This paper is motivated by a vehicle routing problem, arising in the sub-ground analysis field, where the objective is to analyze the subground composition to better understand its properties. In this field, operations cover very wide areas, involving several thousands of points visited each day. To insure an efficient analysis, points can only be visited if they belong to a sensor-covered area, as sensors are needed to get the measures obtained while visiting a point. Since subground analysis operations can last for a very long duration (several months), it is not worth considering to place sensors on the whole area to study. The equipment is then progressively deployed based on the operation progress: when one of the equipment-covered areas is finished, sen- sors are redeployed on the next zone. Point disponibility, based on the sensor-covered area, depends on the period considered. The problem can be modelized as a Team Orienteering Problem, with mandatory and optional points. When every mandatory point of an area is visited, sensors are moved and the sets of mandatory and optional points are modified. The objective of the problem is to maximize the number of visited points, while respecting all field-specific constraints. Among others, vehicles need to be constantly geographically separated, and the mandatory area needs to be quickly covered. We present a LNS-based approach for this problem. The proposed algorithm consists of three phases aiming to get high-quality solutions in a reasonable computational time. First, a clustering algorithm based on k-medoids is applied to the problem, enabling an implicit management of some constraints. Then, several independent sub-problems , based on the initial clusters, are built and optimized simultaneously using a Large Neighborhood Search (LNS). Finally, when all sub-problems are optimized, the initial partitioning is reassessed by applying a local search on nearby routes, not belonging to the same sub-problem . The algorithm is tested on literature instances. On these instances, the proposed metaheuristic is relevant as it finds solutions of similar quality than state of the art algorithms in much less computational time, also improving three best known solutions. When applied to industrial instances, the LNS-based approach finds better solutions than the ones used by the company in a reasonable computational time, confirming the relevance of the pro- posed methodology.