Forecasting Delivery Pattern through Floating Car Data: Empirical Evidence

This paper investigates the opportunities offered by floating car data (FCD) to infer delivering activities. A discrete trip-chain order model (within the random utility theory) for light goods vehicles (laden weight less than 3.5 tons) is hence proposed, which characterizes delivery tours in terms of the number of stops/deliveries performed. Thus, the main goal of the study is to calibrate a discrete choice model to estimate the number of stops/deliveries per tour by using FCD, which can be incorporated in a planning procedure for obtaining a preliminary assessment of parking demand. The data used refer to light goods vehicles operating in the Veneto region. The database contains more than 8000 tours undertaken in 60 working days. Satisfactory results have been obtained in terms of tour estimation and model transferability.

Importance of Spatial Autocorrelation in Machine Learning Modeling of Polymetallic Nodules, Model Uncertainty and Transferability at Local Scale

Machine learning spatial modeling is used for mapping the distribution of deep-sea polymetallic nodules (PMN). However, the presence and influence of spatial autocorrelation (SAC) have not been extensively studied. SAC can provide information regarding the variable selection before modeling, and it results in erroneous validation performance when ignored. ML models are also problematic when applied in areas far away from the initial training locations, especially if the (new) area to be predicted covers another feature space. Here, we study the spatial distribution of PMN in a geomorphologically heterogeneous area of the Peru Basin, where SAC of PMN exists. The local Moran’s I analysis showed that there are areas with a significantly higher or lower number of PMN, associated with different backscatter values, aspect orientation, and seafloor geomorphological characteristics. A quantile regression forests (QRF) model is used using three cross-validation (CV) techniques (random-, spatial-, and cluster-blocking). We used the recently proposed “Area of Applicability” method to quantify the geographical areas where feature space extrapolation occurs. The results show that QRF predicts well in morphologically similar areas, with spatial block cross-validation being the least unbiased method. Conversely, random-CV overestimates the prediction performance. Under new conditions, the model transferability is reduced even on local scales, highlighting the need for spatial model-based dissimilarity analysis and transferability assessment in new areas.

A Test of Species Distribution Model Transferability Across Environmental and Geographic Space for 108 Western North American Tree Species

Predictions from species distribution models (SDMs) are commonly used in support of environmental decision-making to explore potential impacts of climate change on biodiversity. However, because future climates are likely to differ from current climates, there has been ongoing interest in understanding the ability of SDMs to predict species responses under novel conditions (i.e., model transferability). Here, we explore the spatial and environmental limits to extrapolation in SDMs using forest inventory data from 11 model algorithms for 108 tree species across the western United States. Algorithms performed well in predicting occurrence for plots that occurred in the same geographic region in which they were fitted. However, a substantial portion of models performed worse than random when predicting for geographic regions in which algorithms were not fitted. Our results suggest that for transfers in geographic space, no specific algorithm was better than another as there were no significant differences in predictive performance across algorithms. There were significant differences in predictive performance for algorithms transferred in environmental space with GAM performing best. However, the predictive performance of GAM declined steeply with increasing extrapolation in environmental space relative to other algorithms. The results of this study suggest that SDMs may be limited in their ability to predict species ranges beyond the environmental data used for model fitting. When predicting climate-driven range shifts, extrapolation may also not reflect important biotic and abiotic drivers of species ranges, and thus further misrepresent the realized shift in range. Future studies investigating transferability of process based SDMs or relationships between geodiversity and biodiversity may hold promise.

Mapping Rice Paddy Based on Machine Learning with Sentinel-2 Multi-Temporal Data: Model Comparison and Transferability

Rice is an important agricultural crop in the Southwest Hilly Area, China, but there has been a lack of efficient and accurate monitoring methods in the region. Recently, convolutional neural networks (CNNs) have obtained considerable achievements in the remote sensing community. However, it has not been widely used in mapping a rice paddy, and most studies lack the comparison of classification effectiveness and efficiency between CNNs and other classic machine learning models and their transferability. This study aims to develop various machine learning classification models with remote sensing data for comparing the local accuracy of classifiers and evaluating the transferability of pretrained classifiers. Therefore, two types of experiments were designed: local classification experiments and model transferability experiments. These experiments were conducted using cloud-free Sentinel-2 multi-temporal data in Banan District and Zhongxian County, typical hilly areas of Southwestern China. A pure pixel extraction algorithm was designed based on land-use vector data and a Google Earth Online image. Four convolutional neural network (CNN) algorithms (one-dimensional (Conv-1D), two-dimensional (Conv-2D) and three-dimensional (Conv-3D_1 and Conv-3D_2) convolutional neural networks) were developed and compared with four widely used classifiers (random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM) and multilayer perceptron (MLP)). Recall, precision, overall accuracy (OA) and F1 score were applied to evaluate classification accuracy. The results showed that Conv-2D performed best in local classification experiments with OA of 93.14% and F1 score of 0.8552 in Banan District, OA of 92.53% and F1 score of 0.8399 in Zhongxian County. CNN-based models except Conv-1D provided more desirable performance than non-CNN classifiers. Besides, among the non-CNN classifiers, XGBoost received the best result with OA of 89.73% and F1 score of 0.7742 in Banan District, SVM received the best result with OA of 88.57% and F1 score of 0.7538 in Zhongxian County. In model transferability experiments, almost all CNN classifiers had low transferability. RF and XGBoost models have achieved acceptable F1 scores for transfer (RF = 0.6673 and 0.6469, XGBoost = 0.7171 and 0.6709, respectively).

Applying Non-Random Block Cross-Validation to Improve Reliability of Model Selection and Evaluation in Hydrology: An illustration using an algorithmic model of seasonal snowpack

<p>A central question in model structural uncertainty is how complex a model should be in order to have greatest generality or transferability.  One school of thought is that models become more general by adding process subroutines.  On the other hand, model parameters and structures have been shown to change significantly when calibrated to different basins or time periods, suggesting that model complexity and model transferability may be antithetical.  An important facet to this discussion is noting that validation methods and data applied to model evaluation and selection may tend to bias answers to this question.  Here we apply non-random block cross-validation as a direct assessment of model transferability to a series of algorithmic space-time models of April 1 snow water equivalent (SWE) across 497 SNOTEL stations for 20 years.  In general, we show that low to moderate complexity models transfer most successfully to new conditions in space and time.  In other words, there is an optimum between overly complex and overly simple models.  Because structures in data resulting from temporal dynamics and spatial dependency in atmospheric and hydrological processes exist, naïvely applied cross-validation practices can lead to overfitting, overconfidence in model precision or reliability, and poor ability to infer causal mechanisms.  For example, random k-fold cross-validation methods, which are in common use for evaluating models, essentially assume independence of the data and would promote selection of more complex models.  We further demonstrate that blocks sampled with pseudoreplicated data can produce similar outcomes.  Some sampling strategies favored for hydrologic model validation may tend to promote pseudoreplication, requiring heightened attentiveness for model selection and evaluation.  While the illustrative examples are drawn from snow modeling, the concepts can be readily applied to common hydrologic modeling issues.</p>