NPP estimation by fusing geodetector and deep spatio-temporal networks

作者：Xiaohui He, Chi Yuan, Panle Li, Xijie Cheng, Mengjia Qiao, Xiaoyu He, Nan Yang, Guangsheng Zhou, Jiandong Shang · 发表于：International Journal of Digital Earth · 年份：2025 · DOI：10.1080/17538947.2025.2515265 · 被引用次数：4 · 研究领域：Computational Physics and Python Applications、Reservoir Engineering and Simulation Methods、Geophysics and Gravity Measurements

In recent years, deep learning has demonstrated significant potential in net primary productivity (NPP) estimation but the existing methods fall short in fully exploiting the spatio-temporal dependencies inherent in remote sensing data for modeling NPP. To address this limitation, we propose a novel approach named integrate geographic with deep spatio-temporal networks (IGDSNet). Specially, the IGDSNet uses the geodetector to explore geographic mechanism of NPP and then introduces the spatio-temporal long- and short-term memory networks (ST-LSTM) to obtain the deep spatio-temporal feature of NPP. Finally, we develop a novel decoder by combining self-attention unit (SAU) and 3D convolution (Conv3D) to produce the spatio-temporal distribution of NPP. By incorporating ST-LSTM and SAU for capturing long-term dependencies, the proposed IGDSNet is able to model both localized variations and global-scale patterns of NPP distribution. The experimental results demonstrate that the IGDSNet outperforms other deep spatio-temporal modeling methods, with improvement R2 in the range of 0.4%–10.57%. In addition, the IGDSNet is used to estimate the NPP distribution across Tibetan Plateau and create a monthly NPP dataset that the Moran's I index generally remains above 0.75. This study provides vital information for monitoring, modeling and managing the carbon cycle of Tibetan Platea.