Soil moisture (SM) is a crucial land surface variable that links cyclic processes between the land surface and the atmosphere. Accurate SM prediction holds great significance for agricultural production, drought assessment, and global climate forecasting. However, the complex geographical and environmental factors lead to intricate variations and an irregular distribution of SM. Most of the existing multivariate prediction methods predominantly consider localized spatial data, rendering it challenging to capture the heterogeneous dependence effect of SM variations from remote locations. To address this challenge, this study proposes the GCCL model, which integrates two components: GConvLSTM (graph convolutional LSTM) and ConvLSTM (convolutional LSTM). This combined model is designed to forecast SM for the next 7 days. GCCL constructs the connectivity matrix with Pearson correlation coefficient, focusing not only on the surrounding spatial information but also on the broader spatial information and temporal correlation, which results in the ability to capture the heterogeneous dependencies of long-range SM variations. GCCL is evaluated against ConvLSTM, TGC-LSTM, CNN-LSTM, Random Forest, and SARIMAX models for SM prediction using Soil Moisture Active Passive (SMAP) satellite L4 products. Evaluate the model’s performance at different stages using the prediction results for the next 1, 3, 5, and 7 days. The results demonstrate that GCCL outperforms the baseline models, with RMSE of 0.018 m3/m3, 0.031 m3/m3, 0.035 m3/m3, and 0.038 m3/m3 for multistep predictions, respectively. Compared to ConvLSTM, GCCL reduces Root Mean Square Error (RMSE) by 14.3 %, 8.8 %, 7.9 %, and 7.5 %, respectively. In spatiotemporal scenarios, GCCL excels under complex topographical, rainy, and drought conditions. For multistep predictions, GCCL outperforms ConvLSTM in 98 %, 94 %, 86 %, and 83 % of the regions. These findings indicate that the GCCL model demonstrates a superior spatiotemporal performance in multistep SM prediction, with the potential to enhance the accuracy and stability of SM spatiotemporal prediction.

中文翻译：

---

结合图神经网络和卷积 LSTM 网络进行多步土壤水分时空预测


土壤湿度 （SM） 是一个重要的地表变量，它连接着地表和大气之间的循环过程。准确的 SM 预测对农业生产、干旱评估和全球气候预测具有重要意义。然而，复杂的地理和环境因素导致了 SM 的复杂变化和不规则分布。大多数现有的多元预测方法主要考虑局部空间数据，这使得从远程位置捕获 SM 变化的异质依赖效应具有挑战性。为了应对这一挑战，本研究提出了 GCCL 模型，该模型集成了两个组件：GConvLSTM （图卷积 LSTM） 和 ConvLSTM （卷积 LSTM）。此组合模型旨在预测未来 7 天的 SM。GCCL 构建了具有 Pearson 相关系数的连通性矩阵，不仅关注周围的空间信息，还关注更广泛的空间信息和时间相关性，从而能够捕获远程 SM 变化的异质依赖关系。GCCL 根据 ConvLSTM、TGC-LSTM、CNN-LSTM、随机森林和 SARIMAX 模型进行评估，以使用土壤水分主动被动 （SMAP） 卫星 L4 产品进行 SM 预测。使用未来 1 天、 3 天、 5 天和 7 天的预测结果评估模型在不同阶段的性能。结果表明，GCCL 优于基线模型，多步预测的 RMSE 分别为 0.018 m3/m3、0.031 m3/m3、0.035 m3/m3 和 0.038 m3/m3。与 ConvLSTM 相比，GCCL 的均方根误差 （RMSE） 分别降低了 14.3 % 、 8.8 % 、 7.9 % 和 7.5 %。 在时空情景中，GCCL 在复杂的地形、多雨和干旱条件下表现出色。对于多步预测，GCCL 在 98 %、94 %、86 % 和 83% 的区域优于 ConvLSTM。这些发现表明，GCCL 模型在多步 SM 预测中表现出卓越的时空性能，有可能提高 SM 时空预测的准确性和稳定性。