Introduction

Currently, about 60 million anti-personnel (AP) mines are polluting the environment in about 60 countries. Because mine clearance operations proceed much more slowly than mine laying, the number of polluting mines is still increasing. Humanitarian mine clearance operations must be understood and designed correctly, keeping in mind that their main goal is to provide efficient aid to innocent people, who may be severely injured by this dreadful pollution. Further, the analysis of actual demining campaigns primarily reveals the far too long time needed to clear polluted terrain, a far too large false alarm rate, the threat of plastic mines, difficult to detect by classical means (metal detectors), the large variety of the mine clearance scenarios, depending on the country, the region, the climate and the place of the pollution (houses in villages, roads, agricultural fields, etc).
The important parameters which characterize the mine detection problem are the mine occurrence probability, the detection probability of a given material and the false alarm probability of a given material [2]:

• The mine occurrence probability in a given position of a minefield expresses the local mine density of that minefield as well. Obviously, it is impossible to control this parameter because it depends on the terrain reality. Nevertheless, this parameter is very important for assessing the probability of an alarm in a given location of the minefield.
  
• The detection probability is the probability of having an alarm in a given position of a minefield for a given detection material, if there is a mine in that position. This probability gives indirectly a measure of the non-detection probability of that material as well.
  
• The probability of false alarm is the probability of having an alarm, for a given material, in a given location if there is no mine in that location.
  

The two latter definitions are extremely important to understand the humanitarian demining problem and for designing demining systems.
It is indeed particularly important that the detection probability should be as close as possible to one. It is easy to show that evaluating the detection probability also amounts to evaluate the risk of the occurrence of a mine which has not been detected. This risk is concerned with human preservation and is therefore of extreme importance. No such risk is acceptable and it is therefore an absolute requirement that a demining system should decrease the probability of such a risk to the lowest upper bound possible (UN requires 0.4% maximum).
Besides, although one indirectly saves human lives by decreasing the false alarm risk thanks to the acceleration of the demining operations, the false alarm risk is also a question of cost. Indeed, a demining method which minimizes the false alarm rate results in an acceleration of the demining operations which results in a money profit.

Therefore, any demining operation enhancement must result in the highest possible detection probability (close to one) and in the smallest possible false alarm rate and that at the lowest price. Generally, it is accepted that the most efficient way for increasing the detection probability while minimizing the false alarm rate consists in using several complementary sensors in parallel and in fusing the information collected by these sensors.

As a matter of fact, it is imperative to evaluate the detection probability when optimizing the performances of a system. However, the detection probability, as it is defined before, assumes that a mine is present in the considered position. Since, during organized trials, the position of the mines is well known, the condition of the occurrence of a mine in the given position where the performances of a system must be evaluated is always realized. This latter remark is of particular importance because it justifies the organization of trials and the construction of models, to be validated by trials, in order to evaluate the detection probabilities.

Furthermore, assuming in the following as the first approximation that the sensors are independent¹, the detection probability can be maximized by optimizing separately the design of each sensor and of the associated signal processing. Next, it can easily be shown that the detection probability increases if the number of different sensors increases and that maximizing the overall detection probability of a set of independent sensors clearly comes to the same as maximizing the detection capabilities of each individual sensor. This justifies the use of several complementary sensors and of data fusion techniques to increase the detection probability. Among the most cited sensors one finds the metal detectors, the radars and the infrared sensors.

Finally, the false alarm risk, i.e. the probability of having an alarm if there is no mine, cannot be as easily evaluated as the detection probability because of the use of data fusion methods which favor the manual or automatic cancellation of false alarms. Furthermore, it is very difficult to evaluate the risk of false alarm because it is very difficult to define in a general way what is not a mine. In this context, it should be particularly inappropriate that a demining system, whatever it may be, makes decision instead of the final user whose own physical security is involved. Therefore, a well designed system should help the user in the decision making, not by replacing him, but by implementing efficient data fusion methods. For this purpose, methods which are able to deal with uncertainty by making proposals including the doubt to the user seem to be promising.

The rest of the paper tries to fit with the previous reasoning. The first step consists in acquiring knowledge on sensors by means of trials explained in section 2. As explained in section 3, the second step consists in developing models for the description of the ground electromagnetic behaviour, in investigating the capabilities of new sensors (hyper-spectral imagery, nuclear quadrupole resonance, ... and educated rodents) and in enhancing the capabilities of existing sensors (Ground penetrating radars, metal detectors and infrared sensors). The third step means making each of these sensors skilled specialists of their respective domain (e.g. mine metallic content detection for the metal detector), as explained in the section 4.1 which analyses specific preprocessing tools and in section 4.2 which describes some dedicated pattern recognition tools. The last steps sketched out in section 5 consists in fusing the high level information produced by the different experts (the sensors with their dedicated processing tools).