Current Progress in and Future Visions of Key Technologies of UAV-Borne Multi-Modal Geophysical Exploration for Mineral Exploration: A Scoping Review

Abstract

For mineral exploration, an increasing number of geophysical instruments have adopted unmanned aerial vehicles (UAVs) as their carrier platforms. The effective fusion of multi-modal geophysical information will be conducive to further enhancing the reliability of exploration results. However, the integration degree of UAVs and geophysical equipment is still low, and the advantages of UAVs as robots have not been fully exploited. In addition, the existing fusion methods are still difficult to use to establish the spatial distribution model of ore-bearing rock. Therefore, we reviewed the development status of UAVs and the geophysical instruments. We believe that only by integrating the system, designing the observation plan in accordance with the requirements of the fusion method, and treating the hardware part as an external extension of the algorithm, can high-matching data be provided for fusion. Subsequently, we analyzed the progress of the fusion methods, leading us to believe that the cross-dimensional and cross-abstract-level issues are major challenges in the algorithm aspect. Meanwhile, the fusion should be carried out simultaneously with the generation of the ore-bearing rock model, that is, to establish an integrated system of fusion and generation. It is hoped that this research can promote the development of UAV-borne multi-modal observation technology.

1. Introduction

Mineral exploration is a complex process that integrates multiple disciplines, including geology, geophysics, geochemistry, and remote sensing, and can be divided into several stages, such as prospecting, general exploration, and detailed exploration [1]. At each stage, the specific objectives, work content, and associated technical methods differ. In recent years, with the extensive reporting of the continuous application scenario expansion of the unmanned aerial vehicle (UAV) [2,3,4,5], the public’s understanding of the application logic of UAVs has also been constantly developing [6]. In the field of mineral exploration, compared with carrying platforms such as satellites and manned aircraft, employing civilian UAVs equipped with various detection instruments offers advantages in terms of low cost and high flexibility. Therefore, this approach is especially well-suited for the middle and later stages of the mineral exploration process to achieve the fine detection of the spatial distribution and morphology of ore-bearing geological bodies. From this, it can be observed that an increasing number of pieces of geophysical observation equipment, such as optical sensors [7,8,9], LIDAR devices [10,11,12], GPR devices [13,14,15], magnetometers [16], electromagnetic equipment [17], spectrometers [18], and even gravimeters [19], are actively attempting to utilize UAVs as their carrying platforms [20,21,22]. In informatics, the term “modality” refers to a source or form of information [23]. In geoscience, each independent method of observing the Earth represents a way to obtain information about the Earth’s parameters, and can thus be regarded as an independent modality. Within the scope discussed in this paper, the collection of the aforementioned various methods forms a multi-modal observation system. With this low-cost multi-modal observation technology, exploration units can create an exclusive local big data space for each deposit [24], and through the effective processing of multi-modal information, obtain more intuitive and comprehensive information on ore-bearing geological bodies [25,26,27,28,29].

However, what is not temporarily in line with this promising prospect is that, constrained by the performance of equipment and algorithms, UAV-borne multi-modal geophysical exploration is still a simple superposition of various detection technologies, and has not yet demonstrated the system advantages arising from the synergy of multiple detection technologies. In fact, the goal of multi-modal fusion is to form new knowledge and make new judgments by processing, correlating, and analyzing the observed information from different modalities. Within the scope discussed in this paper, how to establish a spatial distribution model of the ore-bearing geological body as accurately as possible is of great significance for subsequent work such as resource estimation and mine planning [30]. However, the existing key technologies in equipment and algorithms still face challenges in supporting the realization of this goal [31].

In terms of instruments, under the current technological status quo, how many types of geophysical equipment can be carried by low-cost UAVs? This actually demarcates the informatics boundaries for subsequent research. Meanwhile, the differences among the geophysical instruments in terms of the sensitivity principles to the features of the earth, the resolution capabilities of observed field quantities, the distribution forms of the observation arrays, and the reliability of the observation systems lead to the differences in the performance of the detection results in describing the features of the earth [32]. Ignoring these differences will have an adverse impact on information integration. Therefore, how to systematically optimize the UAV-borne multi-modal observation system with the goal of enhancing the integration effect is the first issue that needs to be considered in terms of instruments.

In terms of algorithms, multi-modal observations necessarily require fusion algorithms. However, on the one hand, for different geophysical methods, their data may have different dimensions in structure, and it is still difficult to achieve cross-dimensional fusion at present. On the other hand, we also hope to consider other non-geophysical observation results, which will provide support for establishing mapping rules between geophysical parameters and underground rock distribution. However, these observations may be given in non-data forms (such as text form), which means that there are differences in the abstraction levels of information from different sources. How to solve the fusion cross-abstraction levels is another problem that needs to be addressed.

In response to the above issues, this study addresses the development of UAVs and the geophysical instruments that can be carried by them and then reviews the current research progress of multi-modal geophysical information fusion methods. On this basis, in light of the fusion requirements, the innovative observation technologies based on UAVs and fusion methods based on artificial intelligence are discussed and prospected. This study aims to further promote the research and application of UAV-borne multi-modal geophysical methods in mineral exploration through the aforementioned review and prospect.

2. Method

This research was conducted following the PRISMA guidelines [33]. To identify potentially relevant documents, the bibliographic database Web of Science™ (Clarivate™, Philadelphia, PA, USA) was searched. The search strategies (Appendix A) were drafted and further refined through team discussions. As shown in Figure 1, our research follows the logical chain of “exploration demand—exploration equipment—fusion method—output result”. For the middle and later stages of mineral exploration, a precise understanding of the spatial distribution of ore bodies is of great significance for resource assessment and the formulation of subsequent development plans. This is our exploration demand, and it also means that the final result submitted should be a spatial distribution model of the target ore-bearing rock mass. Between the demand and the final goal lies the main body of our research, which includes the related hardware and algorithm technologies. Hardware technology mainly includes UAV technology and geophysical equipment technology, which mainly solve the problem of multi-modal information observation. Due to the key differences between the data dimensions and information abstraction dimensions of the multi-modal observation data, how to effectively fuse multi-modal information is the main issue discussed in the algorithm technology aspect.

