Model-based scenario analysis for effective site-specific weed control on grassland sites

TL;DR

This paper evaluates and compares the effectiveness of tractor-based and robotic weed control systems in grasslands using model-based scenario analysis, considering spatial weed distributions and detection uncertainties.

Contribution

It introduces a simulation framework combining stochastic geometry and Monte Carlo methods to assess site-specific weed control strategies under various spatial and detection scenarios.

Findings

Robotic systems are more efficient for isolated weed distributions.

Tractor systems perform better in highly clustered weed scenarios.

Detection uncertainties significantly impact control strategy effectiveness.

Abstract

The site-specific management of weeds in grassland is often challenging because different weed control strategies have different trade-offs regarding the required resources and treatment efficiency. So, the question arises whether a wide tractor-based system with section control or a small agricultural robot has a higher weed control performance for a given infestation scenario. For example, a small autonomous robot moving from one weed to the next might have much shorter travel distances (and thus lower energy and time costs) than a tractor-mounted system if the locations of the weeds are relatively isolated across the field. However, if the plants are highly concentrated in small areas so-called clusters, the increased width of the tractor-mounted implement could be beneficial because of shorter travel distances and greater working width. An additional challenge is the fact that there is no complete knowledge of the weed locations. Weeds may not have been detected, for example, due to their growth stage, occlusion by other objects, or misclassification. Weed control strategies must therefore also be evaluated with regard to this issue. Thus, in addition to the driving distance, other metrics are also of interest, such as the number of plants that were actually controlled or the size of the total treatment area. We performed this investigation for the treatment of the toxic Colchicum autumnale, which had been detected in drone images of extensive grassland sites. In addition to real data, we generated and analyzed simulated weed locations using mathematical models of stochastic geometry. These offer the possibility to simulate very different spatial distributions of toxic plant locations. Different treatment strategies were then virtually tested on this data using Monte Carlo simulations and their performance was statistically evaluated.