Big data-driven spectral index construction for fine-scale salt pond mapping

• Big data-driven framework generated > 390 K candidate spectral indices of salt pond. • New indices capture subtle spectral signature difference among salt pond subtypes. • Integration of new indices raised RF classification accuracy from 72.93% to 81.14%. • First fine-scale salt pond subtype map produced for the entire Bohai Rim region. Accurate mapping of salt pond systems is crucial for coastal resource management and ecological conservation. However, existing spectral indices for salt pond detection are mostly derived empirically, limiting their ability to differentiate pond subtypes or generalize across diverse coastal environments. This study proposes a big data-driven framework for constructing novel spectral indices that enhance the fine-scale classification of salt ponds in the Bohai Rim region of China. By systematically generating and screening over 390,000 candidate indices derived from Sentinel-2 bands, four new indices were identified to effectively discriminate salt pond subtypes: the Non-Water Body Detection Index (NWBDI), Crystallization Pond Index (CPI), Red-Green Concentration Pond Index (RGCPI), and Evaporation Pond Discrimination Index (EPDI). To rigorously evaluate their performance, three Random Forest configurations were compared, including a baseline model with original spectral bands, an expert-index-enhanced model, and a model incorporating the proposed novel indices. Across the Bohai Rim region, the model with proposed new indices achieved an overall accuracy (OA) of 81.14%, demonstrating clear advantages over both original spectral bands (OA of 72.93%) and expert-index approaches (OA of 76.64%). The resulting fine-scale map revealed a total salt pond area of 2,214.33 km 2 in 2020 across the Bohai Rim, consistent with national statistics. Beyond improving classification accuracy, the data-driven index discovery revealed physically meaningful spectral relationships linked to salinity gradients and hydro-biogeochemical properties, marking a methodological shift from expert-driven hypothesis testing to automated, data-driven feature discovery. This study demonstrates that the data-driven approach can provide a transferable solution for constructing task-specific spectral indices and advancing large-scale environmental monitoring.