Quantifying and tracking the soil organic carbon (SOC) content is a key step toward long-term terrestrial ecosystem monitoring. Over the past decade, numerous models have been proposed and have achieved promising results for predicting SOC content. However, many of these studies are confined to specific temporal or spatial contexts, neglecting model transferability. Temporal transferability refers to a model’s ability to be applied across different periods, while spatial transferability relates to its applicability across diverse geographic locations for prediction. Therefore, developing a new methodology to establish a prediction model with high spatiotemporal transferability for SOC content is critically important. In this study, two large intercontinental study areas were selected, and measured topsoil (0–20 cm) sample data, 27,059 cloudless Landsat 5/8 images, digital elevation models, and climate data were acquired for 3 periods. Based on these data, monthly average climate data, monthly average data reflecting soil properties, and topography data were calculated as original input (OI) variables. We established an innovative multivariate deep learning model with high spatiotemporal transferability, combining the advantages of attention mechanism, graph neural network, and long short-term memory network model (A-GNN-LSTM). Additionally, the spatiotemporal transferability of A-GNN-LSTM and commonly used prediction models were compared. Finally, the abilities of the OI variables and the OI variables processed by feature engineering (FEI) for different SOC prediction models were explored. The results show that 1) the A-GNN-LSTM that used OI as the input variable was the optimal prediction model (RMSE = 4.86 g kg−1, R2 = 0.81, RPIQ = 2.46, and MAE = 3.78 g kg−1) with the highest spatiotemporal transferability. 2) Compared to the temporal transferability of the GNN, the A-GNN-LSTM demonstrates superior temporal transferability (ΔR2T = −0.10 vs. −0.07). Furthermore, compared to the spatial transferability of LSTM, the A-GNN-LSTM shows enhanced spatial transferability (ΔR2S = −0.16 vs. −0.09). These findings strongly suggest that the fusion of geospatial context and temporally dependent information, extracted through the integration of GNN and LSTM models, effectively enhances the spatiotemporal transferability of the models. 3) By introducing the attention mechanism, the weights of different input variables could be calculated, increasing the physical interpretability of the deep learning model. The largest weight was assigned to climate data (39.55 %), and the smallest weight was assigned to vegetation (19.96 %). 4) Among the commonly used prediction models, the deep learning model had higher prediction accuracy (RMSE = 6.64 g kg−1, R2 = 0.64, RPIQ = 1.78, and MAE = 4.78 g kg−1) and spatial transferability (ΔRMSES = 1.43 g kg−1, ΔR2S = −0.13, ΔRPIQS = −0.50, and ΔMAES = 1.09 g kg−1), and the linear model had the higher temporal transferability (ΔRMSET = 1.46 g kg−1, ΔR2T = −0.14, ΔRPIQT = −0.45, and ΔMAET = 1.29 g kg−1). 5) The deep learning models necessitated the OI, whereas the linear and traditional machine learning models necessitated FEI to achieve higher prediction accuracy. This study presents an important step forward in integrating multiple deep learning models to build a highly spatiotemporal transferability SOC prediction model.

中文翻译：

---

一种在多年代际和洲际尺度上可时空转移的 SOC 含量预测模型的新方法


量化和跟踪土壤有机碳 （SOC） 含量是实现长期陆地生态系统监测的关键步骤。在过去的十年中，已经提出了许多模型，并在预测 SOC 含量方面取得了可喜的结果。然而，其中许多研究仅限于特定的时间或空间背景，忽视了模型的可转移性。时间可转移性是指模型在不同时期应用的能力，而空间可转移性与其在不同地理位置的预测适用性有关。因此，开发一种新的方法来建立 SOC 含量具有高时空可转移性的预测模型至关重要。在本研究中，选择了两个大型洲际研究区，并采集了 3 个时期的表土 （0–20 cm） 样本数据、27,059 张万里无云的 Landsat 5/8 图像、数字高程模型和气候数据。基于这些数据，将月平均气候数据、反映土壤特性的月平均数据和地形数据计算为原始输入 （OI） 变量。我们结合了注意力机制、图神经网络和长短期记忆网络模型 （A-GNN-LSTM） 的优点，建立了一种创新的具有高时空迁移性的多元深度学习模型。此外，还比较了 A-GNN-LSTM 和常用预测模型的时空可传递性。最后，探讨了 OI 变量和特征工程 （FEI） 处理的 OI 变量对不同 SOC 预测模型的能力。结果表明，1） 使用 OI 作为输入变量的 A-GNN-LSTM 是最优预测模型 （RMSE = 4.86 g kg-1，R2 = 0.81，RPIQ = 2.46，MAE = 3。78 g kg−1），具有最高的时空转移性。2） 与 GNN 的时间可转移性相比，A-GNN-LSTM 表现出卓越的时间可转移性 （ΔR2T = -0.10 vs. -0.07）。此外，与 LSTM 的空间转移性相比，A-GNN-LSTM 显示出增强的空间转移性（ΔR2S = -0.16 vs. -0.09）。这些发现强烈表明，通过 GNN 和 LSTM 模型的整合提取的地理空间背景和时间依赖信息的融合，有效地增强了模型的时空可转移性。3） 通过引入注意力机制，可以计算不同输入变量的权重，从而提高深度学习模型的物理可解释性。气候数据权重最大 （39.55 %），植被权重最小 （19.96 %）。4） 在常用的预测模型中，深度学习模型具有较高的预测精度 （RMSE = 6.64 g kg-1， R2 = 0.64， RPIQ = 1.78， MAE = 4.78 g kg-1） 和空间可传递性 （ΔRMSES = 1.43 g kg-1， ΔR2S = -0.13， ΔRPIQS = -0.50， ΔMAES = 1.09 g kg-1），线性模型具有较高的时间可传递性 （ΔRMSET = 1.46 g kg-1， ΔR2T = −0.14，ΔRPIQT = −0.45，ΔMAET = 1.29 g kg−1）。5） 深度学习模型需要 OI，而线性和传统机器学习模型需要 FEI 来实现更高的预测精度。本研究在集成多个深度学习模型以构建高度时空可迁移性 SOC 预测模型方面迈出了重要一步。