Intelligent exploration system based on unmanned aerial vehicle multispectral remote sensing and AI mineral identification

Abstract

The invention discloses an intelligent exploration system based on unmanned aerial vehicle multispectral remote sensing and AI mineral recognition, and relates to the technical field of geological exploration. A six-rotor folding platform is adopted, a magneto-rheological damping mounting and multi-load interface is matched, a differential RTK-GPS and a millimeter wave obstacle avoidance radar are integrated, a 20-waveband camera, a thermal infrared camera and a laser radar are included, waveband combination is dynamically adjusted, and multi-sensor data space-time alignment is achieved; the time series data preprocessing module executes three-stage correction, fuses time series data and filters noise; a multi-modal fusion framework is adopted, and a transfer learning library is built in; a three-dimensional model is constructed, and mining simulation and multi-dimensional analysis are supported; and real-time mapping, data transmission and intelligent task planning are realized. According to the invention, exploration efficiency and identification precision are improved, complex environment adaptability is enhanced, data security is guaranteed, ecological protection is considered, seamless connection of exploration and mining is realized, and cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to an intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification. Background Technology

[0002] In the field of geological exploration, the efficiency and accuracy of mineral resource exploration directly determine the cost and benefits of resource development. Traditional mineral exploration mainly relies on manual field surveys. Exploration personnel must carry tools such as geological hammers, compasses, and spectrometers to venture into the field, observing rock outcrops, collecting samples, and conducting laboratory analysis to identify minerals and determine their distribution. This method is severely limited by terrain conditions. In complex terrains such as mountains, canyons, and deserts, not only is personnel movement difficult, but there are also safety risks such as falls and getting lost. Moreover, the exploration cycle for a single area often lasts for weeks or even months, resulting in extremely low efficiency. At the same time, manual identification is highly dependent on the experience of exploration personnel. For minerals with similar compositions (such as hematite and limonite, calcite and dolomite), misjudgment is prone to occur, leading to deviations in subsequent reserve estimates and increasing the uncertainty of resource development. In addition, manual exploration requires a large investment of manpower, including exploration personnel salaries, sample transportation, and laboratory testing costs. For large-scale exploration areas, the cost pressure is significant, making it difficult to meet the needs of large-scale, rapid exploration.

[0003] With the development of remote sensing technology, satellite remote sensing and early UAV remote sensing have been gradually applied to geological exploration. While satellite remote sensing can cover large areas, its spatial resolution is limited by satellite orbit and resolution, typically ranging from a few meters to tens of meters. This makes it difficult to accurately capture the detailed features of small-scale mineral outcrops, and data acquisition is greatly affected by weather and cloud cover, resulting in poor timeliness and failing to meet the needs of urgent exploration missions. Early UAV remote sensing mostly used single-spectral or limited-band cameras, acquiring only surface morphology or limited spectral information, failing to comprehensively reflect the spectral characteristics of minerals. Furthermore, it lacked multi-sensor collaboration capabilities; data from lidar, thermal infrared, and other sources often exhibited spatiotemporal misalignment with multispectral data, leading to low data utilization. Some systems attempted to attach multiple external sensors, but the impact of UAV payload, endurance, and vibration resulted in poor sensor stability and difficulty in guaranteeing data acquisition accuracy. Especially in harsh environments such as high altitudes and low temperatures, equipment failure rates increased significantly, further limiting the application of the technology.

[0004] In recent years, AI technology has begun to be integrated into mineral identification, but existing solutions still have significant shortcomings. Most AI models are trained based on single-spectral data, failing to incorporate topographical information from lidar and temperature characteristics from thermal infrared radiation. This results in poor generalization ability for mineral identification in complex geological contexts, and a significant drop in accuracy is likely in new geological areas. Furthermore, the models lack self-iterative update mechanisms. As exploration areas expand and mineral types increase, a large number of samples need to be manually relabeled and fully trained, which is time-consuming and labor-intensive. In addition, exploration data contains sensitive geographical information and mineral distribution privacy; existing systems lack effective data anonymization methods, which can easily lead to information leaks and disputes over rights in exploration areas. On the other hand, traditional exploration processes often neglect ecological protection; exploration activities may damage vegetation and disturb sensitive ecological areas, which is inconsistent with the current development concept of green exploration. How to improve exploration efficiency while reducing ecological impact has become an urgent problem for the industry. Summary of the Invention

