The Ability Of OCO-2 Sun-induced Chlorophyll Fluorescence Data To Estimate Regional Crop Yield

In recent years,great challenges have been imposing on agricultural development and stability by the increasingly frequent and intensive climate change and extreme weather events.Under such circumstances,it is,therefore,more than crucial to monitor the growth and production of crops in a timely,effectively,and accurately way.Various studies have demonstrated that satellite sun-induced chlorophyll fluorescence(SIF)is promising in observing and monitoring growing conditions and productivity of terrestrial vegetation.With the development of remote sensing techniques,Orbiting Carbon Observatory-2(OCO-2)SIF data achieved a relatively high spatial resolution of 1.3km ×2.25 km.Nevertheless,it is still unclear the ability of OCO-2 SIF data with the nature of spatial discontinuity to monitor regional crop yield,especially when compared with common satellite vegetation indices,such as the MODIS EVI,and climate data.Therefore,this study leveraged the state-of-art deep learning technique(i.e.,transfer learning and DNN)to estimate corn and soybean yield in the U.S.Corn Belt based on OCO-2 SIF,MODIS EVI,and auxiliary climate data.Moreover,the contribution of different combinations of these data to crop yield estimation was explored,and then the uncertainty of the crop yield forecasting model was analyzed.The major conclusions are as follows:(1)The ability of OCO-2 SIF in estimating regional corn and soybean yield.In the study area,both SIF and EVI have strong correlation relationships with corn or soybean yield,and the relationships also shared similar variations.On the other hand,the DNN based on SIF data had lower accuracy than DNNs that inputting either EVI or climate data.The DNN only inputting SIF,EVI,or climate data had mean absolute percentage error(MAPE)of 8.99%,5.59%,and 5.34% for corn,and 8.9%,6.79%,and 5.55% for soybean,separately.Additionally,it was found that the DNN performances were influenced by the level of areal dominance of crops for the yield forecasting models that only input either EVI or SIF,or both EVI and SIF data,while the performances were less impacted when the DNN was built only based on the climate information.(2)The estimating accuracy could be improved when combining SIF,EVI,and climate data in crop yield estimation.The performances of the DNN based on SIF,EVI,and climate data is better than that of the models based on either EVI or climate data.Specifically,the DNN based on SIF and climate data had MAPE values of 5.24% for corn and 5.89% for soybean,while the DNN inputting SIF and climate data had MAPE values of 4.9% for corn and 5.18% for soybean.When inputting SIF,EVI,and climate data,the DNN for corn and soybean yield estimation generated MAPE of 4.08% and 4.45%,respectively.(3)The feature importance analysis demonstrated that during peak growing months(from June to August),MODIS EVI and climate predictors,such as VPD and temperature,is important for the corn and soybean yield estimation in the DNN that inputting SIF,EVI,and climate data.Additionally,SIF showed less importance score than that of the EVI and climate variables.