A Methodological Study On Estimating Crop Yield Gap Based On Remote Sensing Models

Yield gap（YG）analysis between the actual crop yield（Ya）and potential yield（Yp）is important for revealing the possible opportunities to boost farmers’yield on existing farmland,and ensure food security.Up to now,the remote sensing（RS）-based approach for YG estimation is still in its infancy.For Ya simulation,the regionally parametrized RS models remain to be calibrated and validated with field measurements,due to the lack of high-resolution biophysical forcing variables.Moreover,given few studies evaluated the pros and cons of the empirical and the process-based model,the most appropriate model for simulating regional Ya remains to be recommended.Another pivotal issue is to estimate the attainable Yp with spatial heterogeneity,and identify the most potential region for yield improvement.To address the abovementioned issues,this study improves the RS-based framework for regional YG analysis from the following four aspects:（1）retrieve high-resolution crop biophysical variables,（2）explore the feasibility of multi-scale crop modelling with remote sensing,（3）assess the implications of different modelling paradigms for YG analysis and（4）propose the zonation scheme for estimating spatial distribution of attainable Yp.The improved framework is then applied to estimate Ya,Yp and YG for rice in Northeast China（NEC）.The specific themes are:（1）Develop a semi-empirical approach based on the Bayesian theory to retrieve Leaf Area Index（LAI）from multiple decametric satellites.Crop-specific distributions of PROSAIL input variables are first calibrated over a global dataset of ground GAI measurements（for maize,wheat,and rice）and the corresponding reflectance observations from Landsat-8,Sentinel-2,and Quickbird.These distributions are then used as prior information to predict LAI.Results show that the Bayesian approach provides close estimates of LAI to ground truth,with respective RMSE of0.97,1.01 and 1.33 for rice,maize and wheat（R²=0.63,0.67 and 0.76,respectively）.The performances are better than the SNAP algorithm for Sentinel-2.Moreover,the proposed approach possesses unique advantages of fully exploiting large observational datasets from multiple sensors,providing the associated uncertainties of the retrieving results,and compensating the model assumptions that depart from canopy realism.The high spatiotemporal estimation of LAI constitutes an essential data support for driving the RS-based crop model at the field scale.（2）Explore the potential of RS-based photosynthesis model for estimating field-scale biomass over diverse cropping systems.The RS-based photosynthesis model,describing key physiological processes of crop growth（e.g.,photosynthesis）,is a powerful tool for the Ya simulation.However,the regionally parameterized model structure is rarely calibrated or validation in the field.In this section,we force the model at the field scale with high-resolution LAI data,and calibrate a crop-specific parameter set with extensive field measurements.Results show that local calibration prevents the overestimation presented in those crude calibration cases.No significant bias is presented in the calibrated model,with the bias of 15.2g m^-2（1.8%）,and an overall RMSE of 261.9g m^-2（31.4%）,R² of 0.88 for all crops over the entire growing season.Simulated time series of biomass agrees well with the observed dynamics.This study explores the great potential of RS-based model to achieve multiscale crop modelling,which forms a solid model foundation for subsequent regional Ya simulation.（3）Comparative assessment of the process-based（PB）and machine learning（ML）models for regional Ya estimation.In this section,we first evaluate multiple ML algorithms for Ya estimation,in which a rotation scheme is proposed to correct the systematic model error caused by the underrepresented yield extremes.Then the corrected random forest model is chosen as the representative ML model,to be compared with the process-based photosynthesis model.The consistency and applicability of these two types of models are systematically evaluated in the simulation of rice yield over NEC.The county-level validation from 2006 to 2016 reveals that the simulated yields of both models are in good agreement with the census data,with the Pearson correlation coefficient（r）values of 0.63 and 0.70,and RMSE of 1355.3 kg ha^-1（17.3%）and 1205.5 kg ha^-1（15.4%）for the PB model and ML model,respectively.The PB model shows better accuracy in producing the probability distribution of Ya,while the ML model trained by the county average yields exhibits significant homogeneity over adjacent pixels.（4）Propose an environmental zonation scheme to portray the spatial patterns of attainable Yp and YG,and a quantitative method to identify the potential region for future yield improvement.We divide the entire study region into domains having similar climatic,geomorphic,and edaphic context,within which the high percentile of Ya is taken as the attainable Yp.Based on abovementioned scheme,the spatial distributions of Yp and YG are derived,and implications of different RS yield models for YG analysis are assessed.Results suggest the PB model as a more reliable and robust approach for estimating both Ya and Yp,and consequently YG,and illustrates the advantages of process-based modelling of crop growth at pixel scale with the aid of satellite data.Whereas the ML model presents significant underestimation of Yp and YG,which emphasizes the need to train ML models using data with field representativeness.For rice in NEC,the exploitable YG was 2454kg ha^-1,amounting to24.0%of Yp.Southern NEC possessed substantial YG with the primary priority to be explored.In this region,optimizing the farmer management practices will significantly improve current rice yields.While in Sanjiang Plain,Ya already approximated Yp（Ya accounting for>80%）.The room for yield gain is limited,which requires further technological progress or breeding breakthroughs to raise the upper limit Yp.