[0005] The present invention proposes an intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification, comprising: The UAV adaptive mounting module adopts a six-rotor folding platform, equipped with magnetorheological shock absorption mounts, and has a compatible interface to support the simultaneous mounting of multispectral cameras, LiDAR, and thermal infrared cameras; it also integrates millimeter-wave obstacle avoidance radar and terrain following function. Dynamic multispectral acquisition module: includes a 20-band tuned multispectral camera, a thermal infrared camera and a pulsed lidar; equipped with a spectral feature sensing unit, it dynamically adjusts the camera band combination according to the target mineral's preset spectral library, which contains 32 common mineral feature bands; the module achieves spatiotemporal alignment of data from each sensor through a 1PPS synchronization clock. The time-series data preprocessing module performs three levels of correction on the raw data: radiometric correction, atmospheric correction, and geometric correction. Radiometric correction is based on real-time monitoring of dark current. Atmospheric correction integrates the MODTRAN model with real-time meteorological data to eliminate the dynamic effects of water vapor and aerosols. Geometric correction combines lidar point clouds and ground control points. A new time-series data fusion function has been added, which uses weighted average fusion of multispectral data from 3-5 time periods in the same area. The weights are allocated according to the data signal-to-noise ratio, and the final output is a standardized time-series multispectral dataset. Cross-modal AI mineral recognition module: It adopts a multispectral + LiDAR + thermal infrared cross-modal fusion architecture; the basic network is ResNet-101, the input is a 512×512 pixel image, and a cross-modal attention mechanism is added to the network to dynamically assign weights to multispectral reflectance, LiDAR elevation, and thermal infrared temperature features; the module has a built-in mineral feature transfer learning library to support incremental learning of new mineral types. The exploration-mining linkage visualization module constructs a 3D terrain model based on LiDAR point clouds, overlays mineral identification results onto the 3D terrain model, generates a dynamic heat map, and supports real-time updates. The module adds a mining simulation function. After inputting mining process parameters, it can automatically generate the mining range and reserve estimation. The reserve estimation uses Kriging interpolation. The module supports horizontal slicing, vertical profile, and reserve trend prediction analysis. Digital Twin Control and Feedback Module: Builds a digital twin model of the UAV and exploration area, mapping the UAV's location, sensor status, and mineral distribution in real time; the module achieves real-time data transmission and command issuance through 5G / satellite dual-mode transmission; the module is equipped with task replanning and health monitoring functions, and also realizes the automatic generation of exploration reports, which include mineral distribution, reserve estimation, and mining recommendations.

[0007] Furthermore, it also includes: Mineral Feature Enhancement Module: Enhances mineral differentiation by fusing the integral features of multispectral reflectance and thermal infrared radiation values. The calculation formula is as follows: Where G is the enhanced multimodal eigenvalue; λ1 and λ2 are the characteristic band intervals of the target mineral; R(λ) is the multispectral reflectance at wavelength λ; T(λ) is the normalized value of thermal infrared radiation temperature at wavelength λ, which is obtained by converting the measured temperature from the thermal infrared camera; z1 and z2 are the reference depths from the surface to the subsurface in the exploration area; and μ(λ,z) ​​is the spectral attenuation coefficient at wavelength λ and depth z, which is obtained by querying the geological database.

[0008] Dynamic trajectory optimization module: Adjusts the flight path based on real-time mineral identification results. The optimized trajectory curvature is calculated as follows: Where κ(t) is the trajectory curvature at time t, the smaller the curvature, the smoother the trajectory; x(t) and y(t) are the planar coordinates of the UAV at time t, dx(t) / dt and dy(t) / dt are the velocities of the UAV in the x and y directions at time t, and d 2 x(t) / dt 2 d 2 y(t) / dt 2 Let t be the accelerations of the UAV in the x and y directions.

[0009] Furthermore, the cross-modal AI mineral identification module also includes a mineral purity inversion unit. First, a mineral sample with known purity is collected, and the actual purity of the sample is determined using an X-ray fluorescence spectrometer. Then, the average multispectral reflectance of the sample in the characteristic band is extracted.

