Towards a true conservation zero tillage system: “A proposed solution based on computer vision to herbicide resistance”

Abstract

The Zero Tillage (ZT) or Non-Tillage system constitutes an agricultural production approach designed for improved soil conservation, reduced fossil fuel usage, mitigation of waterway pollution, water and wind erosion, and addressing soil compaction, among other objectives. In this way, ZT promises more sustainable agriculture. However, current ZT production systems depend on herbicide applications for weed control. The use of herbicides with the same modes of action over many years has led to numerous resistant weed species, which end up threatening the success of the herbicide weed control and, consequently, the overall success of the ZT system. Being able to automatically detect and classify weed species directly under the sprayer during the chemical application is an alternative to control populations of weeds since it would allow selecting the herbicide and dose for those particular species. This study, conducted over 15 years in commercial fields utilizing Zero Tillage (ZT) management, examined various ground covers intrinsic to the ZT system originating from previous crops or residues post-harvest. ZT features a broad spectrum of contexts due to the complexity of different ground cover types, which challenge computer vision techniques. The assessment focused on the automatic detection and classification of among the most problematic monocotyledonous and dicotyledonous weed species in the province of Cordoba in Argentina, with the objective of enabling a more targeted and selective approach to their control. These are Amaranthus palmeri, Echinochloa crus-galli, Eleusine indica, Parietaria debilis and Conyza sumatrensis. Additionally, some species belong to the same botanical families with several morphological resemblances that defy vision algorithms. This work proceeded through a machine learning approach applied to computer vision features such as SIFT (scale-invariant feature transform), K-means (clustering), and Bag of Features over the Support Vector Machine for classification. The results showed an accuracy between 89 % and 98 % over the species, allowing a significant reduction of herbicide while applying the adequate active ingredient. Moreover, virtual binary maps from weed patches are devised to be implemented through ISOBUS protocol. Thus the current research contributes significantly to the issue of controlling populations of weeds in the ZT agriculture system.