A Transformer-Based Model for Time Series Prediction of Remote Sensing Data

In the field of ecology, remote sensing technology is widely used to acquire time-series data of surface changes, and time-series prediction algorithms are employed to forecast the future ecological conditions of the surface. Traditional time-series prediction models fail to effectively capture the dependencies of relative positions in the data and lack perception of periodic, seasonal, and trend information within the data. To address these limitations, we propose a novel network architecture called the RSformer model for time-series prediction of remote sensing data. Firstly, the model effectively extracts temporal relationships and positional information of time-series through a feature extraction module, enhancing the understanding of dependencies between time-series. Secondly, by introducing multiple sequence decomposition mechanisms, the model can more accurately extract periodic, seasonal, and trend features within the time-series. Finally, we add a denormalization module to prevent excessive normalization of data during training, reducing the loss of original data information. We collect remote sensing data of different spectral bands in open-pit mining areas and validate the model's time-series prediction performance on this data. Results showthat compared to the baseline Transformer, the RSformer increases the MSE indicator by 26% and the MAE by 28%, significantly improving the prediction accuracy of remote sensing data forecasting tasks.