Assuring the quality of water is crucial for the growth and survival of fish in aquaculture ponds. Traditional methods of water quality monitoring can be inefficient which makes real-time monitoring and decision is a challenging one. Some deep learning techniques have shown apparent in improving water quality monitoring and assessment process, but encounter some limitations like data-overfitting, interpretability, and finds difficulties in capturing complex spatial and temporal dynamics that have hindered their effectiveness. To overcome these challenges, we propose an enhanced Dilated Spatial-temporal Convolution Neural Network (DSTCNN) for water quality monitoring in aquaculture, which uses an IoT system setup for capturing real-time data inputs from aqua ponds. The water quality data captured through the IoT sensors is labeled as per the water quality index (WQI) standards for analysis. This labeled data is effectively classified into two categories by the proposed DSTCNN model based on their suitability for fish growth or potential to cause fish mortality. By the leveraging power of dilated convolutions, the DSTCNN architecture accurately handles the intricacies of both spatial and temporal data, enabling it to capture essential features and patterns across multiple snapshots. This capability empowers the model to truly comprehend the complex relationships inherent in spatiotemporal data. Furthermore, to address the concerns like overfitting due to complexity of data and enhance generalization, the proposed model employs a hybrid activation function that synergistically combines ReLU and sigmoid during the activation process. The proposed DSTCNN model has been implemented on real-time and public datasets and obtained 99.28% and 99.02% accuracy respectively, whereas the state-of-the-art PCR-GB model obtains 96.97% and 97.11% accuracy on real-time and public datasets respectively.

保证水质对于水产养殖池塘鱼类的生长和生存至关重要。传统的水质监测方法效率低下，这使得实时监测和决策成为一项挑战。一些深度学习技术在改善水质监测和评估过程方面显示出明显的效果，但遇到了一些限制，例如数据过度拟合、可解释性，并且在捕获复杂的时空动态方面存在困难，从而阻碍了其有效性。为了克服这些挑战，我们提出了一种用于水产养殖水质监测的增强型扩张时空卷积神经网络（DSTCNN），它使用物联网系统设置来捕获来自水产池塘的实时数据输入。通过物联网传感器捕获的水质数据根据水质指数（WQI）标准进行标记以进行分析。所提出的 DSTCNN 模型根据这些标记数据对鱼类生长的适合性或导致鱼类死亡的可能性，将其有效地分为两类。通过利用扩张卷积的力量，DSTCNN 架构可以准确处理空间和时间数据的复杂性，使其能够跨多个快照捕获基本特征和模式。这种能力使模型能够真正理解时空数据固有的复杂关系。此外，为了解决由于数据复杂性而导致的过度拟合等问题并增强泛化能力，该模型采用了混合激活函数，在激活过程中协同结合 ReLU 和 sigmoid。所提出的 DSTCNN 模型已在实时和公共数据集上实现，分别获得了 99.28% 和 99.02% 的准确率，而最先进的 PCR-GB 模型在实时和公共数据集上获得了 96.97% 和 97.11% 的准确率分别为数据集。