As shown in Figure 2, 258 articles were retrieved from Web of Science™, and combined with other sources, resulting in a total of 280 articles being obtained. After an initial screening, 236 articles were retained. Based on the themes involved, we conducted a detailed review of each of these documents and further excluded 75 articles. Among the excluded literature, in terms of hardware, it mainly involved optical, LiDAR, hyperspectral, and ground-penetrating radar sensors, and other equipment. As the research focus of this paper is on the middle and later stages of mineral resources, with the aim of establishing a spatial distribution model of underground ore-bearing rock bodies within a certain depth range, the above equipment has certain limitations in meeting the above goals. In terms of algorithms, a considerable number of studies have investigated fusion methods based on map structure, and the technical solutions have overlapping aspects. Therefore, the relatively less representative studies have been excluded. Ultimately, 162 articles were included in this study.

We conducted a preliminary analysis of this literature. As shown in Appendix B, the scope of consideration for references is divided into the following: within 5 years, within 10 years, and others, with a primary focus on the latest achievements within 5 years. The search range for references mainly includes important journals in the fields of geophysics, remote sensing, and UAVs, as well as high-impact-factor journals in comprehensive journals.

After completing the data collection, we first randomly selected 60 articles (approximately 25% of the total number of articles) and assigned them to each co-author for independent research. First, they determined whether the selected articles met the research requirements; then, they analyzed the articles according to the research objectives, compared the conclusions of each author, and discussed any differences.

Currently, research on enhancing the fusion effect of multi-modal information more effectively remains in its infancy, particularly within the geophysics domain. The influence of differences in the information carrying capacity among various observation modes on fusion outcomes has yet to be extensively discussed. To address this issue, we systematically analyzed the current state of equipment and methodologies, evaluating their shortcomings against the aforementioned criteria. Through summarizing common problems, we established a comprehensive understanding and outlined potential future technological directions.

Furthermore, it should be pointed out that UAV-borne multi-modal geophysical exploration and information fusion is a cutting-edge technology in an interdisciplinary field, making it very difficult to conduct a scoping review. Additionally, our search results only extend to early 2025, and there are new technologies that had not yet been disclosed by this time point. This is the limitation of this research.

3. UAVs and Geophysical Instruments

The development history of UAVs can be traced back to the early 20th century [34]. The development of high-performance UAVs and miniaturized geophysical equipment has laid the foundation for creating a private local big data space. Compared to manned aircraft, the advantage of UAVs lies in their flexibility, while the disadvantage lies in their carrying performance, which makes it difficult to carry some of the heavier and/or power-demanding equipment. In recent years, with the rapid improvement in UAV performance and the miniaturization of geophysical equipment, the technology of UAV-borne geophysical observation has become increasingly advanced [35].

3.1. UAVs Used in Geophysical Observation

UAVs constitute a vast and complex industrial category. From the perspective of geophysical observation methods, we will review the development of UAVs used in the geophysical industry in recent years.

In the civilian sector, UAVs typically use the gas recoil force as their takeoff power source [36], and more than 90% of them employ vertical takeoff methods [37].

Table 1. Three types of UAVs.

Multirotor UAVs	Tandem Rotor UAVs	Rotary Wing and Fixed-Wing Compound UAVs
Power: Motor	Power: Turboshaft Engine	Power: Motor
Maximum Payload: 30 kg	Maximum Payload: 310 kg	Maximum Payload: 6 kg
Maximum Range: 28 km	Maximum Range: 800 km	Maximum Range: 300 km
Maximum Speed: 20 m/s	Maximum Speed: 50 m/s	Maximum Speed: >20 m/s
Wind Resistance: 12 m/s	Wind Resistance: 17 m/s	Wind Resistance: 12 m/s
Maximum Ceiling: 6000 m	Maximum Ceiling: 6500 m	Maximum Ceiling: 7500 m

presents typical examples of three types of vertical takeoff UAVs that have entered the Chinese geophysical exploration market in the past five years. Table 1 presents five important quantifiable parameters, which are significant in the following aspects: (1) Maximum Payload, which limits the weight of geophysical equipment that can be carried, thereby restricting the types of methods supported by this type of UAV; (2) Maximum Range, which limits the reuse rate of the UAV’s takeoff and landing fields, which is particularly important for complex terrain survey areas; (3) Maximum Speed, which limits the maximum work efficiency of a single UAV flight; (4) Wind Resistance, which limits the key meteorological conditions for UAV operation; and (5) Maximum Ceiling, which limits whether the UAV can perform specific flight tasks in high-altitude environments, such as calibration flights using aeromagnetic methods and background noise observations using airborne electromagnetic methods. Additionally, the table also includes the power forms of these three types of UAVs, which in practice means the probability of on-site repair after a UAV failure.

The first category is the currently more common multirotor UAV, with the example in the table being the Flycart 30 from DJI [38]. In fact, compared to many industrial-grade multirotor UAVs launched by little-known companies, the performance parameters of the Flycart 30 are not even the most outstanding. Multirotor UAVs are the earliest type of UAVs to enter the field of mineral and engineering geophysical exploration. The most common instrument for them is fluxgate magnetometers or optical pumping magnetometers, both of which are usually relatively light in weight [39]. As the load-carrying capacity of such UAVs improves, they are also being used in other methods, such as the semi-airborne electromagnetic method [40,41]. The second category is heavy-lift tandem rotor UAVs, with the example in the table being the MK450 model from Xcontrol [42]. Similar products include those from AviDrone, a Canadian company [43]. Compared to multirotor UAVs, these UAVs not only have a greater carrying capacity, but also offer a larger internal space for equipment installation. They are expected to serve as platforms for carrying equipment for airborne electromagnetic methods, radiometric surveys, and gravity observations. The third category is the rotary wing and fixed-wing compound UAVs, with the example in the table being the V10 model from Feima Robotics [44]. Compared to the first two types of UAVs, this type combines the excellent aerodynamic performance of fixed-wing aircraft with the vertical takeoff and landing capabilities of rotor UAVs. It is particularly suitable for carrying optical–hyperspectral sensors for surveys, with significantly improved efficiency compared to multirotor UAVs carrying similar payloads. In addition, we also note that, apart from carrying capacity, these UAVs all focus on the capability of gust resistances and service ceilings. Improvements in these metrics will help expand the areas where survey work can be conducted. These advancements indicate that nearly all airborne geophysical survey equipment can or will soon be able to be mounted on UAVs.

