Improving the ability of Landsat 8 NDVI in-season yield prediction by integrating the sequences of weather data to support smart farming

Abstract:

Smart farming requires ongoing monitoring of crop growth and development, and the ability to predict the effects of various management choices on crop yield and profit. Remote sensing, with its synoptic characteristics across multiple spatial and temporal scales, provides an opportunity to perform real-time monitoring and in-season yield prediction to support farm decision making. NDVI sequences from satellite images can be used to detect the timing of phenological events and calculate measures of crop growth. However, crop growth is driven by winter rainfall when clouds are common. The number of cloud contaminated images during the growing season limits the ability of remote sensing to provide accurate yield predictions.

In this study, we aimed to determine if weather data for the 16-day revisit period prior to the Landsat 8 data acquisition can be used to predict NDVI when Landsat 8 images are obscured by cloud. The weather data include: mean solar radiation, accumulated rainfall, accumulated growing degree days, number of cold days, number of hot days, mean vapor pressure deficit, and mean evapotranspiration deficit. We collected Landsat 8 datasets, daily weather records, and yield maps for 32 wheat paddocks in Western Australia in 2013 to 2019. We used random forests to predict missing NDVI using the integrated weather data and then assessed the use of the entire NDVI sequence for in-season yield prediction. 5-fold cross-validation was applied to estimate the prediction ability. Accuracy was measured by the correlation between observed and predicted yields (R) and mean absolute error (MAE). The results indicated that: 1) weather data from the 16-day period prior to Landsat 8 image acquisition can be used to predict Landsat 8 NDVI (R = 0.84; MAE = 0.05 averaged across the test datasets); 2) the in-season yield prediction ability of weather-estimated NDVI sequences were higher than the reconstructed NDVIs from August (the 14th time step for paddock-year) and onward.