Enhancing River Water Level Prediction Using CNN with Attention and Squeeze-and-Excitation Blocks

River Water Level Prediction (RWLP) serves as a critical tool for running river water management along with modeling hydrological processes and conserving sustainable water resources. Manual readings and ultrasonic sensors, which are traditionally used for measurement, face operational problems while also being affected by environmental conditions. A deep learning framework handles these issues by using image data to create classifications, which sort river water levels into high, medium and low categories. A multi-stage process handles a workflow that starts with organizing labeled images and subsequently acquires data for use in the pipeline. The analysis identified the distribution across classes as well as the different characteristics in the provided images through Exploratory Data Analysis (EDA). Previous data processing techniques normalized images while applying resizing methods and enhancing training stability through augmentation procedures. The research establishes seven convolutional neural network (CNN) models including a baseline CNN as well as VGG16, MobileNet, InceptionV3, Xception, along with two custom models integrated with attention mechanisms and Squeeze-and-Excitation (SE) blocks. The evaluation involved standard performance metrics like precision and accuracy together with recall and F1-score while preserving constant trainer and validator parameters. The model interpretation was improved by the use of Gradient-weighted Class Activation Mapping (Grad-CAM) for showing the key areas that influenced each decision. MobileNet outshined other models in the experiment by achieving 0.9918 accuracy alongside 0.9911 precision and 0.9918 recall and 0.9917 F1-score with only seven misinterpretations in the validation set. MobileNet succeeds in delivering accurate classifications while using limited resources. The SE block-based CNN model outperformed Xception and InceptionV3 in addition to surpassing both the attention-based CNN model and its baseline version. The visual regions focused by MobileNet's Grad-CAM outputs were consistently identified because the model paid attention to important components such as gauge readings and water lines, which strengthened both its reliability and interpretability. The research outcomes show that MobileNet among lightweight pre-trained CNNs excels for river water level detection while providing efficient, accurate and transparent operations for real-time hydrological observation.