Integrating Zero-Shot Learning with Convolutional neural network in predicting dissolved oxygen & Salinity in river water

Abstract

In this paper, we propose the machine learning model which apply Zero-Shot Learning (ZSL) with CNNs to predict river water DO and salt without tagged data or additional environmental factors. CNNs determine important water quality and meteorological factors. ZSL adaptability forecasts new situations. The proposed model can project accurately without direct training data by modelling these features in a semantic space with domain expertise and variable linkages. CNN analyses raw input data to find complicated patterns and connections to understand water quality changes. In the proposed model, temperature, pH, and flow rate affect DO and salinity. This model forecasts unexpected events using semantic linkages. This proposed model improves real-time predictions and environmental adaptation. Use semantic linkages to estimate dissolved oxygen (DO) and salinity effects in severe weather or locations with poor monitoring systems with the ZSL-CNN model. This aids fast, accurate forecasts. Adaptability makes the model powerful for water quality management, where quick and precise decision-making is essential to handle environmental challenges and preserve aquatic ecosystems. Zero-shot learning (ZSL) and convolutional neural networks allow the model to adapt to new input and forecast without retraining. This proposed model enables environmental monitoring systems adapt to new data and conditions. Proposed CNN model improve performance from RMSE 0.5 to RMSE 0.4 and R² 0.7, while GRU models improve performance to RMSE = 0.35 and R² = 0.8. The CNN-GRU model can lower RMSE to 0.3 and boost R² to 0.85. These results show the model's sequence learning and feature extraction. This proposed model leverages CNNs' feature extraction and Zero-Shot Learning's flexibility. Water resource management and environmental protection improve.