Use of an AI-based annotation tool reduces inter-observer error linked to cover estimation of arable crops and weeds

Vegetation surveys are employed in agro-ecosystems to understand weed eco-biology and develop integrated weed management solutions. Vegetation cover through visual estimation is the most common, non-destructive measurement but is often associated to inter- and intra-observer variability. The present study investigates whether the inter-observer error linked to experience can be reduced after using an AI-based annotation tool that we developed. The tool combines general and domain-specific knowledge using Segment Anything Model and VegAnn to assist the annotator. Two vegetation surveys were performed at different growth stages of a lentil-buckwheat intercropping trial. Vegetation and soil cover were evaluated independently by two observers (expert and novice) on two sampling areas per plot. Each sampling area was photographed (Samsung Galaxy A42 smartphone) and the pictures were processed by the novice with the AI-based tool. The effect of the observer type (expert, novice, AI-tool) on the cover of crops, weeds and bare soil was tested using generalized least squares-mixed effect models. Overall, after using the AI-tool the gap between expert and novice decreased by 44 % for weeds but increased by 8 % for lentil, by 11 % for buckwheat, and by 4 % for soil. Nonetheless, when the spatial layout and the growth stage of the crops were considered, the gap between expert and novice was reduced in almost all cases. We also proved that the AI tool is useful for novice observer training prior to entering the field and which can be further developed to aid researchers and farmers in estimating vegetation cover.