Search

目的/意义人工智能（Artificial Intelligence，AI）技术已在学术和工程应用领域掀起了研究高潮，在地球物理参数和农业气象遥感参数反演方面也表现出了强大的应用潜力。目前大部分AI技术在地学和农学的应用还是“黑箱”，没有物理意义或缺乏可解释性及通用性。为了促进AI在地学和农学的应用和培养交叉学科的人才，本研究提出基于AI耦合物理和统计方法的地球物理参数反演范式理论。方法首先基于物理能量平衡方程进行物理逻辑推理，从理论上构造反演方程组，然后基于物理推导构建泛化的统计方法。通过物理模型模拟获得物理方法的代表性解以及利用多源数据获得统计方法代表性的解作为深度学习的训练和测试数据库，最后利用深度学习进行优化求解。［结果和讨论］判定形成具有通用性和物理可解释的范式条件包括：（1）输入与输出变量（参数）之间必须存在因果关系；（2）输入和输出变量（参数）之间理论上可以构建闭合的方程组（未知数个数少于或等于方程组个数），也就是说输出参数可以被输入参数唯一确定。如果输入参数（变量）和输出参数（变量）之间存在很强的因果关系，则可以直接使用深度学习进行反演。如果输入参数和输出参数之间存在弱相关性，则需要添加先验知识来提高输出参数的反演精度。此外，本研究以农业气象遥感中的关键参数地表温度、发射率、近地表空气温度和大气水汽含量联合反演作为案例对理论进行了证明，分析结果表明本理论是可行的，并且可以辅助优化设计卫星传感器波段组合。结论本理论和判定条件的提出在地球物理参数反演史上具有里程碑意义。

The land surface temperature (LST) changes in North America are very abnormal recently, but few studies have systematically researched these anomalies from several aspects, especially the influencing forces. After reconstructing higher quality MODIS monthly LST data (0.05° * 0.05°) in 2002-2018, we analyzed the LST changes especially anomalous changes and their driving forces in North America. Here we show that North America warmed at the rate of 0.02 °C/y. The LST changes in three regions, including frigid region in the northwestern (0.12 °C/y), the west coast from 20°N-40°N (0.07 °C/y), and the tropics south of 20°N (0.04 °C/y), were extremely abnormal. The El Nino and La Nina were the main drivers for the periodical highest and lowest LST, respectively. The North Atlantic Oscillation was closed related to the opposite change of LST in the northeastern North America and the southeastern United States, and the warming trend of the Florida peninsula in winter was closely related to enhancement of the North Atlantic Oscillation index. The Pacific Decadal Oscillation index showed a positive correlation with the LST in most Alaska. Vegetation and atmospheric water vapor also had a profound influence on the LST changes, but it had obvious difference in latitude.

Soil moisture (SM) and land surface temperature (LST) are entangled, and the retrieval of one of them requires a priori specification of the other one. Due to insufficient observational information, retrieval of LST and SM from passive microwave remote sensing data is often ill-posed, and the retrieval accuracy needs to be improved. In this study, a novel fully-coupled paradigm is developed to robustly retrieve SM and LST from passive microwave data, which integrates deep learning, physical methods, and statistical methods. The key condition of the general paradigm proposed by us is that the output parameters of deep learning can be uniquely determined by the input parameters theoretically through a certain mathematical equation. Firstly, the physical method is deduced based on the energy radiation balance equation. The nine unknowns require the brightness temperatures of nine channels to construct nine equations, and the solutions of the physical method equations are obtained by model simulation. Based on the derivation of the physical method, the solution of the statistical method is constructed using multi-source data. Secondly, the solutions of physical and statistical methods constitute the training and test data of deep learning, which is used to obtain the solution curve of physical and statistical methods. The retrieval accuracy of LST and SM is greatly improved by smartly utilizing the mutual prior knowledge of SM and LST and cross iterative optimization calculations. Finally, validation indicates that the mean absolute error of the retrieved SM and LST data are 0.027 m3/m3 and 1.38 K, respectively, at an incidence angle of 0–65°. A model-data-knowledge-driven and deep learning method can overcome the shortcomings of traditional methods and provide a paradigm for retrieval of other geophysical variables. The proposed paradigm not only has physical meaning, but also makes deep learning physically interpretable, which is a milestone in the retrieval of geophysical remote sensing parameters based on artificial intelligence technology. [ABSTRACT FROM AUTHOR]

Accurate predictions of sea surface temperature (SST) are crucial due to the significant impact of SST on the global ocean-atmospheric system and its potential to trigger extreme weather events. Many existing machine-learning-based SST predictions adapt the traditional iterative point-wise prediction mechanism, whose predicting horizons and accuracy are limited owing to the high sensitivity to cumulative errors during iterative predictions. Therefore, this paper proposes a novel granulation-based long short-term memory (LSTM)-random forest (RF) combination model that can fully capture the feature dependencies involved in the fluctuation of SST sequences, reduce the cumulative error in the iteration process, and extend the prediction horizons, which includes two sub-models (adaptive granulation model and hybrid prediction model). They can restack the one-dimensional SST time-series into multidimensional feature variables, and achieve a strong forecasting ability. The analysis shows that the proposed model can achieve more accurate prediction-hours in nearly all prediction ranges from 1 to 125 h. The average prediction error of the proposed model in 25–125 h is 0.07 K, similar to that (0.067 K) in the first 24 h, which exhibits a high generalization performance and robustness and isthus a promising platform for the medium- and long-term forecasting of hourly SSTs. [ABSTRACT FROM AUTHOR]

Ta (Near-surface air temperature) is an important physical parameter that reflects climate change. Although there are currently many methods to obtain the daily maximum (Tmax), minimum (Tmin), and average (Tavg) temperature (meteorological stations, remote sensing, and reanalysis data), these methods are affected by many factors. In order to obtain daily Ta data (Tmax, Tmin, and Tavg) with high spatial and temporal resolution in China, we fully analyzed the advantages and disadvantages of various existing data (reanalysis, remote sensing, and in situ data). Different Ta reconstruction models are constructed for different weather conditions, and we further improve data accuracy through building correction equations for different regions. Finally, a dataset of daily temperature (Tmax, Tmin, and Tavg) in China from 1979 to 2018 was obtained with a spatial resolution of 0.1°. For Tmax, validation using in situ data shows that the root mean square error (RMSE) ranges from 0.86 °C to 1.78 °C, the mean absolute error (MAE) varies from 0.63 °C to 1.40 °C, and the Pearson coefficient (R2) ranges from 0.96 to 0.99. For Tmin, RMSE ranges from 0.78 °C to 2.09 °C, the MAE varies from 0.58 °C to 1.61 °C, and the R2 ranges from 0.95 to 0.99. For Tavg, RMSE ranges from 0.35 °C to 1.00 °C, the MAE varies from 0.27 °C to 0.68 °C, and the R2 ranges from 0.99 to 1.00. Furthermore, a variety of evaluation indicators were used to analyze the temporal and spatial variation trends of Ta, and the Tavg increase was more than 0.0 °C/a, which is consistent with the general global warming trend. In conclusion, this dataset had a high spatial resolution and reliable accuracy, which makes up for the previous missing temperature value (Tmax, Tmin, and Tavg) at high spatial resolution. This dataset also provides key parameters for the study of climate change, especially high-temperature drought and low-temperature chilling damage, which is publicly available with the following DOI: https://doi.org/10.5281/zenodo.5502275 (Fang et al., 2021a). [ABSTRACT FROM AUTHOR]