Improving soil organic carbon estimation in paddy fields using data augmentation algorithm and deep neural network model based on optimal image date

•We proposed a novel SOC estimation method using augmented dataset based on optimal image date.•The heading to flowering stage of rice was determined as optimal image date for SOC estimation.•The dataset augmented with the Generative Adversarial Network algorithm improved the accuracy of the SOC estimation model.•The proposed method is potential for estimating regional-scale SOC in paddy fields. Soil organic carbon (SOC) is the largest carbon reservoir in terrestrial ecosystems. Rapid monitoring of SOC is essential for sustainable agricultural management and environmental protection. However, affected by the cost and time of soil sampling, limited datasets remain a main challenge for satellite-driven SOC estimation. Moreover, current studies on satellite-driven SOC estimation in vegetation covered areas rarely consider the optimal image date, which may hinder the accuracy of SOC estimation. Therefore, this study proposes a new method to improve SOC estimation. Firstly, on the basis of the sample dataset, generative adversarial networks algorithm was used to augment crop spectral variables and annotated data. Secondly, the optimal image date was determined by the partial least squares regression fit degree based on the characteristic crop spectral variables screened by the extreme gradient boosting algorithm. Then, the support vector regression, random forest and deep neural networks algorithms were used to construct SOC estimation models and the accurate model was determined by comparing the performance of the models. Finally, the accurate model was used to obtain the SOC at regional-scale. The results of our study are as follows:(1) 120 sample data were generated based on 120 original sample data by using the generative adversarial networks algorithm. (2) The optimal image date for SOC estimation was determined as heading and flowering stage of rice, with seven characteristic spectral variables. (3) The deep neural networks model was determined as the accurate model with a determination coefficient of 0.76 and a normalized root mean square error (NRMSE) of 7.75%. (4) The deep neural networks model based on augmented data provided superior estimation accuracy at regional-scale than that based on original data, significantly decreasing the NRMSE by 3.17%. This study offers a promising method to address the limited datasets and highlights the importance of optimal image date in SOC monitoring within paddy fields.