Biology-Informed inverse problems for insect pests detection using pheromone sensors

Most insects have the ability to modify the odor landscape in order to communicate with their conspecies during key phases of their life cycle such as reproduction. They release pheromones in their nearby environment, volatile compounds that are detected by insects of the same species with exceptional specificity and sensitivity. Efficient pheromone detection is then an interesting lever for insect pest management in a precision agroecological culture context. A precise and early detection of pests using pheromone sensors offers a strategy for pest management before infestation. In this paper, we develop a biology-informed inverse problem framework that leverages temporal signals from a pheromone sensor network to build insect presence maps. Prior biological knowledge is introduced in the inverse problem by the mean of a specific penalty, using population dynamics PDE residuals. We benchmark the biological-informed penalty with other regularization terms such as Tikhonov, LASSO or composite penalties in a simplified toy model. We use classical comparison criteria, such as target reconstruction error, or Jaccard distance on pest presence-absence. But we also use more task-specific criteria such as the number of informative sensors during inference. Finally, the inverse problem is solved in a realistic context of pest infestation in an agricultural landscape by the fall armyworm ( Spodoptera frugiperda ).