Sunflower Yield Estimaton Using Synthetic Aperture Radar and Optical Data
Özet
The scope of the presented thesis, it is aimed to obtain highly accurate yield estimation by applying Artificial Neural Networks (ANN) and Simple Linear Regression (SLR) methods on sunflower parcels in a certain area using Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical satellite images. The 2018 Sentinel-1 and Sentinel-2 satellite images of the sunflower farming region in the Zile district of Tokat province were examined. First, the Normalized Difference Vegetation Index (NDVI), Normalized Difference Vegetation Red-Edge-1 (NDVIR1), Optimized Soil-Adjusted Vegetation Index (OSAVI), and Inverted Red-edge Chlorophyll Index (IRECI) were obtained from Sentinel-2 satellite images. Then, the correlations between these indices and reference yield data were examined separately. The highest correlation in all examined vegetation indices examined was determined on 30 June, 8 July, and 10 July. Using the SLR method, OSAVI (R2=0.75 RMSE=22.47 kg) gave the best result during the heading period on 30 June. However, when the indices are processed as a single input value in the ANN method, the best result is obtained from NDVI (R2=0.76 RMSE=22.07 kg) in the 30 June heading period. The ANN method gave better results than the linear regression method on all dates. The best estimation study was made with the ANN method in the NDVI and NDVIR1. When the four indices were processed together as input values, the best result was again obtained during the heading period on 30 June (R2=0.78 RMSE=21.59 kg). In the ANN method, it was concluded that using the indices together as inputs reduces the error rate in yield estimation. The phenological stages and yield values of the sunflower plant were determined with high accuracy by the ANN method. Secondly, backscattering and coherence values obtained from Sentinel-1 satellite images produced close values in both VH and VV polarizations in most of the dates considered, but in the ANN method, VV polarization produced better values than VH polarization. When the yield estimation was made with the backscatter values, the best result in the ANN method was obtained in the VV polarization on June 29 (R2=0,09 RMSE=42,46 kg). When the yield is estimated with the coherence values, the best result was obtained on the 5th of July-11th in the VV polarization (R2=0.01 RMSE=46.24 kg). The results showed higher errors compared to optical data. When the R2 value and the average RMSE are compared with the values obtained from the indices values, it is seen that the contribution of the backscatter and coherence values is very little or not in the yield estimation. The use of backscatter and coherence values alone in the yield estimation is not recommended for this study. Lastly, when the values obtained from optical and SAR satellites are used together, the best result was obtained during the heading period between 29 June - 5 July (R2=0.67 RMSE=26.83 kg). The estimated yield values were found with acceptable accuracy. However, the values did not improve.