When geophysical instruments are carried by UAVs, the impact of the UAV on the observation system needs to be considered. This mainly includes electromagnetic interference and vibration interference [45]. Regarding the above issues, the related research shows the following characteristics: (1) research on fixed-wing UAVs and unmanned helicopters is more abundant, while research on multirotor UAVs is relatively scarce; (2) research on magnetic field sensors is more prevalent, while studies on other sensor types are relatively few. This is mainly because, traditionally, aeromagnetic survey technology based on manned aircraft has a relatively high level of maturity, and the related technologies are easily adapted to UAVs with similar structural forms. Generally speaking, the basic solution to avoid electromagnetic noise is to increase the distance between the sensor and the UAV [46], while the basic solution to avoid vibration interference is to use flexible or elastic materials to connect the payload and the UAV [47]. In practice, different sensors have different requirements for suppressing the above-mentioned interference, and the forms of the carrying structures will also vary accordingly. Due to the differences in control logic and the performance of flight controllers among different UAVs, if a relatively complex structure is to be carried beneath the UAV through a rigid connection, the center of gravity position and stability of the structure need to be considered. Especially for multirotor UAVs, the addition of complex carrying structures may compress the performance margin for attitude adjustment, thereby causing flight safety risks. On the contrary, if a simple flexible structure is adopted, new problems arise for many sensors that require precise orientation and positioning. Moreover, the motion-induced noise caused by swinging will also seriously interfere with the observation [48,49]. In addition to the common problems mentioned above, for aeromagnetic surveys, the noise sources from UAVs mainly include permanent magnetization, induced magnetization, and eddy currents [50,51]. The related magnetic compensation techniques are relatively mature on fixed-wing UAVs and unmanned helicopters [52], but for multirotor UAVs, they are still in the initial stage [53,54]. To establish the relationship between the maneuver of UAVs and the corresponding magnetic field changes, calibration flights need to be conducted in a specific manner [55]. Even if it is assumed that there are no technical issues, there may still be regulatory restrictions [56]. Overall, there seems to be a contradiction here: on the one hand, the design of UAVs is a highly specialized business—if UAV suppliers are reluctant to get involved in the optimization design of the carrying structure for geophysical equipment, it is actually very difficult to solve the above problems by relying solely on sensor suppliers or users; on the other hand, since UAVs are recognized as a low-cost carrier platform, from the perspective of general users, there seems to be a lack of motivation to solve the above problems to the best of their ability. This contradiction also outlines the inherent logic of the current low integration degree between UAVs and geophysical instruments from one aspect.

3.2. Advances in the Geophysical Instruments

Traditionally, airborne geophysical exploration encompasses several key techniques: aeromagnetic surveys, aerial gravity surveys, aerial radiometric surveys, and airborne electromagnetic surveys. From the perspective of acquiring Earth observation data, optical/hyperspectral and radar sensors, which belong to the field of remote sensing, are also closely related to airborne geophysical exploration. For mineral exploration, GPR is relatively less applied due to its limited detection depth. The first type of equipment to be widely used was optical/hyperspectral equipment [57,58,59,60,61]. Such equipment is generally used to identify surface geometry and mineral characteristics and is typically applied in the earlier stages of the exploration process or to provide auxiliary information for establishing the spatial model of ore-bearing geological bodies in the middle and later stages. For this reason, we will discuss the related fusion algorithms in the subsequent section.

The development of airborne geophysical exploration equipment has been ongoing for over half a century, with manned fixed-wing aircraft and helicopters primarily serving as the carrying platforms. Among this equipment, the miniaturization of airborne magnetic survey sensors has developed relatively rapidly, and the detection accuracy (based on the magnetic compensation model) has gradually been able to reach the sub-nT level [62], which has been widely reported in recent years. Taking China as an example, according to incomplete statistics, over 30 companies and institutions are engaged in the research on, development, and application of equipment with different magnetic sensors and structures in 2024. In contrast, the development of UAV-borne equipment related to aerial gravity, aerial radioactivity, and airborne electromagnetic surveys has been relatively slow, and these technologies are still some distance away from market maturity. In the following text, we will address the research progress of these types of equipment.

3.2.1. Aeromagnetic Equipment

Aeromagnetic survey equipment was the earliest geophysical equipment to introduce UAVs as carrying platforms [63,64,65,66]. In these early studies, the sensors used were mostly heavier than 15 kg, so fixed-wing UAVs or unmanned helicopters were mainly adopted. At this stage, due to the high structural similarity between UAVs and traditional aircraft, the integration of UAVs with sensors was relatively high. Some UAVs were even developed as subsystems of the entire aeromagnetic system. However, as UAV technology continues to advance, the R&D model through which the suppliers of aeromagnetic systems develop UAVs has become increasingly unsustainable [67,68]. Meanwhile, the gradual maturation of multirotor UAV technology has provided an opportunity to lower the cost threshold for aeromagnetic detection [46,69]. Due to the limited carrying capacity of such UAVs, the main logic of the system’s development has shifted to reducing the weight of the equipment (which to some extent means sacrificing performance) in exchange for the feasibility of the system. In addition, for this type of UAV, through reasonable system structure design, lightweight bracket material selection, and even the introduction of certain aerodynamic designs, the gradient measurement system similar to the traditional structure can also be constructed [70,71,72]. From this, we can see that a large number of lightweight and miniaturized optical-pumped magnetometers and fluxgate magnetometers have been widely applied [16,73,74,75]. After the development of the above two generations, the development of aeromagnetic technology has come to a crossroads: one option calls for stronger-performance UAV that can carry the existing high-precision but heavy detection systems; the other option continues to insist on innovation in sensors, developing lightweight and high-precision sensors. The superconducting full-tensor magnetic gradient measurement technology based on traditional aircraft has now matured [76,77], with a measurement accuracy reaching the level of ±10 pT/m/√Hz. However, the sensor of such equipment, the superconducting quantum interference device (SQUID), needs to be cooled with liquid helium, so the weight and volume of its working condition maintenance system are both relatively large. In recent years, with the rapid development of 100 kg-class heavy-lift UAVs, UAV-borne SQUID equipment has gradually become practical [78,79,80]. For another technical route, the main development in recent years has been the miniaturization of atomic magnetometers [81], mainly referring to coherent population trapping atomic magnetometers (CPTs) [82], while spin-exchange relaxation-free atomic magnetometers (SERFs) [83] and nonlinear magneto-optical rotation atomic magnetometers (NMORs) [84] are more suitable for near-zero magnetic field environments. Compared with traditional optical-pumped magnetometers, CPTs can overcome the measurement dead zone problem when the magnetic field lines are nearly horizontal in low-latitude regions, and its dynamic range can cover from 100 to 100,000 nT [85]. The theoretical predicted sensitivity limit of CPTs is 0.1 fT/√Hz [86], and the sensitivity of several commercialized products has reached the level of several 10 pT/√Hz [87,88], indicating that there is still room for the development of CPTs. In terms of the miniaturization of quantum instruments, diamond quantum vector magnetometers have the advantages of a small size, low power consumption, and high sensitivity [89]. As shown in Figure 3, this is the working principle of such sensors [90]. The nitrogen-vacancy (NV) center is a photoluminescent defect in the diamond lattice [91]. The negatively charged NV center is sensitive to various physical quantities and can be read using the optical detection of magnetic resonance (ODMR) method. When in a magnetic field, the degeneracy of the $\left|m_S=\pm 1〉\right.$ states is lifted, resulting in Zeeman splitting, which is the detection mechanism. Green laser excitation of the NV center generates red photoluminescence, which serves as the readout mechanism of the magnetic resonance. Some media reports have indicated that they have been applied in the detection scenarios of UAVs [92]. In addition, in recent years, with a deeper understanding of the performance of yttrium iron garnet (YIG) films, the theoretical sensitivity of the new fluxgate sensors based on them will be less than 1 fT/√Hz, and they have the advantages of small size and high stability, which is a revolutionary development of the traditional fluxgate technology [93]. Currently, how to further improve the performance of YIG films by doping other elements is still being researched, and it will also provide new technical possibilities for unmanned aerial vehicle aeromagnetic detection in the future.

