To address the high costs of environmental pollutant detection and the reliance on animal experiments, quantitative structure-activity relationship (QSAR) models have been widely used in predicting the molecular activity and ecotoxicity of typical environmental pollutants (such as persistent organic pollutants and endocrine-disrupting chemicals). However, traditional QSAR methods often suffer from the "activity cliff" and the scarcity of small-sample data when identifying characteristic pollutants in complex environmental samples, resulting in insufficient generalization capability of the models. In view of these issues, this paper systematically reviewed the mainstream training paradigms of deep learning in environmental molecular modeling, including strategies such as transfer learning, pre-training and fine-tuning, and self-supervised learning, and it discussed their theoretical foundations and technical implementations in feature extraction, model robustness, and cross-task generalization. Literature analysis and method comparison show that compared with traditional machine learning methods, these training paradigms demonstrate stronger generalization capability, cross-task adaptability, and identification advantages for structural features of unknown compounds under small-sample conditions across different types of self-constructed environmental datasets. This paper further proposes a molecular modeling workflow suitable for environmental pollutant modeling and risk assessment and summarizes the key points of the model evaluation system, providing a systematic pathway for constructing highly generalizable and interpretable environmental deep learning models. This research can provide structured technical guidelines for model development in the environmental field and a reference framework for model comparison, promote the systematic evaluation of different deep learning training strategies, and offer support for the industry to establish more unified and reproducible intelligent modeling standards, thereby promoting the development of environmental pollutant screening and risk assessment toward data-driven and intelligent directions.