Super Learner for Survival Data Prediction

Survival analysis is a widely used method to establish a connection between a time to event outcome and a set of potential covariates. Accurately predicting the time of an event of interest is of primary importance in survival analysis. Many different algorithms have been proposed for survival prediction. However, for a given prediction problem it is rarely, if ever, possible to know in advance which algorithm will perform the best. In this paper we propose two algorithms for constructing super learners in survival data prediction where the individual algorithms are based on proportional hazards. A super learner is a flexible approach to statistical learning that finds the best weighted ensemble of the individual algorithms. Finding the optimal combination of the individual algorithms through minimizing cross-validated risk controls for over-fitting of the final ensemble learner. Candidate algorithms may range from a basic Cox model to tree-based machine learning algorithms, assuming all candidate algorithms are based on the proportional hazards framework. The ensemble weights are estimated by minimizing the cross-validated negative log partial likelihood. We compare the performance of the proposed super learners with existing models through extensive simulation studies. In all simulation scenarios, the proposed super learners are either the best fit or near the best fit. The performances of the newly proposed algorithms are also demonstrated with clinical data examples.

Evaluation of Machine Learning Algorithms for Surface Water Extraction in a Landsat 8 Scene of Nepal

With over 6000 rivers and 5358 lakes, surface water is one of the most important resources in Nepal. However, the quantity and quality of Nepal's rivers and lakes are decreasing due to human activities and climate change. Despite the advancement of remote sensing technology and the availability of open access data and tools, the monitoring and surface water extraction works has not been carried out in Nepal. Single or multiple water index methods have been applied in the extraction of surface water with satisfactory results. Extending our previous study, the authors evaluated six different machine learning algorithms: Naive Bayes (NB), recursive partitioning and regression trees (RPART), neural networks (NNET), support vector machines (SVM), random forest (RF), and gradient boosted machines (GBM) to extract surface water in Nepal. With three secondary bands, slope, NDVI and NDWI, the algorithms were evaluated for performance with the addition of extra information. As a result, all the applied machine learning algorithms, except NB and RPART, showed good performance. RF showed overall accuracy (OA) and kappa coefficient (Kappa) of 1 for the all the multiband data with the reference dataset, followed by GBM, NNET, and SVM in metrics. The performances were better in the hilly regions and flat lands, but not well in the Himalayas with ice, snow and shadows, and the addition of slope and NDWI showed improvement in the results. Adding single secondary bands is better than adding multiple in most algorithms except NNET. From current and previous studies, it is recommended to separate any study area with and without snow or low and high elevation, then apply machine learning algorithms in original Landsat data or with the addition of slopes or NDWI for better performance.

Comparison between Random Forests, Artificial Neural Networks and Gradient Boosted Machines Methods of On-Line Vis-NIR Spectroscopy Measurements of Soil Total Nitrogen and Total Carbon

Accurate and detailed spatial soil information about within-field variability is essential for variable-rate applications of farm resources. Soil total nitrogen (TN) and total carbon (TC) are important fertility parameters that can be measured with on-line (mobile) visible and near infrared (vis-NIR) spectroscopy. This study compares the performance of local farm scale calibrations with those based on the spiking of selected local samples from both fields into an European dataset for TN and TC estimation using three modelling techniques, namely gradient boosted machines (GBM), artificial neural networks (ANNs) and random forests (RF). The on-line measurements were carried out using a mobile, fiber type, vis-NIR spectrophotometer (305-2200 nm) (AgroSpec from tec5, Germany), during which soil spectra were recorded in diffuse reflectance mode from two fields in the UK. After spectra pre-processing, the entire datasets were then divided into calibration (75%) and prediction (25%) sets, and calibration models for TN and TC were developed using GBM, ANN and RF with leave-one-out cross-validation. Results of cross-validation showed that the effect of spiking of local samples collected from a field into an European dataset when combined with RF has resulted in the highest coefficients of determination (R²) values of 0.97 and 0.98, the lowest root mean square error (RMSE) of 0.01% and 0.10%, and the highest residual prediction deviations (RPD) of 5.58 and 7.54, for TN and TC, respectively. Results for laboratory and on-line predictions generally followed the same trend as for cross-validation in one field, where the spiked European dataset-based RF calibration models outperformed the corresponding GBM and ANN models. In the second field ANN has replaced RF in being the best performing. However, the local field calibrations provided lower R² and RPD in most cases. Therefore, from a cost-effective point of view, it is recommended to adopt the spiked European dataset-based RF/ANN calibration models for successful prediction of TN and TC under on-line measurement conditions.