The increasing problem of environmental pollution requires a new level of innovation going beyond the scope of existing monitoring systems. In this paper, we propose EnviroSense-ML – an end-to-end architecture leveraging IoT sensors together with machine learning algorithms for environmental monitoring and predictions. Our solution consists of a combination of inexpensive electrochemical sensors, LoRaWAN-based communication channels, and novel approaches in the field of hybrid machine learning techniques, which include the spatiotemporal GCN-LSTM model and CNN-BiGRU model using 8-bit quantization. The performance evaluations performed using the real-world dataset showed that our GCN-LSTM model demonstrated the highest interpolation accuracy (R² = 0.96), due to the inclusion of additional information about altitude and land cover into graph connections of the sensors. At the same time, 8-bit quantization resulted in 66% compression of the model\\\'s size with less than 1% degradation of its accuracy. Moreover, experiments showed that ML algorithms can improve sensor measurements\\\' accuracy up to 46%. Also, our two-stage approach based on XGBoost reached near-perfect Air Quality Index prediction results (R² = 1.00, MAE = 0.35).
Dr. Dolley Srivastava (Thu,) studied this question.