Integrating Landsat-8 imagery and spectroscopic data for accurate soil lead content estimation using the LeadNet model based on transfer learning

Heavy metal contamination in soils is a significant environmental issue . Traditional chemical assays are adequate but costly and geographically limited. Remote sensing offers a cost-efficient alternative for mapping soil lead, yet extensive training samples limit its utility. Model transfer across regions is gaining traction, but further refinement is needed to reduce sampling demands for digital soil lead mapping. This paper presents a novel model that combines spectrometry with LeadNet, an advanced deep learning model, to address key objectives—identifying spectral indices, constructing a simulated image, developing an inversion model, and mapping soil lead content. This model integrates spectroscopic and Landsat-8 spectra to improve soil lead estimation. Additionally, it employs a domain adaptation strategy to transfer the feature knowledge, securing new sites compatibility and preventing overfitting. In this regard, the results' study reveal that Normalized Pigment Difference Indices (NPDI) surpass Normalized Difference Spectral Indices (NDSI) in identifying spectral features relevant to lead content due to their more pronounced correlation with lead content. NDPI-2 (885, 1875, and 2095 nm) with R and VIP values of 0.66 and 1.27, respectively, was the most significant index, while NDSI-1 (1005 and 1455 nm) with R and VIP values of 0.55 and 0.07 was the least important. LeadNet demonstrated superior performance over traditional multi-layer neural networks (MLNN), achieving an R 2 of 0.52 in reference regions (137 samples) and 0.37 in target regions (80 samples). Using optimal spectral indices from Landsat-8, LeadNet's initial mapping was refined with an interpolated residual map via the Inverse Distance Weighting (IDW) method. The final aggregated map, achieved through algebraic summation of corresponding pixels, exhibited an R 2 value of 0.73, affirming the model's high accuracy and its practical utility in environmental monitoring, providing a reliable tool for soil lead concentration assessment.