Detection and attribution analysis of surface temperature anomalies based on MODIS thermal infrared data (<italic>invited</italic>)

ObjectiveUrban thermal environment issues have become increasingly prominent with accelerated urbanization. Traditional site-based temperature monitoring methods suffer from limitations such as uneven spatiotemporal coverage and poor continuity. To address these challenges, this study aims to detect and attribute land surface temperature anomalies on a national scale using Moderate-resolution Imaging Spectroradiometer thermal infrared brightness temperature data, supported by a deep learning framework, thereby providing technical support for precise monitoring and scientific regulation of urban thermal environments.MethodsThis research constructs monthly land surface temperature data across China based on MODIS thermal infrared brightness temperature data from 2002 to 2022 (Tab.1). A Long Short-Term Memory encoder-decoder framework is proposed to perform sequence prediction and reconstruction using a 12-month sliding window (Fig.1). A semi-supervised anomaly detection mechanism is developed by combining reconstruction error and prediction residuals to automatically identify temperature anomalies (Fig.2). Additionally, multi-source data including population density, normalized difference vegetation index, and evaporation are integrated to conduct multivariate regression analysis within localized time windows, quantifying the contributions of various factors to temperature anomalies.Results and DiscussionsThe LSTM prediction model demonstrates optimal performance under a 12-month forecasting step, with a mean absolute error of 1.1503 ℃, root mean square error of 1.6638 ℃, and a coefficient of determination of 0.9765 (Tab.2). The anomaly detection model successfully identifies temperature mutation points in multiple typical cities across China (Fig.3). Attribution analysis reveals significant spatial heterogeneity in the driving factors of temperature anomalies: population density generally exhibits a warming effect (contributing 14%-39% in most cities), while the roles of vegetation and evaporation vary across regions and urbanization stages (Fig.4)ConclusionsThis study develops an LSTM-based deep learning model that effectively integrates prediction and anomaly detection tasks, enabling the identification and interpretation of large-scale, long-term land surface temperature anomalies. The findings highlight the regional heterogeneity and interactive complexity of factors influencing urban thermal environments. However, limitations include uneven anomaly sample distribution across cities and constrained extrapolation capability under extreme climate events. Future work should incorporate physical constraints and more refined spatiotemporal data to enhance the model's explanatory power and generalizability.