Integrating Time-Series Nighttime Light Data With Static Remote Sensing and Village View Images for Hollow Villages Identification

Accurately identifying hollow villages (HVs) has long been a challenge in rural governance and revitalization. Traditional field surveys require significant human and material resources, making large-scale identification difficult. This study develops a model that integrates static and dynamic data for HV identification. The model uses a ResNet18 with an attention module to extract static features of villages from remote sensing imagery and village view images, and employs an LSTM-FCN to analyze periodic human activity changes from nighttime light (NTL) data to extract dynamic features. Evaluated in four Guangdong counties, the multisource data approach outperforms single-source models, achieving a test overall accuracy of 0.8451, a kappa index of 0.6391, and an F1 score of 0.8880. The human activity patterns reflected by time-series NTL data play a significant role in the identification of HVs. The multisource data model helps to mitigate the biases inherent in individual data types. This approach provides a reliable solution for the rapid identification of HVs.