3.2.2. Gravimeter

The concept of aerial gravity measurements dates back to the mid-20th century [94]. Overall, it has evolved from scalar gravity measurements based on zero-length spring technology [95] to the development of full-tensor gravity gradient meters based on quantum technology [96] and atomic interferometry technology [97]. Currently, the most mature and widely used equipment types in the exploration market are three-axis inertial stabilized platform-based gravimeters and strapdown inertial navigation system-based gravimeters [98]. The three-axis inertial stabilized platform-based aerial gravimeters are represented by the AIRGrav system from Sander Geophysics Limited of Canada and the GT system from Gravimetric Technology, Russia. The main parameters of both systems are quite similar [99]. However, these types of gravimeters are typically heavy (often exceeding 100 kg), power-hungry (in the order of hundreds of watts), expensive, and require operation during use. As a result, they place very high demands on the performance of the carrying platform. By contrast, strapdown gravimeters are better suited as they are lightweight, making them more suitable for integration into UAV platforms. The University of Calgary in Canada was the pioneering institution in the development of strapdown gravimeters [100]. Since then, numerous institutions in Russia, Germany, Denmark, and China have undertaken significant and productive research in this field [101,102,103,104,105]. The fundamental factor limiting UAVs from being the carrying platform for gravity meters is still the weight of the gravity meter, as parameters such as power consumption and volume ultimately translate into weight. Reducing the weight of the gravity meter will ultimately affect the detection accuracy, typically resulting in a decrease in accuracy by nearly one order of magnitude compared to manned systems with the same technological setup. However, the trade-off is the more competitive system acquisition and operational costs. In fact, the use of fixed-wing UAVs [106,107] as platforms for carrying gravity meters was an early choice, reflecting this balancing approach within the technological conditions of the time. However, fixed-wing UAVs typically have poor low-altitude and low-speed performance, which still imposes limitations on detection accuracy. With the leap-forward development of IMU technology (iMAR), rotary wing UAVs (initially unmanned helicopters), which offer good low-altitude and low-speed performance, but have a relatively limited carrying capacity, began to be used as carrying platforms [108]. In ideal conditions, the precision of INS/GNSS systems (when combined with advanced processing techniques) can achieve sub-mGal accuracy [19]. As the weight of the IMU constitutes a major part of the weight of such lightweight and miniaturized systems, and the total weight of the entire system can generally be controlled under 20 kg, this has in fact fallen within the carrying capacity range of industrial-grade UAVs in recent years. As shown in Figure 4, a system with a small, unmanned helicopter as the carrier platform was reported in reference [109]. Since 2024, reports have emerged of UAVs being used as carrying platforms for gravity measurement systems, including the system developed by DroneSOM based on the Arcsky X55 UAV [110], as well as the gravity measurement system product released by China State Shipbuilding Corporation Limited at the 2024 Zhuhai Airshow, which is based on DJI UAVs. In addition to the aforementioned equipment types, despite some technical challenges, MEMS-based gravity meters, which can balance both miniaturization and accuracy, hold significant development potential [111,112]. They could potentially bring new breakthroughs to gravity observations based on UAV platforms in the future.

3.2.3. Gamma Spectrometer

Compared to other airborne geophysical methods, aerial radiometric surveying (aerial gamma spectrometry) plays an irreplaceable role in the detection of radioactive minerals. In fact, beyond geophysics, aerial radiometric surveying also includes nuclear radiation detection tasks. These primarily involve source localization and signature collection, as well as real-time mapping of environmental radiation distribution around nuclear facilities or the radioactive plume following a radiological event. Therefore, different application scenarios present significant differentiated requirements for both the detector and the UAV in terms of detection accuracy, operational efficiency, working environment, and coverage area [113]. Based on the different ways in which detectors interact with radiation, they are generally classified into three types: gas detectors, scintillation detectors, and semiconductor detectors. In geophysical observations, the primary detectors used are scintillation detectors and composite detectors, primarily based on scintillation detectors [114]. The main parameters affecting the performance of scintillation detectors are the sensitive volume of the detector crystal and the crystal resolution. Improving the resolution requires selecting crystals with larger volumes and higher resolutions, which in turn increases the weight of the detection system and raises the performance demands on the flight platform. For example, a 4L NaI crystal weighs approximately 30 kg, which is already beyond the capacity of most current civilian UAVs. Gamma spectrometers based on multiple NaI crystals generally exceed 100 kg, and traditionally, they can only be carried by manned aircraft. Since the maximum penetration distance of gamma rays in the air is approximately 200 m, low-altitude and even ultra-low-altitude slow-flight observations can, to some extent, reduce the demands on the crystal performance. Nevertheless, it is still necessary to consider whether the UAV has sufficient carrying capacity. Therefore, UAV-based aeroradiometric surveying technology essentially requires finding a balance between detection requirements, crystal performance, and the flight platform. Due to the large “dynamic range” of these three factors, there is more than one potential balance point between them, leading to the formation of key advancements in UAV-based aerial radiometric surveying technology in recent years. Sanada & Torii [18] reported the development of a 6.5 kg detection system using three LaBr3:Ce scintillation detectors (diameter 38.1 mm × length 38.1 mm) on a UAV helicopter platform. Šálek et al. [115] developed a system with two 103 cm³ Bismuth Germanium Oxygen (BGO) scintillation detectors, with a total system weight of approximately 4 kg, carried by a hexacopter UAV. Parshin et al. [116] and others developed a series of lightweight gamma spectrometers (weighing less than 1 kg) using CsI (T1) crystals combined with Silicon Photomultiplier (SiPM) technology for specific low-altitude detection tasks. Similarly, Van der Veeke et al. [117] employed larger CsI scintillators, with the detector weight reaching the 2 kg range. As shown in Figure 5, the system developed by Kunze et al. has two detector systems (installed beneath the flight control compartment of the UAV) [118]: one is solely based on a CeBr₃ scintillation detector, and the other is a combined system of CeBr₃ and NaI. To further enhance detection performance under the dual constraints of limited weight and cost, multiple small, high-resolution scintillators (mainly CeBr₃) can be arrayed [119], using list-mode measurement [120] to eliminate the adverse effects of Compton scattering on detection efficiency and energy resolution. Further research is still needed to address related issues such as correcting and compensating for the scintillator’s own radioactive background [121] and the cross-response effects between adjacent scintillators. In conclusion, considering the cost constraints, both the detection requirements and crystal selection in the aforementioned balance must be compromised to align with the actual capabilities of the UAV. Therefore, as the performance of low-cost UAVs continues to improve, it will undoubtedly drive the development of more powerful and cost-effective aerial radiometric equipment technologies.

