Floods are one of the most destructive natural disasters. They cause serious damage to property, infrastructure, and the economy every year across the world. To address problems related to accuracy and interpretability, this study proposed a hybrid attention-based framework for flood analysis. The framework integrated a customized Attention Vision Transformer (AttViT) with a MultiAttNet classifier. In addition, SHAP-based feature selection was incorporated to improve interpretability. The proposed AttViT used an enhanced attention block. This helped the model capture global contextual information more effectively. MultiAttNet combined three complementary attention-based branches. These were an attention-based MLP (AttMLP), an attention-based CNN (AttCNN), and an attention-based BiLSTM (AttBi-LSTM). Together, they captured both spatial and sequential patterns from the extracted features. In addition, SHapley Additive exPlanations (SHAP) was applied. It was used to identify and rank the most influential features. This reduced feature redundancy and improved model transparency. Experiments on the AlleyFloodNet dataset showed an accuracy of 98.4%. The proposed method outperformed both convolutional and transformer-based baseline model. The integration of attention-guided feature extraction, multi-branch learning and explainable feature selection established a robust and transparent approach for complex environmental image analysis.