A Mechanism-Coupled Split Window Network for Medium- to High-Resolution Land Surface Temperature Retrieval

Researchers propose PCD-Net, a neural network framework that combines physics-based split window algorithms with machine learning to improve land surface temperature retrieval from satellite thermal infrared data. The approach adaptively learns dynamic coefficients for atmospheric correction, addressing limitations of traditional fixed-coefficient methods and enhancing generalization across diverse environmental conditions.

This research advances Earth observation science by tackling a persistent challenge in remote sensing: accurately measuring land surface temperature across varying atmospheric and surface conditions. Traditional split window algorithms use fixed mathematical coefficients that fail when atmospheric water vapor spikes or surface temperatures exceed design assumptions, limiting their applicability in extreme climates or rapidly changing environments. The PCD-Net framework represents a paradigm shift by embedding physics-based equations within neural network architecture, allowing coefficients to dynamically adapt to real-time conditions rather than remaining static.

The innovation addresses a fundamental tension in machine learning: data-driven models achieve impressive accuracy but often fail catastrophically outside training data distributions, while physics-based models remain interpretable but inflexible. By decoupling the split window equation into parallel subnetworks that learn individual physical components—constant terms, brightness temperature differences, and coupling corrections—PCD-Net maintains physical transparency while gaining adaptive capacity. This component-level decoupling explicitly models relationships between surface emissivity and atmospheric water vapor, two variables that traditionally confound retrieval accuracy.

For climate science and environmental monitoring, this advancement enables more reliable thermal mapping of urban heat islands, agricultural stress, and ecosystem changes globally. The framework's improved generalizability particularly benefits emerging economies where validation datasets remain sparse. Applications extend beyond climate research into precision agriculture, urban planning, and disaster response. Developers building Earth observation platforms gain a more robust tool for integrating satellite data into decision-support systems. The work exemplifies how physics-informed machine learning can overcome longstanding barriers in geophysical parameter estimation without sacrificing model interpretability or computational efficiency.

• →PCD-Net combines physics-based split window algorithms with neural networks to dynamically adapt coefficients rather than using fixed empirical parameters.
• →The framework explicitly models component-level relationships between surface emissivity and atmospheric water vapor to improve accuracy under complex conditions.
• →Parallel subnetwork architecture enables better generalization to out-of-distribution samples compared to conventional data-driven models.
• →Improved land surface temperature retrieval enhances applications in climate monitoring, urban heat mapping, and precision agriculture.
• →The approach demonstrates how physics-informed machine learning can overcome traditional constraints of either purely empirical or purely statistical methods.