Despite yielding considerable degrees of accuracy in landslide predictions, the outcomes of different landslide susceptibility models are prone to spatial disagreement; and therefore, uncertainties. Uncertainties in the results of various landslide susceptibility models create challenges in selecting the most suitable method to manage this complex natural phenomenon. This study aimed to propose an approach to reduce uncertainties in landslide prediction, diagnosing spatial agreement in machine learning-based landslide susceptibility maps. It first developed landslide susceptibility maps of Cox’s Bazar district of Bangladesh, applying four machine learning algorithms: K-Nearest Neighbor (KNN), Multi-Layer Perceptron (MLP), Random Forest (RF), and Support Vector Machine (SVM), featuring hyperparameter optimization of 12 landslide conditioning factors. The results of all the four models yielded very high prediction accuracy, with the area under the curve (AUC) values range between 0.93 to 0.96. The assessment of spatial agreement of landslide predictions showed that the pixel-wise correlation coefficients of landslide probability between various models range from 0.69 to 0.85, indicating the uncertainty in predicted landslides by various models, despite their considerable prediction accuracy. The uncertainty was addressed by establishing a Logistic Regression (LR) model, incorporating the binary landslide inventory data as the dependent variable and the results of the four landslide susceptibility models as independent variables. The outcomes indicated that the RF model had the highest influence in predicting the observed landslide locations, followed by the MLP, SVM, and KNN models. Finally, a combined landslide susceptibility map was developed by integrating the results of the four machine learning-based landslide predictions. The combined map resulted in better spatial agreement (correlation coefficients range between 0.88 and 0.92) and greater prediction accuracy (0.97) compared to the individual models. The modelling approach followed in this study would be useful in minimizing uncertainties of various methods and improving landslide predictions.

Since the dawn of human civilization, forced migration scenarios have been witnessed in different regions and populations, and is still present in the twenty-first century. The current largest population of stateless refugees in the world, the Rohingya people, reside in the southeastern border region of Bangladesh. Due to rapid expansion of refugee camps and lack of suitable locations, a large proportion of the infrastructure are at risk of landslides. This study aims to use machine learning for predicting landslide risk of camp infrastructure using geospatial features. Four supervised classification algorithms have been employed viz., (i) Logistic Regression (LR), (ii) Multi-Layer Perceptron (MLP), (iii) Gradient Boosted Trees (GBT) and (iv) Random Forest (RF) and applied on preprocessing varied versions of features. Results show that RF achieves accuracy of 76.19% and AUC of 0.76 on un-scaled features which is higher than all other algorithms. The applications of the study reside in refugee management and landslide susceptibility mapping of Rohingya camps, which can both potentially save refugee lives and serve as a case study for global applications.

Compared with numerous existing forced migration scenarios across the globe, migration from Myanmar to Bangladesh through southeastern border region is unique at least for three reasons – (i) very large number of migrants have been displaced to (ii) a very small area in (iii) a relatively short period of time, creating an obvious cumulative impact on forest cover area of the host country. Therefore, this study aims to analyze the dynamics of refugee migration and deforestation in Bangladesh. Satellite images of Landsat-5 & 8 and Sentinel-2 were classified via four different classification algorithms (SVM, Random Forest, CART, and Max Entropy) to measure major land use and land cover changes, namely, (i) dense forest, (ii) sparse forest, (iii) open area, and (iv) settlement from 1988 to 2018. The analysis revealed a declining trend of dense forest area, majority of which took place from 2016 to 2018 triggered by Rohingya migration. As a whole, the dense forest cover has been effectively halved (8531 ha in 2016 to 4498 ha in 2018) in the span of just two years while refugee settlement has increased nine-folds (271 ha in 2016 to 2679 ha in 2018). Aggregated and indisputable conclusion has been derived indicating that forced Rohingya migration and deforestation are indeed positively correlated.

Label noise is a commonly encountered problem in learning building extraction tasks; its presence can reduce performance and increase learning complexity. This is especially true for cases where high resolution aerial drone imagery is used, as the labels may not perfectly correspond/align with the actual objects in the imagery. In general machine learning and computer vision context, labels refer to the associated class of data, and in remote sensing-based building extraction refer to pixel-level classes. Dense label noise in building extraction tasks has rarely been formalized and assessed. We formulate a taxonomy of label noise models for building extraction tasks, which incorporates both pixel-wise and dense models. While learning dense prediction under label noise, the differences between the ground truth clean label and observed noisy label can be encoded by error matrices indicating locations and type of noisy pixel-level labels. In this work, we explicitly learn to approximate error matrices for improving building extraction performance; essentially, learning dense prediction of label noise as a subtask of a larger building extraction task. We propose two new model frameworks for learning building extraction under dense real-world label noise, and consequently two new network architectures, which approximate the error matrices as intermediate predictions. The first model learns the general error matrix as an intermediate step and the second model learns the false positive and false-negative error matrices independently, as intermediate steps. Approximating intermediate error matrices can generate label noise saliency maps, for identifying labels having higher chances of being mis-labelled. We have used ultra-high-resolution aerial images, noisy observed labels from OpenStreetMap, and clean labels obtained after careful annotation by the authors. When compared to the baseline model trained and tested using clean labels, our intermediate false positive-false negative error matrix model provides Intersection-Over-Union gain of 2.74% and F1-score gain of 1.75% on the independent test set. Furthermore, our proposed models provide much higher recall than currently used deep learning models for building extraction, while providing comparable precision. We show that intermediate false positive-false negative error matrix approximation can improve performance under label noise.