Streamlined multilayer perceptron for contaminated time series reconstruction: A case study in coastal zones of southern China

Abstract

Time series reconstruction is pivotal for enabling continuous, long-term monitoring of environmental changes, particularly in rapidly evolving coastal ecosystems. Despite the array of developed reconstruction methods, they often fail to be effectively applied in coastal zones. In coastal zones, the dynamic environment and frequent cloud cover undermine the effectiveness of existing methods, making it challenging to accurately capture time series variations. Additionally, the need for long-term, large-scale monitoring demands methods that are both efficient and adaptable. To address these challenges, a streamlined multilayer perceptron (SMLP) method is proposed to reconstruct contaminated and long-term time series in coastal zones, consisting of three steps. Firstly, to mitigate the impact of anomalies, we constructed a frequency principle theory (FPT)-based filtering module. Subsequently, to capture variations within the time series, we proposed a frequency domain representation (FDR)-based decomposition module. Finally, considering gaps in time series, we applied an implicit neural representation (INR)-based reconstruction module. SMLP was evaluated using dense Landsat time series data from 1999 to 2019 in southern China, where the data face challenges from noise, gaps, and variations. Qualitative results show that the $\overline{{\text{RMSE}}_c}$ of SMLP is 0.028, lower than other methods ranging from 0.02 to 0.05. Furthermore, quantitative analysis demonstrates that SMLP is more effective than existing approaches in mitigating the impact of anomalies and accurately capturing variations in time series. Additionally, the rapid operational speed and high transferability of SMLP makes it well-suited for long-term and large-scale applications, providing valuable support for coastal zone research.