TSA-GAIN method for greenhouse environment data imputation based on encoder and dynamic hint matrix

Missing data in greenhouse environmental monitoring may substantially compromise the accuracy of predictive models and the effectiveness of control strategies. To address long-gap missing data and the fragmentation of cross-sensor coupling features in greenhouse environmental datasets, this study improves the Generative Adversarial Imputation Network (GAIN) and proposes a Two-Stage Adaptive Generative Adversarial Imputation Network (TSA-GAIN). The main innovations are as follows: an adaptive hint-rate function is designed to optimize the missing-value localization guidance mechanism. By leveraging the dynamic balance between generator and discriminator adversarial losses, the hint rate is updated adaptively, replacing the fixed hint-rate setting in conventional GAIN and thereby constructing a dynamic hint matrix to improve the training process. In addition, the model architecture is enhanced by introducing an encoder module at the front end of the generator. By embedding the distributional characteristics of the original data and the spatiotemporal correlations among variables into the latent noise vector, the model can generate samples more realistically and more stably, effectively addressing the slow convergence and unstable imputation caused by GAIN’s reliance on random noise. Experimental results show that TSA-GAIN significantly outperforms GAIN on three imputation metrics, with MAE reduced by 51.8%, RMSE reduced by 48.7%, and the correlation coefficient increased by 61.7%. Moreover, under high missing rates, the proposed method outperforms traditional approaches such as RF, KNN, and MICE, which can better adapt to the nonlinear relationship in greenhouse climate characteristics.