Abstract: Dissolved oxygen (DO) in water is one of the most important ecological factors for the healthy growth of aquaculture organisms. The too high concentration can easily cause the fish bubble disease. The low concentration can slow down the growth of fish, where the prolonged floating head can cause the fish to float. There are temporal and spatial variations in the DO of aquaculture water under various environmental factors in large-scale aquaculture. The existing predictions cannot accurately describe the spatial distribution of DO, due to the time series. In this study, an improved prediction model of spatiotemporal DO was presented for the intelligent control of water quality in ponds. The high aquaculture efficiency with fewer risks was then realized using an improved gated recursive unit using a dual attention mechanism (IDA-GRU) and an improved inverse distance weighted interpolation algorithm (IIDW). Firstly, the feature and temporal attention mechanism were introduced to improve the prediction accuracy of DO using the gated recurrent unit (GRU), according to the correlation between parameters and the dependence relationship of temporal information. The key environmental features were also strengthened using the feature attention mechanism. The contribution rate of each environmental feature was then calculated to continuously modify the weight of each environmental feature in real time. The time attention mechanism autonomously extracted the key historical moments, and then enhanced the expression of the key moment, thus improving the stability of DO time prediction. Secondly, an improved inverse distance-weighted interpolation (IDW) was applied to achieve spatial prediction using the prediction of DO time series. The distance weight correction coefficient was also introduced using the traditional inverse distance-weighted interpolation. The interpolation weights were adjusted in real time, according to the DO content at each monitoring point. An improved accuracy was achieved in the three-dimensional spatial interpolation of DO. The local environment was considered around the monitoring points. As such, the interpolation weights were changed concurrently rather than that in the traditional algorithm, once the interpolation distance was determined. Finally, the improved model was validated at the digital eco-farming site of Shanghai City electric development Co. The experimental results showed that the mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) of the evaluation indexes were 0.0687 mg²/L², 0.2621 mg/L, and 0.2051 mg/L, respectively, in the IDA-GRU model for the DO time series prediction. Additionally, the MSE, RMSE, and MAE of the IIDW were 0.2088 mg²/L², 0.4570 mg/L, and 0.3835 mg/L, respectively, for the spatial prediction of DO. A better performance was achieved, compared with the previous models. The present model had improved the accuracy of spatiotemporal DO prediction. The finding can also provide a strong reference for disaster prevention and intelligent regulation of water quality in large-scale aquaculture.