Machine learning and task disambiguation in hand-picked agriculture

Abstract

Although GPS-based travel data has been studied by many mainly for automated travel mode detection, the area of activity mode detection during harvest still remains an open technical challenge. This thesis proposes and tests a pattern recognition approach to harvest mode recognition from GPS travel data collected from 4 volunteers for 2 days in Oxnard, California. Three profiles were created to characterize activities performed during harvest. Piecewise quadratic interpolation was used on smoothened data to detect segments in trips taken by workers. Trip segments are then evaluated with the different profiles to find the best fitting profiles and the associated optimal parameters. Results indicated that the proposed framework performs well under data discrepancies. Identification of different modes during harvest is of relevance for assessing productivity of different workers and addressing any mismatch in vehicle scheduling. In our assessment, this proof-of-principle study demonstrates a use case for using GPS data in disambiguating different activities conducted during harvest; scalability of the methodology remains a challenge - programming GPUs to take advantage of independence in the different processes has been proposed to reduce the code runtime.