基于人工智能的农业气象遥感关键参数联合反演方法

The Joint Inversion Method for Key Parameters of Agricultural Meteorological Remote Sensing Based on Artificial Intelligence

关键词:  人工智能  热红外遥感  农业气象遥感参数  联合反演

Key words:  Artificial Intelligence  Thermal Infrared Remote Sensing  Agricultural Meteorological Remote Sensing Parameters  Joint Inversion

摘要:

【目的】为了提升农业气象遥感关键参数—地表温度(LST)、地表发射率(LSE)、大气水汽含量(WVC)和近地表空气温度(NSAT)的反演精度,解决传统方法未充分利用参数间相互依赖关系导致整体精度不足的问题。【方法】本文提出了一种基于人工智能(AI)的热红外遥感多参数联合反演方法。该方法通过深度学习耦合物理辐射传输方程和统计方法,首先推导物理反演模型,明确参数反演需满足的两个条件,获得物理代表性解；然后基于多源数据构建统计代表性解,组合两者形成深度学习训练与测试数据库。反演框架先直接反演LST和LSE,再将其作为先验知识,通过交叉迭代优化WVC和NSAT的反演精度,消除不确定性,实现多参数联合反演。【结果】利用MODIS 27、28、29、31、32波段进行实际反演应用分析。数据分析表明,5个波段组合下LST相对反演误差MAE和RMSE分别为0.42K和0.71K。LSE31和LSE32对应相对反演误差MAE分别是0.004、0.003,RMSE都是0.005。WVC的反演相对误差 MAE和RMSE分别0.46g/cm₂^和0.62 g/cm₂^。NSAT的反演相对误差 MAE和RMSE分别1.26 K和1.86 K。NSAT反演引入LST和LSE后,反演精度得到了提升,并更加稳定。地面观测站点数据验证也得到了相似的结论。【结论】联合反演策略下,LST和LSE精度稳定,WVC和NSAT通过先验知识迭代优化克服了传统方法精度不足的局限。分析表明,基于AI的热红外遥感多参数联合反演方法充分利用参数间关系,大幅提升整体反演精度,对农业气象遥感和卫星传感器设计具有重要意义。

Abstract:

[Purpose] To improve the inversion accuracy of key agricultural meteorological remote sensing parameters—Land Surface Temperature (LST), Land Surface Emissivity (LSE), Atmospheric Water Vapor Content (WVC), and Near-Surface Air Temperature (NSAT)—and address the issue of insufficient overall accuracy caused by traditional methods not fully utilizing the interdependencies between parameters. [Method] This paper proposes an artificial intelligence (AI)-based multi-parameter joint inversion method for thermal infrared remote sensing. The method integrates deep learning with physical radiation transfer equations and statistical methods. First, a physical inversion model is derived, and the two conditions required for parameter inversion are clarified to obtain a physically representative solution. Then, a statistical representative solution is constructed based on multi-source data, combining both solutions to form a deep learning training and testing database. The inversion framework directly inverts LST and LSE, using them as prior knowledge to iteratively optimize the inversion accuracy of WVC and NSAT through cross-iteration, eliminating uncertainty and achieving multi-parameter joint inversion. [Results] Practical inversion application analysis was performed using MODIS bands 27, 28, 29, 31, and 32. Data analysis showed that the relative inversion errors for LST under a five-band combination were 0.42K for MAE and 0.71K for RMSE. The relative inversion errors for LSE31 and LSE32 were MAE values of 0.004 and 0.003, with RMSE values of 0.005 for both. The inversion relative errors for WVC were 0.46 g/cm2 for MAE and 0.62 g/cm2 for RMSE. The relative errors for NSAT inversion were 1.26K for MAE and 1.86K for RMSE. After introducing LST and LSE into NSAT inversion, the inversion accuracy was improved and became more stable. Similar conclusions were also confirmed by ground observation station data. [Conclusion] Under the joint inversion strategy, the accuracy of LST and LSE remained stable, while WVC and NSAT overcame the limitations of traditional methods" accuracy by iterative optimization with prior knowledge. Analysis indicates that the AI-based multi-parameter joint inversion method for thermal infrared remote sensing effectively utilizes the relationships between parameters, significantly enhancing the overall inversion accuracy, which is of great importance for agricultural meteorological remote sensing and satellite sensor design.