A methodology to assess the accuracy and reliability of yield-monitor data

摘要

The ability to accrue full benefits from the adoption of precision agriculture (PA) technologies depends on having complete confidence in the data layers on which decisions are based. Previous research has shown that yield data can reliably show spatial patterns in within-block yield, but the process of filtering, manipulating and attributing yield-sensor data to sugar mill records for the block from which it was derived can have considerable bearing on the utility and reliability of the resultant yield maps. This paper reports on the production of data handling tools to evaluate, clean and prepare the collected sensor data (harvester position/speed, chopper pressure/speed, feed roller separation, elevator pressure/speed and weigh pad) for input into a yield-mapping protocol for rigorous map generation. The tools developed use Python scripts have been developed using freely available Python libraries and Excel macros. The data manipulation steps include aggregating data packets, clipping to block boundaries, identifying individual harvest events, aligning with mill records, predicting yield and kriging the data. Applying the filtering protocol in an automatic fashion vastly reduced the total time required for the task, while ensuring that all harvest events were processed consistently and resulting in improved confidence in the resultant maps. The block mean yields derived from each yield map generated compares well with the mill tonnages for the individual blocks. In comparing the mill tonnage with the sensor-derived average yields for each block, all sensor yields are within the 95% confidence interval of the mill tonnage for all except one block. This is well within the level of accuracy in commercial yield sensors sold for other crops such as grains or wine grapes. The time to process the data has also been greatly reduced - from weeks to hours. The procedures and tools reported on in this paper have gone a long way towards automating the analysis of yield monitor data. They have enabled the datasets to be treated in a consistent and regimented fashion, with only limited manual input. This has improved confidence in the data on which the derived yield maps are based.