Aviation plays a significant role in global warming through both CO2 and non-CO2 emissions, with persistent contrails and aviation-induced cloudiness serving as key drivers of radiative forcing. Accurate detection of these phenomena is therefore essential for evaluating aviation's environmental impact. This work introduces the Contrail Detection and Segmentation Ensemble Model (CDSEM), combining six neural networks for semantic and instance segmentation, trained on the Open-Contrails dataset and our modified polygon-based variant (PB-OpenContrails). Detection metrics show that CDSEM accurately identifies 93% of targeted contrail features, with false positive detections under 3%. Segmentation metrics were found to be significantly constrained by imperfections in human-delineated contrails used as ground truth. To address this, we propose the Boundary Soft (BS γβ) framework, which accounts for minor pixel deviations (γ and β). With this framework, CDSEM achieves an 81.25% Global Dice Score for γ = β = 1. Despite strong detection and segmentation performance on OpenContrails, our analysis revealed that CDSEM struggles with older, diffused, and nonlinear contrails, a limitation underrepresented in the metrics due to the scarcity of such cases in the dataset. To enhance its applicability in detecting a broader range of aviation-induced cloudiness, we propose the Optical Flow Correction (OFC) algorithm as an additional layer for CDSEM. This integration improves the detection of aged contrails by incorporating information from previously identified ones. Collectively, the proposed methods serve to effectively detect both linear and older diffused contrails, making them valuable tools for climate studies. Furthermore, the proposed evaluation framework and the insights from this study establish a solid foundation for enhancing the performance and interpretability of machine learning models applied in this domain.