<p>Optimal water quality is a prerequisite for sustainable aquaculture operations; however, the existing approaches for evaluating the water quality are inadequate in addressing the complex and interacting factors that affect aquatic environment over time. Current approaches offer cross-sectional data with no consideration for dynamic fluctuations and relationships between water quality and the fish environment that is essential for their well-being. Other traditional deep learning models are also challenged by long-term dependencies and interactions among those variables. Considering such challenges, we introduce a new framework that integrates IoT technologies with deep learning algorithms for real-time assessment of water quality in aquaculture. In our approach, strategically positioned IoT sensors collect an uninterrupted stream of water quality and environmental data. This data is then labelled through the Aqua-Enviro Index (AEI) which is a new index that uses water quality parameters, fish death rates, and environmental parameters to give better representation of the aquatic ecosystem. The framework proposed here is based on the Temporal FlexNet (TFN) model, a novel deep learning architecture that addresses the identified issues. 2D Convolutions are used for spatial features of the water body whereas Long Short-Term Memory (LSTM) to capture the temporal evolution of water quality with respect to the environmental factors. The proposed TFN model yields one of the outstanding performances as it has 99.38% accuracy on public datasets and 98.79% on real-time datasets while Hybrid CNN-LSTM and ANN-MLR have 93.57% and 95.76% accurate respectively. This high accuracy, combined with better precision, recall, and F-score further makes the TFN as a new state-of-the-art for water quality classification. Thus, our framework can be applied at large scale by aquaculture managers to monitor water conditions, predict problems, and prevent them to improve fish health and production in various environments.</p>