Multistep Prediction of Land Cover From Dense Time Series Remote Sensing Images With Temporal Convolutional Networks

Abstract

Time series prediction (TSP) of land use/land cover (LULC) is an important scientific issue, but forecasting LULC changes at lead times of multiple time steps at fine time scales remains problematic. Especially in the context of current rapid economic and social development, the traditional one-step prediction models with a five-year or ten-year cycle cannot meet the application needs of land management departments. Temporal convolutional networks (TCNs) outperform other traditional TSP approaches. Therefore, we have proposed a pixel-level multistep TSP (pMTSP) approach that employs TCNs to carry out multistep prediction of land cover from dense time series remote sensing images, making up for the shortcomings of low accuracy, coarse time granularity, and labor-consuming of the current LULC prediction approaches. The results of comparative experiments with seasonal-trend decomposition procedure based on LOcally wEighted regreSsion Smoother and autoregression (STL-AR), seasonal autoregressive integrated moving average, and dynamic harmonics regression using single enhanced vegetation index time series, as well as the comparative experiment with the cellular automata-Markov model using real moderate resolution imaging spectroradiometer image time series, showed that the pMTSP can accurately extrapolate the change trend of the time series in fine-scale and obtain highly consistent prediction results with actual data, performing better than the other four contrasting algorithms in 23-step LULC prediction. The pMTSP can be used for multistep, fine-time-scale, and long time-series land cover prediction, which is of great guiding significance for the sustainable development and utilization of land resources.

Keywords

• Dense time series
• land use/land cover (LULC)
• multistep prediction
• pixel-level
• temporal convolutional networks (TCNs)