A Multi-Scale ConvLSTM with Seasonal Encoding for Monthly Land Surface Temperature Prediction

Abstract

: Land Surface Temperature (LST), as a key indicator for characterizing surface energy exchange and urban thermal environments, plays a crucial role in climate change research and sustainable urban development. To address the challenges of strong spatial heterogeneity, pronounced temporal periodicity, and difficulty in modeling long-term dependencies in monthly LST sequences, this study proposes an MSG-ConvLSTM model that integrates multi-scale spatial feature extraction with a periodic awareness mechanism. The model enhances spatial representation through parallel multi-scale convolutions and incorporates a sinusoidal positional encoding–based periodic gating mechanism to explicitly capture seasonal variations, thereby improving its ability to model long-term dependencies. Experimental results based on the Pearl River Delta dataset from 2003 to 2023 demonstrate that the proposed model outperforms several mainstream methods in multi-step forecasting tasks. Compared with ConvLSTM, the proposed model reduces MSE by approximately 10%–15%; compared with CNN-BiLSTM, by 12%–18%; compared with Swin-Transformer, by 8%–12%; and compared with PredRNN, by 9%–14%. Similar improvements are also observed in terms of MAE. Notably, in medium-and long-term forecasting, the model exhibits slower error accumulation and stronger stability. Ablation studies further verify the effectiveness of the multi-scale feature extraction and periodic modeling mechanisms in improving prediction accuracy and mitigating error propagation. This study provides an effective and robust approach for monthly LST forecasting using only historical LST data.