Wheat yield estimation using satellite images in Golestan province

Early prediction of wheat yield is major challenge in agrictultural management and food security. This requires regional estimation of yield which is costly and time consuming. Remote sensing (RS) is a fundamental approach for achieving practical and effective solutions for this problem. The existing differences in the spectral reflectance characteristic of agiven phenomena such as crop canopies, can serve as an identifier attribute . Wheat farms can have different yields based on their various growth and management conditions. These variations can affect the spectral reflectance. In this research, the yields data for year 2017 of 200 wheat farms located in Gonbad-Kavus and Voshmgir Dam regions, Golestan province north of Iran have been studied. After correcting the Landsat 8 satellite images of month of May, spectral bands data were extracted, and 15 spectral indices were calculated. The relation between spectral indices and wheat yield were worked out using multivariate linear regression and M5 regression tree model. According to the nonlinear relation between wheat yield and spectral reflectance, multivariate linear regression did not perform well. The correlation coefficient was about 63%, and the mean error was about 425 kg ha-1. The M5 regression tree model predicted wheat yield with an accuracy of 89 % and root mean square error (RMSE) lower than 325 kg ha-1. This accuracy was obtained using bands 1, 3, and NMDI and NDWI indices and band 4 to band 3 ratios (B43). Inaccurate yield data the non-uniformity and small area of the farms are the main sources of error.