AUTOMATIC DIGITIZATION OF PAKISTAN’S AGRICULTURAL FARM BOUNDARIES VIA DEEP LEARNING

Abstract

Pakistan is primarily an agricultural country suffering from the lack of a national digitized field boundary dataset. Pakistani companies, both private and governmental are spending years upon years, dedicating resources to manual digitization procedures that are kept for organizational purposes only. With the advent of deep learning, digitization has become a process of mere seconds. But for Pakistan, again, this spells difficulties such as the lack of training data and a small and irregular fielding system that is common to the country, rendering it difficult to train a deep network that can accurately delineate Pakistan’s fields. Within this paper, we have trained a ResUNet model that delineates the fields of National Agricultural Research Center, Islamabad with an accuracy of 84%, a mean IoU of 0.65, and a loss of 0.40. We have managed to achieve this in a highly cost-effective manner, without utilizing Pakistani imagery or obtaining high resolution imagery. Finally, we have incorporated this model into a web dashboard that can take in an image uploaded by a user, predicts its mask, and delivers a downloadable shapefile of the digitized fields to the user. We have compared our model to the FarmSAM model, our finetuned version of Meta’s Segment Anything Model on field boundaries and shown that our model has exceedingly satisfactory results in face of the cost-effective methodology we opted for.