3.2.4. Electromagnetic Equipment

In the field of electromagnetic methods, there are currently two kinds of airborne electromagnetic exploration equipment: airborne systems (mainly time domain methods) and semi-airborne systems (which combine both time domain and frequency domain methods). Airborne methods require all equipment to be mounted on the flight platform. The system is compact, flexible, and efficient. However, due to the limitations of the platform’s power supply and carrying capacity, the transmitted power and antenna aperture are restricted, resulting in a typical detection depth of around 600 m. To address the above issues, the transmitting part of the system is placed on the ground, using a long-grounded wire source as the transmitting antenna, while the flight platform only carries the sensors and receivers for airborne observations. This approach ensures high operational efficiency while achieving an ultra-high detection depth of 1000 m. This mode is referred to as the semi-airborne electromagnetic method. In addition, there is a method based on the magnetic field transfer function, where a horizontal magnetic field observation device is installed on the ground, and a vertical magnetic field observation device is mounted on the flight platform. The data is processed using the tipper calculation principle to obtain the detection results at great depths. Because this method requires simultaneous observations on the ground and in the air, and may even involve deploying sources on the ground to augment natural fields, it is sometimes categorized as a semi-airborne method [41]. In comparison, the semi-airborne approach is more suitable for deployment on UAVs, as it only requires the mounting of sensors and receivers [122], as shown in Figure 6. For tipper measurements using natural sources as the field source, the required sensors are typically large in size and heavy in weight. Additionally, since the sensors used in such methods generally have lower resonant frequencies, the achievable spatial resolution is relatively limited. The combination of these two factors makes it difficult for conventional short- endurance electric multirotor UAVs to serve as suitable platforms. Therefore, hybrid-powered rotorcraft UAVs or unmanned helicopters should be employed instead. For airborne methods, due to the need to deploy larger ring-shaped structures to carry the transmission lines, the system’s weight is relatively heavy. It is generally considered that manned aircraft are the only viable platforms for such setups. If the size and number of turns of the transmitter are reduced, the detection depth will rapidly decrease, thereby limiting the system’s application scenarios. Current reports primarily focus on unexploded ordnance detection [123,124]. To conduct deep-depth detection, the flying platform still needs to possess a strong power supply and carrying capacity. Such UAVs do exist, but their application cost is nearly equivalent to that of manned helicopters. In addition, the technology for directly observing electric fields in the air in a non-contact manner is still under development [125]. With continuous technological advancements, this could drive significant progress in airborne electromagnetic detection technology in the future. Therefore, based on the current technological landscape, there is still no effective solution for using UAVs as platforms for deep-depth fully airborne electromagnetic methods under cost-controlled conditions. Further research is needed to address this challenge.

3.2.5. Seismic Detector

In traditional airborne geophysical surveys, seismic methods are not included, unless there is a disruptive technology. From a theoretical standpoint, non-contact seismic observation is not feasible. However, we have also noted that in recent years, some studies have integrated UAVs with seismic sensors based on node seismometers [126]. The core issue addressed is the rapid deployment and recovery of the node seismometers. Sudarshan et al. [127] reported a design where four geophone sensors (with spikes) are used as the landing gear of an UAV, while the system proposed by Yashin et al. [128] features a data acquisition configuration that is more akin to a traditional node seismometer. Currently, although UAV swarm flight control technology has become relatively advanced, if the goal is to achieve the scale of seismic sensor deployment commonly seen in oil and gas exploration, UAV-based seismic detection methods and field technologies still require further refinement. In addition, it is also very interesting that the seismic acquisition device, which is the most difficult with which to achieve non-contact detection, as shown in Figure 7, appears to be the one with the highest degree of integration with the UAV, which may be due to the emphasis on the detection scale in seismic observation itself.

4. Fusion Method

In the middle and later stages of mineral exploration, the core task is to establish a spatial distribution model of ore-bearing geological bodies based on observational results. For multi-modal observations, it is necessary to have fusion algorithms to build a bridge between observational results (physical parameters) and the three-dimensional distribution model of rocks, so as to provide support for subsequent work such as drilling and resource assessment. This is different from the fusion requirements in the early stage of mineral exploration, which aim to delineate target areas, and the data used are mostly areal data, that is, the data structure is basically two-dimensional (map). In this context, if the issue of information fusion is to be discussed, the main focus should be on the effectiveness of different methods.

However, for information fusion aimed at extracting the spatial distribution of ore bodies, the existing map-oriented fusion techniques actually struggle to directly meet its requirements. Traditional fusion algorithms with key technologies include feature extraction and fusion rules, but for the research scenario in this paper, the data dimensions are different, and the requirements for feature extraction of different modalities will also vary; how to establish cross-dimensional and cross-abstract-level fusion rules is also rarely involved in traditional research. The difference in data dimensions essentially reflects the disparity in the capacity of the data to carry earth information. The differences in information-carrying capacity [32] here mainly result from the distinctions among different geophysical methods in terms of their sensitivity principles to the features of the earth, the resolution capabilities of the observed field quantities, the distribution forms of the observation arrays, and the reliability of the observation systems, which lead to the differences in the performance of the detection results in describing the features of the earth. Therefore, the greatest challenge in achieving cross-dimensional fusion lies in aligning the information-carrying capacity of different observations. In fact, if we examine the fusion problem solely from the perspective of data dimensions, we can clearly see that a large number of studies rely on two-dimensional data structures, and that the research on fusion with a three-dimensional data structure is far less than that with two-dimensional data, while cross-dimensional fusion research is even rarer. Clearly, the demand for three-dimensional and cross-dimensional fusion has long existed, but the research results are scarce. This situation may imply that the increase in data dimensions is a key factor leading to the increased difficulty of fusion.

