Quantifying the impact of remotely sensed photosynthetically active radiation retrievals on empirical crop models in the United States

Abstract

Photosynthetically active radiation (PAR), is an essential component for life onEarth and one of the essential climate variables. Due to the differences in biochemistry, cell structure, and photosynthetic pathways, different plant species absorb PAR with varying efficiency and have evolved to thrive in different conditions, such as direct, intense sunlight or indirect, diffuse light conditions. Ground-based measurements allow for direct estimation of PAR; however, those are available in select locations, e.g. through the Surface Radiation Budget (SURFRAD) Network. Remote sensing-based methods, on the other hand, enable spatially explicit estimates of PAR on a regular basis. Current methods and models for satellite-based PAR retrievals require many ancillary atmospheric datasets as well as a large computing infrastructure. PAR, as one of the parameters influencing plant productivity, has not been previously used in the empirical crop yields and as such can lead to better satellite-based yield estimates. Having the advantages of spatially explicit PAR estimates, spatial and temporal patterns of the PAR can reveal differences in the land uses and the level of crop productivity. Therefore, the overarching goal of my dissertation is to advance the science of satellite-based PAR estimation and agricultural applications. This is done through the use of machine-learning models to reduce data input requirements for PAR estimation from daily Moderate Resolution Imaging Spectroradiometer (MODIS) acquisitions and by incorporating PAR into the empirical crop yield models over the US. In order to obtain satellite-based PAR estimates without the need for ancillary atmospheric data, I developed an empirical approach making use of machine learning methods as an efficient way to capture the non-linear relationship between top of atmosphere radiance and PAR at the surface. I found that the bootstrap aggregated decision tree (Bagged Tree), Gaussian Process Regression (GPR), and Multilayer Perceptron (MLP) yielded the best results with minimal input and training data requirements with an R2 of 0.77, 0.78, and 0.78 respectively, and a relative RMSE of 22-23%. While these results underperform compared with the look up table (LUT) approach, it does not require the same atmospheric parameters as input, such as atmospheric water vapor, aerosol optical depth, and others that might not be available in near real time or are only available at coarser spatial resolution. I incorporated MODIS-based PAR estimates into empirical corn and soybean yield models over the US. By explicitly adding PAR into the crop yield models, I found a maximum R2 of 0.81 and 0.80 for corn and soybean, respectively, whereas models that do not include PAR showed a maximum R2 of 0.60 for corn and soybean. By adding PAR directly into the empirical yield model and demonstrating additional explained variability, I show that my model is in closer agreement with process-based models than previous empirical models. I found that MODIS- derived coefficient of absorption of PAR (αPAR), which corresponds to the plant canopy chlorophyll content (CCC) and consequently productivity, corresponds to the ground-based αPAR measurements. Specifically, I found that for the US-Ne sites of corn and soybean fields in Eastern Nebraska R2 was 0.97 and RMSE was 1.34 (11%) when comparing MODIS-derived αPAR with the in situ measurements. I also found that the relationships between MODIS-based αPAR and CCC for corn and soybean corresponded to the ones obtained from in situ data. The relationships between αPAR and CCC for corn and soybean are distinct due to the different photosynthetic pathways of corn (C4) and soybean (C3), differences in cell structure, and chloroplast distribution between the two crops. Crop yield and productivity are also related to CCC, meaning αPAR can be used as a crop specific indicator of yield. Through this research, I have demonstrated the added value of incorporating PAR directly into crop yield models, by improving crop yield estimates over empirical models based on vegetation indices or surface reflectance alone. The research also provides the basis for further work using crop specific measures of the absorption of PAR into the same empirical models at large spatial scales that were previously impractical due to the spatial discrepancies between in situ- and MODIS- derived measurements.