"Road transport is a major contributor to world energy consumption and emissions. The validity of models developed for environmental assessment of transport projects when used out of their origins is questionable as they are only validated for the prevailing conditions at their origin. This study starts by the validation of one of the most popular transportation environmental assessment models, MOVES, for use in non-U.S. regions such as Canada through performing on-road measurements. Distinct differences between the ground-truth and MOVES predictions are revealed. MOVES underestimates fuel and CO2 rates by 17% and 35%, respectively. Nitrogen Oxides (NOx) and Particulate Matters (PM) predictions set overestimation records of up to +420%. Furthermore, MOVES output is biased for vehicle groups with specific attributes. The results of MOVES validation emphasized the need for using alternative local fuel and emission models. However, many of the existing vehicular fuel and emission modeling methodologies are criticized in aspects such as ignoring real-world training data, low diversity of test fleet, impracticality in real-world applications (such as instrument-independent eco-driving or use alongside with traffic microsimulation), and low prediction power in the non-linear multi-dimensional space of fuel consumption and emission generation. Hence, a machine learning modeling methodology relying on on-road data from a fleet of 35 vehicles is proposed. The accuracy of the proposed instrument-independent models is tried to be improved by introducing estimates of influential engine variables to the feature set through a cascaded modeling procedure. As a result, the R-squared metric reached 83%, while score improvements as high as 37% are achieved depending on the vehicle class and the machine learning technique used.Despite the considerable scores achieved by utilizing fully-connected neural networks architectures, use of techniques compatible with the serially-correlated nature of vehicular operation seems more promising in achieving higher accuracy and robustness. Moreover, generalizing the models developed for particular vehicles to more aggregate levels is a need for diversifying models’ use cases. To this end, a two-stage ensemble learning methodology based on vehicle-specific Recurrent Neural Network (RNN) models is proposed.Long Short-Term Memory (LSTM) cell architecture resulted in the best lag-specific modeling scores (compared to the other RNN cell types). Vehicle-specific ensemble models developed by combining predictions from lag-specific RNN models showed score improvement records of up to 28% compared to the best component model (4% on average). In addition, the category-specific ensembles developed on top of metamodels achieved score improvements of up to 32% compared to the best component metamodel (6% on average). Linear regression dominantly resulted in the best score improvements for NOx and PM rates at both forecast combination stages, while random forests and gradient boosting methods dominantly worked the best for fuel and CO2 rates"--