Based on the above factors, in this paper, we did not adopt the classification method that is commonly used and driven by “methods”, but instead focused on overcoming the dimensional differences. We have reviewed the cases of information fusion for mineral exploration in recent years. We use the data dimension as the classification criterion, and each classification essentially represents an application scenario. Based on this, we discuss the ability of existing fusion methods to solve problems in different scenarios.

4.1. Imaged-Based Fusion

At present, the rise of information fusion technology is generally attributed to research by the U.S. Department of Defense in the late 1980s [129]. However, in the field of remote sensing, the research in and exploration of information fusion technology can be traced back to as early as the late 1970s [130]. Subsequently, with the rapid advancement of computer technology, fusion research based on GIS technology for geological applications developed swiftly [129]. Some fusion algorithms that are still widely used today, such as IHS transform, HPF transform, linear weighting, and PCA transform, also originated during this period. In the late 20th century to the early 21st century, research on the fusion of remote sensing and geochemical data began to emerge [131]. Similarly to geological data, the spatial distribution of elements (concentration levels) on the Earth’s surface forms a pattern akin to remote sensing images, providing a basis for integration with remote sensing imagery. During this period, numerous fusion models were developed, including knowledge-driven models (e.g., fuzzy logic models [132] and evidence theory models [133]), data-driven models (e.g., logistic regression models [134], weights-of-evidence models [135], extended weights-of-evidence models [136], and neural network models [137]), as well as hybrid models combining both approaches (e.g., fuzzy neural network models [138] and fuzzy weights-of-evidence models [139]). The extent to which human experience is involved is the main difference among the above-mentioned different types of fusion. With the development of detection methods and equipment in terms of information-carrying capacity, new methods such as support vector machines (SVMs) [140], random forest [141], Boltzmann machines [142], and decision trees [143] have also been introduced into fusion algorithms to establish deep information-oriented fusion strategies for more complex data.

In the above brief historical review of fusion methods, there were actually not many cases related to geophysics. This is because traditional ground-based geophysical observations have sparse data with complex structures. It was not until the beginning of this century that various kinds of airborne geophysical observation equipment gradually matured and the research on the fusion of geophysical methods gradually increased. Incorporating geophysical information into the fusion process also begins with methods that facilitate the construction of the aforementioned “map”. In earlier research, Choi et al. [132] integrated aeromagnetic, radiometric, geological, and geochemical data using a unique fuzzy logic fusion method to achieve data fusion. Nordebo et al. [144] utilized the Fisher analysis method to conduct fusion studies on electrical resistivity and electromagnetic tomography data, addressing issues related to tunnel engineering. Kashani et al. [145] presented a case study on constructing a mineral potential map using a fuzzy logic fusion method. In this case, the authors collected eleven evidence layers, including seven geophysical layers, three geological layers, and one geochemical layer. Eldosouky et al. [146] conducted an integrated study using Principal Component Analysis, Minimum Noise Fraction transform, and Band Ratio methods on geological, remote sensing, and magnetic data to map lithological units, hydrothermal alteration zones, and structural elements in the study area. The above cases demonstrate that, despite differences in application fields and observation methods, if geophysical results are also presented in the form of two-dimensional feature maps, then within the framework of traditional methods, geophysical results can be integrated with non-geophysical results of the same data structure and achieve good results.

Zhou et al. [147] utilized gamma operators to construct fuzzy rules, integrating a variety of data including geographical, geological, geophysical, geochemical, and dynamic data. This integration was employed to create a hydrocarbon potential map for the Carboniferous period in the survey area. As shown in Figure 8, the tree-like fusion process structure used in reference [147] is quite typical in such fusion algorithms. Firstly, the data is preprocessed, with the core step being the definition of the membership grade of geological variables. To avoid information loss, this article selects the gamma operation to construct the fuzzy fusion rules. Figure 8 shows the feature fusion process for the C1v strata, involving multiple sets of evidence such as the thickness of the source rock, hydrocarbon generation index, and membership grade of the source rock evidence, etc. The characteristics of this case lie in the following: The hierarchical fusion of information is achieved using a tree-like fusion structure. This indicates that researchers have recognized the need to balance evidence with different information-carrying capacities to enhance the fusion effect.

In geophysical observations, in addition to the methods mentioned above that provide results in the form of planar maps, there are also some surveys that typically present profiles along measurement lines, forming characteristic maps on vertical planes. Grandjean et al. [148] proposed a fusion strategy based on fuzzy set theory to study landslide structures, integrating seismic measurements and electrical resistivity imaging results. Li et al. [149] proposed a fusion method based on the Tandem Inverse Filter to enhance the resolution of seismic data with logging data. Dezert et al. [150] proposed an information fusion strategy based on belief functions to integrate geotechnical information with electrical resistivity tomography data, addressing issues related to levee characterization. Dong et al. [151] used a backpropagation neural network optimized via particle swarm optimization to fuse seismic data with transient electromagnetic data, achieving the precise localization of water-bearing collapse columns. Compared with the previous cases of fusion based on planar maps, these cases generally no longer include magnetic, gravity, or other non-geophysical observation results. The observation methods adopted here can penetrate deep into the earth and form a three-dimensional data structure. Observations presented in the form of planar maps have a relatively limited information-carrying capacity in the depth direction. This also implies that one core task for cross-dimensional fusion is to ensure that the information-carrying capacity of different observation modals is roughly consistent in the depth direction.

Moreover, we should also note that, according to the fusion strategy described in the above case, a considerable amount of follow-up work is still required to establish a three-dimensional geological target body model.

4.2. Voxel-Based Fusion

Research on complex geoscience problems urgently requires fusion under three-dimensional spatial conditions. Voxel is essentially a three-dimensional pixel and is the fusion carrier under three-dimensional conditions.

For mineral exploration, Guo & Yan [152] proposed a fusion method based on three-dimensional wavelets. They decomposed gravity and magnetic models using a three-dimensional discrete wavelet transform, separated and fused approximate and detailed components in the wavelet domain, and finally reconstructed the fused model through an inverse discrete wavelet transform. The adoption of wavelet transform-based methods highlights the significance of multi-scale variation information of parameter morphology in three-dimensional space, which should be regarded as a critical factor requiring prioritized consideration.

