Assessment of the effectiveness of spatiotemporal fusion of multi-source satellite images for cotton yield estimation

Deficiencies in the spatiotemporal resolution of remote sensing (RS) images limit crop yield estimation at the farm and field scale. These deficiencies may be alleviated by fusion of high spatial and temporal resolution images such as MODIS and Landsat. In this study, a new daily MODIS NDVI product (reconstructed MODIS) was generated from 16-day composite images using the Extreme Model, which integrates the NDVI value with the corresponding specific date information at each pixel. The Flexible Spatiotemporal Data Fusion (FSDAF) model was then used to create two fused, high-resolution time-series products (fused MODIS and fused reconstructed MODIS) in order to enhance the spatial and temporal effectiveness of satellite images for field-scale applications. Three yield estimation models were then built using time-series data of Landsat NDVI, predicted NDVI from fused MODIS, and predicted NDVI from fused reconstructed MODIS. The methodology was tested on a farm field over the cotton growing season in the San Joaquin Valley of California. Results showed that: (1) the time trend of NDVI over the growing season for the fused reconstructed MODIS was more similar to that of Landsat than were either of MODIS or fused MODIS, indicating that the specific date of MODIS pixels is important for time-series analysis; (2) the NDVI from fused reconstructed MODIS provided the best correlation with Landsat NDVI, with R2 and RMSE values 15% higher than for fused MODIS; (3) correlation between cotton yield and all three datasets at the pixel level was statistically significant for all image dates, and (4) the accuracy of the cotton yield estimation model using predicted NDVI from fused reconstructed MODIS (R2 = 0.79; RMSE = 488.01) was higher than with fused MODIS (R2 = 0.77; RMSE = 513.96) and only slightly lower than with Landsat (R2 = 0.84, RMSE = 463.12). This study improved the accuracy of MODIS-based yield estimation using fusion images, and the results can be applied to improve vegetation monitoring and quantitative modeling using MODIS NDVI at the field scale.