*3.3. Identification of Threshold for When to Blend*

The incomplete time-series pixels of Landsat are gap-filled with the L–M blended pixels (*LLM*). The threshold to indicate when blending is beneficial is identified by quantifying the impact of the fraction of missing data on the prediction accuracy at 25 m. Firstly, we group MODIS and *LLM* time series in eight groups, based on the fraction of Landsat missing data (<10%, 10–20%, 20–30%, 30–40%, 40–50%, 50–60%, 60–70%, and 70–80%). Then, the accuracy of each group is analyzed by calculating the *R*<sup>2</sup> and the *RMSE*. The performance of C-Crop using MODIS and *LLM* is compared by the fraction of missing data in Landsat across time. The threshold can be identified, when the model provides the same *R*<sup>2</sup> and *RMSE* using *LLM* as with MODIS (Figure S3, Supplementary Materials). This threshold determines when, where, or how much L–M blended data improves crop yield prediction when the fraction of missing data in Landsat is lower than the identified threshold.

#### *3.4. Evaluation of the Improvement in Yield Prediction Accuracy*

The improvement in prediction accuracy using the identified threshold for multiple spatio-temporal data selection is statistically quantified. More specifically, the threshold is applied to the Landsat observations (2000–2018). Firstly, we compute the temporal probability of optimally using MODIS, Landsat, and *LLM* images for nationwide crop yield predictions during the past two decades, and then map the results to illustrate the spatial variability of multi-sensor data selection for 25-m pixel-level yield prediction across the wheatbelt. We then evaluate the area percentage of the data sources on a yearly basis and analyze their potential correlation to the annual precipitation (mm/year). Finally, the improvement in the accuracy of predicted yields is evaluated on the field level using MODIS and *LLM* for Western and eastern Australia, against the reported data [58]. The growing season of 2015 is selected due to the availability of a larger quantity of observed yield data. The incomplete 2015 Landsat series are gap-filled with the blended values corresponding to the threshold value.