In the study of the Makran subduction zone, Nasri et al. [153] first performed a 3D inversion of the acquired aeromagnetic and gravity data to obtain 3D models. They then applied wavelet transform rules to fuse the two models, resulting in dimensionless and normalized fusion results (which can be considered a form of “pseudo physical parameter”). This approach revealed more details of the Makran zone structure, the location of the subduction zone, and steep dip faults. In general understanding, the detection results of magnetic fields and gravity fields are mostly presented in the form of two-dimensional maps. Transforming them into three-dimensional data structures through three-dimensional inversion (despite the challenge of multi-solutions) is actually a typical approach, which is helpful for unifying the data dimensions of different observation modes.

In the research on 3D mineral prospectivity modeling, Zheng et al. [154] observed the modal differences among various datasets, including data sources, representation, and information abstraction levels. They proposed a method utilizing correlation analysis regularization to maximize the correlation of features across different models, thereby guiding the network to effectively learn features related to mineralization. Figure 9a presents the fusion network proposed in this paper, which is a typical architecture of this kind of method. This network first employs a coordination module based on convolutional neural networks and a multilayer perceptron to extract features from geological models and simulation data, and uses canonical correlation analysis to balance the differences in the information-carrying capacity among different modalities. The coordinated features are then fused within the joint module to ultimately predict the probability of mineralization. Based on the above fusion strategy, three potential areas (spatial distribution pattern) were delineated in the Jiaojia gold deposit, as shown in Figure 9b. These ore bodies are located at the deep and peripheral parts of the known ore bodies, providing important guidance for future exploration work.

To propose a single model that can accommodate multiple geophysical observations, Erfanian-Norouzzadeh & Fathianpour [155] introduced a three-dimensional fusion algorithm based on the use of a two-dimensional contourlet transform, thereby achieving the rational integration of multiple independent geophysical models.

For the study of the thermo-rheological stratification beneath active volcanic areas, Gola et al. [156] proposed a multi-disciplinary data fusion approach. This method integrates various geological and geophysical data for comprehensive analysis, enabling information analysis and extraction based on a three-dimensional model.

In establishing the mapping relationship between geophysical models and rocks’ physical property distributions, Ramdani et al. [9] employed a generative adversarial network (GAN) to generate visual features of the target rock body based on geophysical observations. This solution is highly insightful and may become an indispensable component of the fusion method framework in the future.

In addition, as we previously discussed, the differences in information-carrying capacity among different observation modes pose significant challenges to fusion research. Friedel [32] recognized this issue in his study on the estimation and scaling of hydrostratigraphic units, where he needed to address the fusion problem between discrete hydrophysical measurements and large-scale 3D airborne electromagnetic survey data. To address this, Friedel introduced a self-organizing map to uncover the nonlinear relationships between multi-modal data with significantly different scales, thereby achieving the integration of the two types of observational data.

From the above review, it can be seen that for the research on three-dimensional fusion methods, how to align the information from different observations is the fundamental issue that determines the technical path.

Furthermore, we also note that in the data preprocessing stage of the aforementioned fusion algorithm, due to the scale differences among different observations, it will inevitably be necessary to align the data space through interpolation. The interpolation process may introduce errors, which could be further amplified in the subsequent fusion. Therefore, whether interpolation should be part of the fusion algorithm or whether the interpolation and fusion functions should be deeply integrated might also be a potential direction for future research.

5. Discussion and Prospect

For mineral resource exploration, UAVs equipped with different geophysical devices are used to obtain multi-modal observation data, and an underground rock mass spatial distribution model is established through information fusion algorithms. Figure 10 describes the important details of the above process: Firstly, an overall design will be carried out based on the detection purpose, and the task will then be decomposed. On this basis, the design and development of the fusion algorithm will be carried out. In fact, although the general development of the fusion algorithm is not related to the specific detection purpose, there may still be targeted development tasks for specific survey areas. Based on the development idea of the fusion algorithm, the multi-modal observation process will be designed to achieve a higher level of matching between the obtained data and the algorithm. After completing the multi-modal observation, the data will be integrated and data processing will be carried out. At the same time, preparations for subsequent fusion need to be made. If the developed fusion algorithm is based on a certain neural network, parameter pre-tuning or pre-training of the network is required. After completing the data and algorithm preparations, information fusion will be carried out and a spatial distribution model of the underground ore-bearing rock will be generated. Finally, based on this model, related geological interpretation work can be carried out. According to the above review, there are still difficulties in completing the above process at present.

5.1. UAV-Borne Observation

Through the above review, we can see that the demand for using UAVs for geophysical observations has always been urgent. The biggest competitive advantage of UAVs will be the possibility of intelligent and large-scale data collection at a low (economic, time) cost. Although this viewpoint has been extensively acknowledged, the comprehension of the so-called large-scale data still, on many occasions, remains merely at the level of the sheer volume of data. In fact, directly fusing data with different information-carrying capacities may, under extreme conditions, lead to one type of observation dominating while others become ineffective. On the other hand, due to the continuation of the traditional observation method based on survey lines, the density of measurement points on and between survey lines varies greatly, which is significantly different from the process of obtaining results based on optical sensors. Since the distribution of measurement points within the survey area is not uniform, this is not very favorable for algorithms aimed at integrating spatial information. Furthermore, gradient observations hold significant importance for enhancing the effectiveness of detection and information fusion. When there are discontinuous interfaces of physical parameters in geological structures, the response process will produce gradient mutations near the interfaces. Therefore, gradient observation will play an important role in enhancing the sensitivity of perception of geological body interfaces. Through UAV formations, multi-dimensional and variable-scale gradient observations can be realized, but reports on this aspect are still very rare. It can be seen that, with the goal of information fusion, there is still considerable room for improvement in the existing UAV-borne multi-modal observation technologies.

At the same time, we also noticed that in most current research cases, the UAVs and the geophysical instruments are still isolated systems from each other; the UAV is still regarded as a mere carrying platform, and there is basically no or even avoided information interaction between the two. This is of course not difficult to understand. But from the perspective of application, if multi-modal observation equipment and algorithms are regarded as the infrastructure of one complete information system [157,158], then integrating UAVs with geophysical instruments and ensuring smooth data and instruction flow between the equipment and computing power (cloud deployment) can maximize the scale effect of this model. And this integration of hardware will also lead to the coordination of detection methods and the unification of data formats among different observation modes, thereby promoting multi-modal geophysical observation and methods to become an independent technical category.

