Methane (CH₄) is the second most important anthropogenic greenhouse gas (GHG) after carbon dioxide (CO₂), and paddy fields are among the largest sources of CH₄ emissions. Owing to the scarcity of observational data, the characteristics and influencing factors of CH₄ fluxes in paddy fields at different timescales need to be further investigated. Observations of CH₄ fluxes via eddy covariance (EC) data were performed over four seasons in two paddy fields in Nanchang, Jiangxi Province, China. The pattern, magnitude and biophysical controls of CH₄ emissions were explored by wavelet analysis and stepwise multiplicative modelling. The results revealed a distinct, single-peak diurnal pattern in CH₄ fluxes during the vegetative stage in all four rice growing seasons and the reproductive stage of early rice. Large seasonal variations in daily CH₄ emissions were observed in the two double-cropping paddy fields, and the soil temperature at a depth of 5 cm (T_s5) explained most of the seasonality of the CH₄ fluxes. At the inter-seasonal scale, under local farms’ traditional field management method, reducing the amount and frequency of irrigation during the vegetative stage could decrease CH₄ emissions in southern China. This study improves the understanding of CH₄ emissions and helps in developing GHG management strategies for paddy fields.

Accurate and efficient extraction of rice planting structures, coupled with comprehensive analysis of their spatiotemporal dynamics and driving factors, is crucial for rice yield estimation and optimized water resource management in the Poyang Lake region. However, traditional approaches face significant limitations: single machine learning models often yield insufficient classification accuracy, while existing fusion models typically involve complex processing workflows and exhibit low computational efficiency. To address these challenges, this study developed an efficient and simplified fusion model based on a scoring strategy to determine rice planting structures from 2018 to 2023, followed by an in-depth analysis of their spatiotemporal patterns and underlying drivers. The evaluation results demonstrated that four individual classification models of K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting Decision Tree (GBDT) achieved Overall Accuracy of 85.29%–90.07%, Kappa coefficients of 0.786–0.855, User Accuracy of 80.51%–93.02%, and Mapping Accuracy of 80.87%–92.63%. The proposed scoring-based fusion model significantly enhanced these metrics, improving Overall Accuracy by 3.36%–9.16%, Kappa coefficient by 5.15%–14.38%, User Accuracy by 0.37%–11.13%, and Mapping Accuracy by 0.48%–10.71%. Spatiotemporal analysis revealed distinct trends in rice cultivation patterns: single-cropping rice and regenerated rice showed consistent expansion, both in planting area and proportion, with a spatial tendency towards flat regions. Conversely, double-cropping rice exhibited a gradual decline, with its cultivation areas contracting towards the central lake region. These shifts were primarily driven by socioeconomic factors, particularly rural labor migration and rising fertilizer prices, which have incentivized farmers to adopt less labor-intensive and lower-input cultivation systems, such as single-cropping and regenerated rice. The findings offer a novel methodological framework for precise extraction of crop planting structures, and a scientific foundation for local governments to develop targeted water resource management strategies.