Tiny ML-based reconfigurable IoT platform design for brackish water aquaculture monitoring.

Abstract

An important sector of India's fishing industry is prawn farming. Gross prawn exports came to 5,90,275 MT (metric tonnes) and were worth $4,426.19 million. White-leg prawn exports decreased from 5,12,204 MT to 4,92,271 MT in 2020–21. To show the problem with the traditional approach of monitoring brackish water prawn aquaculture, a study was conducted. After further investigation, it became clear that the farmers were up all night trying to maintain the perfect water quality needed to produce healthy shrimp. This is because shrimp ponds have a tolerance level for a variety of environmental factors, including temperature, pH (potential of hydrogen), and humidity. As a result, the farmers must continuously check on the pond's status throughout the night. An Intelligence forecasting approach would address the complexity of farmer's crop monitoring issues. A hybrid intelligence mechanism for forecasting efficiently and handling a large amount of streaming data is achieved through Auto regressive long short-term memory integrated moving averages (ARLSTMIMA ). The intelligent algorithm is embedded into Tiny ML an IoT device developed for getting real-time data. As a result of this procedure, the pond's water quality and environmental conditions are checked to guarantee a healthy prawn crop. To get real-time environmental data and weather behavior of brackish water shrimp aquaculture as well as Real-time data collection over some time from various shrimp farming locations, the goal of this work is to construct a small ML-based resilient and wireless sensor network. For additional ML training, the appropriate sensor data are gathered and input into Google Sheets. Tiny ML thereby forecasts the sensor data before it is displayed. With this method, aquaculture producers can save time and money by receiving information at the right times from pre-established designs and achieving an efficiency of 95.16 in the prediction in comparison with the existing forecasting mechanisms such as Random Forest regressor, ARIMA, and LSTM by outperforming with the highest accuracy in forecasting values. In achieving higher accuracy, an optimal architectural design is achieved by inducting the ARLSTMIMA which has a decreased time complexity in comparison with other architectures in forecasting with time complexity of the order O(n2) with the least time for forecasting is achieved.