5.2. Fusion Based on AI

The current research hotspot in 3D multi-modal data or information fusion is not only limited to the field of geosciences. In daily life, what ordinary people have more contact with is the relevant research and technology regarding self-driving vehicles. The optical and radar sensors outfitted on vehicles can generate planar and spatial data (point clouds) for different fields of view and detection distances. Based on computer vision (CV) technology, the dynamic semantic extraction of basic road conditions, traffic prompt information, and traffic participants is accomplished, and correct decisions are made through information fusion. The capacity of an on-board system to realize autonomous driving is one of the crucial criteria for differentiating a vehicle as a traditional transportation tool or a robot, which holds significant reference significance for the intelligent geophysical exploration system constructed based on UAVs. Employing UAV-borne geophysical instruments to obtain large-scale detection data and form a private big data space at the survey area level establishes the foundation for the subsequent implementation of multi-modal information fusion. Simultaneously, we should recognize that the capacity for data collection of the existing equipment significantly surpasses the ability to extract and interpret information from the collected data [24]. Although existing fusion methods can better characterize Earth’s features compared to traditional single geophysical methods, they still do not fully harness the potential of multi-modal observations.

The difficulty in realizing the above-mentioned scheme lies in resolving the fusion difficulties brought about by the complex disparities among multi-modal geophysical and non-geophysical observations in aspects such as target sensitivity, spatial resolution, feature resolution, data structure, and abstraction level. For instance, suppose that the gravity and magnetic field observations (two-dimensional data structure) and electromagnetic observations (three-dimensional data structure) need to be fused. In the existing fusion strategies, the above-mentioned data are usually inverted to construct three-dimensional models of their respective physical quantities, achieving alignment in the three-dimensional space. However, the challenge lies in that the non-uniqueness problem of the inversion of gravity and magnetic field data is more severe compared to the electromagnetic method. The cost of this cross-data dimension fusion solution is the loss of objectivity to a certain extent. For another instance, suppose we have successfully integrated various geophysical observations and obtained a single geophysical parameter (or pseudo physical parameter) model. In the subsequent fusion process with non-geophysical information such as geological, remote sensing, geochemical, or borehole data, the abstraction levels of different models of information vary, meaning that the knowledge logic they represent is different.

The introduction of artificial intelligence (AI) approaches (multi-modal machine learning, MMML) appears to be a solution trajectory. Particularly with the reinforcement of the attention mechanism [159], the abilities to discriminate the significance of features and establish feature relationships have been notably advanced. Consequently, this has brought about the vigorous growth of research such as Vision Language (VL) bimodal models [160,161]. Under the MMML architecture, there are multiple challenges in the fusion of geophysical and geological data, including representation, alignment, co-learning, and generation [162]. The multi-modal representation primarily aims to excavate the complementarity and redundancy among different modal data, organically integrate various pieces of heterogeneous information, and realize information complementarity and redundancy filtering between modals. In this process, the attention mechanism (cross-attention) is introduced to model the cross-modal relationship by determining the “participation degree” of different modal information. Multi-modal alignment is used to construct the corresponding relationships between different pieces of modal information, which could be difficult in the fusion of multi-modal geophysical information, as in many cases, the alignment relationship of different observations based on parameters (values and spatial positions) is ambiguous. At this point, scenarios similar to “Text-Guided Visual Tasks” can be considered, introducing descriptive texts of the observed regions into multi-modal representation and alignment, that is, enhancing the overall performance of the system through co-learning. The ultimate objective of carrying out multi-modal fusion is to achieve the generation of the spatial distribution model of rocks. If the “rock distribution model” to be generated is also regarded as a modal (target modal), then the process from the observation modals to the target modal is actually a cross-modal generation process, which necessitates the introduction of relevant construction techniques for generative networks. Additionally, distinct from NLP or CV, constrained by the complexity of the geoscience data acquisition process and the non-disclosure of a substantial amount of measured data, the more modals that need to be taken into account, the greater the difficulty of constructing large-scale multi-modal datasets for supervised training. This represents a challenge for a large number of general research institutions, and even the employment of distillation techniques may not essentially address the issue. Hence, whether to adhere to the universality of the model and which type of fusion model architecture to select will be the primary considerations when constructing the system. In reality, the problem of datasets is also one of the difficulties that constrains the extensive application of AI methods in the geoscience domain. Thus, it can be seen that obtaining data through UAV-borne geophysical instruments can at least allow more researchers to construct relatively comprehensive multi-modal datasets within the limited areas of interest in a more flexible manner. As the number of such multi-modal fusion studies targeting relatively specific scenarios grows, more generalized technical approaches will gradually surface. Standardized open-source multi-modal datasets tailored to the fine exploration requirements of minerals or underground engineering will gradually form, and they will exert a significant role in deepening the research on informatization and intelligence in geoscience.

6. Conclusions

With the rapid development of drone technology, an increasing number of pieces of geophysical exploration equipment are choosing to use UAVs as their deployment platform. Due to the significant reduction in costs and substantial increase in efficiency, it has become possible to create a private local big data space within a limited survey area. Enhancing the reliability of comprehensive exploration through multi-modal observations is a natural extension of the aforementioned trend. However, the current multi-modal observation and processing procedures are mainly simple superimpositions of single modalities. On the one hand, the integration degree of UAVs and geophysical equipment is still low, and the advantages of UAVs as flight robots have not been fully exploited; on the other hand, the existing fusion methods are still difficult to use to establish the spatial distribution model of underground ore-bearing rocks. To explore the above problems, we systematically reviewed the current development status of UAVs and geophysical instruments suitable for UAVs. We believe that only by integrating the system, designing the observation plan in accordance with the requirements of the fusion method, and treating the hardware part as an external extension of the algorithm, can high-matching data be provided for information fusion. Subsequently, we analyzed the latest progress of existing fusion methods, which led us to believe that cross-dimensional and cross-abstract-level problems are major challenges in the algorithm aspect. At the same time, the fusion process should be carried out simultaneously with the generation of the spatial distribution model of ore-bearing rocks, that is, to establish an integrated system of fusion and generation. Improving the observation and integration logic framework will not only facilitate the development of refined mineral exploration in the future, but also play a significant role in areas such as the construction of underground works and monitoring of geological disasters in mines. Through the above review and prospect, this study aims to further promote the research in and application of systematic AI methods in the field of geophysical exploration.