Advances in multispectral and hyperspectral inversion for soil heavy metal contamination: Mechanisms, machine...

Heavy metal (HM) contamination in soil exhibits insidious and cumulative effects, posing long-term risks to ecosystems and human health. Traditional field sampling and laboratory analysis are increasingly insufficient for large-scale continuous monitoring, driving the adoption of multispectral (MS) and hyperspectral (HS) remote sensing. Bibliometric analysis reveals clear research trends: target elements are primarily copper, lead, and zinc, while data acquisition has progressively shifted from laboratory spectroscopy to portable devices and satellite platforms, reflecting an expansion from local to regional scales. This has led to increasing data complexity and greater demands on model robustness and generalization. However, expanding the spatial scale and transitioning to satellite observations introduce fundamental challenges. Mixed pixels and moisture-induced spectral distortions reduce signal purity, while the indirect spectral response of HMs further complicates quantitative inversion. Sample scarcity and spatial heterogeneity also limit cross-regional generalization, constraining model robustness and stability. In response, models have evolved from traditional linear regression to ensemble learning methods such as Extreme Gradient Boosting (XGBoost), and further to deep learning frameworks, including Convolutional Neural Networks (CNN) and Transformers, enabling hierarchical feature extraction and task-oriented structural design. This paper reviews the key technical bottlenecks in soil HM spectral inversion, integrating bibliometric insights with methodological advances to provide a comprehensive framework for understanding current progress and guiding future developments in large-scale, high-precision inversion.