An Encoder-Decoder Model with Interpretable Spatio-Temporal Component
for Soil Temperature Prediction
Abstract
Soil temperature (ST) is a crucial land-surface parameter and accurate,
interpretable ST predictions are essential for Earth system science
applications. While deep learning methods have shown excellent
performance in ST prediction, they are often referred to as “black box
optimizers”, making it difficult to extract physical knowledge and gain
interpretability. To address this issue, we developed the
Encoder-Decoder Model with Interpretable Spatial-Temporal Component
(ISDNM) to improve predictive accuracy and provide spatial-temporal
interpretation of ST. The ISDNM model combines a CNN-encoder-decoder and
LSTM-encoder-decoder to enhance the representation of spatial-temporal
features and applies linear regression and UMAP to provide interpretable
spatial-temporal insights into ST. The ISDNM outperforms traditional
deep learning models such as Convolutional Neural Network, Long
Short-term Memory, and Convolutional LSTM, making it a valuable tool to
improve our understanding of ST’s spatiotemporal characteristics.