test dataset. The best model for classification achieved an average of 0.778 top-one, 0.890 top-two and 0.942 top-three accuracy, 0.798 F1-score, 0.766 recall and 0.838 precision for the first test dataset and 0.737 top-one, 0.906 top-two, 0.924 top-three, 0.729 F1-score, 0.676 recall and 0.790 precision for the second test dataset. The results demonstrated that the proposed ZSL is a promising tool for land cover mapping based on high-resolution photos.