<p>Depression is a significant concern for mental health that requires accurate and efficient detection for timely intervention. Millions of users share their thoughts and emotions on social media, providing a valuable data source for early depression detection. Although Bengali is the sixth most widely spoken language worldwide, only a few studies have focused on this language. In this study, we present a Bangla depression detection dataset that is categorized into”depressive” and”non-depressive,”. This work focuses on the early detection of depression, particularly in extroverted social media users. Our proposed hybrid cascaded model with attention approach (CNN-BiLSTM), as well as advanced Deep learning models (LSTM, Bi-LSTM, GRU, BiGRU) and pre-trained transformer models (BERT, BanglaBERT, SahajBERT, BanglaBERT-Base). According to the experiment results, the proposed model, CNN-BiLSTM, outperformed several deep learning models and achieved an accuracy of 91.6%. Among the pre-trained transformer-based models, BanglaBERT demonstrated the highest accuracy of 89%. We also categorized the severity of depression in social media posts into four levels—non-depressed, mild, moderate, and severe. BanglaBERT outperformed other models, achieving an accuracy of 79%. Furthermore, we integrated the Explainable Artificial Intelligence (XAI) technique LIME (Local Interpretable Model Agnostic Explanations) to analyze model predictions and model decisions by providing insights into how the models identify depressive content. This explainability adds transparency and trustworthiness to the system, particularly in sensitive applications like mental health. Our findings underscore the effectiveness of hybrid architectures, augmented methods, and XAI-driven interpretability in advancing depressive content detection for resource-constrained languages like Bangla.</p>