Deep Representation Learning‐Driven Self‐Supervised Label Generation for Automated Long‐Term Land Use/Land Cover Classification Using Remote Sensing and GIS

ABSTRACT In this study, we reconstruct multi‐decadal Land Use and Land Cover (LULC) dynamics (1990–2025) in the Pearl River Delta (PRD) using a deep representation learning–driven, self‐supervised framework applied to harmonized Landsat archives. High‐confidence pseudo‐label cores were derived from percentile‐based spectral–phenological rules, refined with robust Median Absolute Deviation (MAD) filtering, then embedded with an era‐specific denoising autoencoder and classified with Random Forest (RF). Over the span of four Landsat generations, classification performance proved consistent, with overall accuracy frequently above 0.80 and peaking at 0.914 ± 0.060 during the Landsat‐9 era. The PRD noticed substantial and diverse land redistribution between 1990 and 2025. The highest relative expansion was in urban land, which increased from 1931.43 to 4547.26 km 2 (+2615.83 km 2 ; +135.4%), and in forest area, which increased from 15,863.85 to 20,509.68 km 2 (+4645.83 km 2 ; +29.3%). A notable spike in agricultural land was observed, from 10,495.05 to 12,193.43 km 2 (+1698.38 km 2 ; +16.2%). On the other hand, the amount of water reduced from 8004.79 to 4675.04 km 2 (−3329.75 km 2 ; −41.6%), and the amount of bare land decreased marginally from 18,041.47 to 17,270.08 km 2 (−771.39 km 2 ; < 4.3%). A systematic, uneven distribution was also evident in the intensity analysis, with the two main transition pathways being barren‐to‐water (34.92%) and barren‐to‐urban (9.29%). These findings demonstrate that self‐supervised representation learning (SSRL) enables temporally consistent LULC reconstruction while diagnosing directional land reallocation in highly heterogeneous megaregions.