Innovative Sunshine Duration Observations with AI: Bridging the Gap in Climatological Data
Liu, Dongwei ; Tan, Jianguo ; Wang, Yadong ; Shi, Jun ; Ao, Xiangyu ; He, Xiaochuan ; He, Qianshan ; Mu, Haizhen ; Hou, Wenxuan ; Sun, Juan ; Peng, Jie ; Liu, Miaomiao
Année de publication
2025
Sunshine duration is an important parameter to characterize local climatology and has been observed for a long time. We propose a novel method for observing sunshine duration using video images, based on a Large Vision-Language Model (LVLM) named Vision-and-Language Transformer (ViLT). The method can directly identify weather phenomena from video images without the need for fine-tuning and can determine whether sunshine is present based on ViLT's assessment of whether it is a sunny day or not. Evaluation results indicate that the cumulative percentage errors of the annual sunshine duration derived from video observations are 1.3% and 1.1% at the Chongming and Xujiahui stations, respectively. These errors are well below the World Meteorological Organization's (WMO's) achievable measurement uncertainty for sunshine duration (2% or 0.1 h). Since the method only requires images from standard video equipment as input, it offers a cost-effective and convenient technology for observing sunshine duration, particularly in areas lacking sunshine observations but with available video data. Further studies demonstrate that when combined with the Ångström-Prescott formula, the method can be utilized to estimate total solar radiation from video data. Compared to observations from solar radiometers, the percentage error in the estimated total annual solar radiation is around 4.3%, indicating promising applications of the method in solar resource assessment and climate studies.</div>
Accès à la notice sur le site du portail documentaire de Météo-France