[0010] Furthermore, the UAV adaptive mounting module also includes a harsh environment protection unit, a waterproof camera compartment with a built-in positive pressure dust removal system that uses a HEPA filter to filter dust, a 20W constant temperature heating element in the battery compartment that automatically activates when the temperature inside the battery compartment is below 0°C, and a carbon fiber-glass fiber composite structure for the propeller.

[0011] Furthermore, the time-series data preprocessing module also includes an exploration data desensitization unit, which uses a combination of differential privacy and k-anonymity techniques to achieve data desensitization. The specific process is as follows: Gaussian noise is added to the coordinate data, with a mean of 0. The noise is dynamically adjusted according to the regional sensitivity to meet the privacy protection budget ε=1.5. The desensitized data maintains the relative positional relationship and trend characteristics of mineral distribution, and can be safely used for multi-team collaborative analysis, while avoiding the risk to exploration area rights caused by the leakage of geographic information.

[0012] Furthermore, the cross-modal AI mineral recognition module also includes a model self-iterative update unit, which updates the model every 100km. 2 The exploration task automatically triggers the sample update process, which selects new samples from the identification results. The new samples include high-confidence samples and low-confidence verification samples. The identification probability of high-confidence samples is not less than 90% and the spectral similarity with surrounding samples is less than 3%, and the number is not less than 1,000. The identification probability of low-confidence verification samples is between 40% and 60%, and the number needs to be manually verified on-site.

[0013] Furthermore, the exploration-mining linkage visualization module also includes a three-dimensional fracture analysis unit. First, the lidar point cloud data is preprocessed, including noise point removal and downsampling. Noise point removal uses statistical filtering. Then, an improved Hough transform is used to detect the linear characteristics of fractures. Next, the fracture density is calculated using a 10m×10m sliding window. The fracture density is the ratio of the sum of all fracture lengths within the window to the window area. Then, the correlation between fracture density and mineral abundance is analyzed using the Pearson correlation coefficient. When the absolute value of the correlation coefficient is not less than 0.7, it is considered a strong correlation. Finally, the potential exploration area is marked in red in the three-dimensional model.

[0014] Furthermore, the digital twin control and feedback module also includes an energy consumption optimization unit to construct an energy consumption prediction model. The input parameters of the model include terrain complexity, flight speed, and sensor operating mode. Terrain complexity is divided into three categories: flat, gentle slope, and steep slope. Based on the energy consumption prediction model, an algorithm is used to automatically plan the optimal path, dynamically adjust the sensor operating mode, and monitor the remaining battery power in real time.

[0015] Furthermore, it also includes: The exploration ecological impact assessment module first calculates the Normalized Difference Vegetation Index (NDVI), calculated as NDVI = (NIR - R) / (NIR + R), where NIR is the near-infrared reflectance and R is the red reflectance. Areas with an NDVI exceeding 0.6 are classified as high vegetation cover areas. Then, sensitive ecological zones are identified using multispectral data, including wetlands and rare plant areas. Next, an ecological impact index E is constructed: E = 0.4 × (1 - vegetation cover) + 0.3 × sensitive area distance factor + 0.3 × terrain damage risk. Vegetation cover is the percentage of areas with an NDVI exceeding 0.3, the sensitive area distance factor is 1 - distance / 1000 and is effective when the distance is less than 1000m, and the terrain damage risk is the product of the percentage of areas with a slope exceeding 25° and 0.5. Finally, ecological protection recommendations are generated, suggesting avoiding high-impact areas with an E exceeding 0.6.

[0016] Compared with existing technologies, the beneficial effects of this invention are: In terms of exploration efficiency, this invention completely changes the inefficient mode of traditional manual exploration. The UAV adaptive mounting module supports the simultaneous operation of multiple sensors, enabling rapid coverage of large exploration areas. Combined with dynamic trajectory optimization, it can accurately collect data in high-potential mineral areas, avoiding ineffective exploration and significantly shortening the exploration cycle. The digital twin control and feedback module enables real-time data transmission, intelligent task planning, and automatic report generation, reducing manual intervention. Compared with traditional manual exploration, the overall efficiency is significantly improved, making it particularly suitable for large-scale, urgent exploration tasks.

[0017] In terms of recognition accuracy, the cross-modal AI mineral recognition module comprehensively captures the spectral, topographic, and temperature characteristics of minerals by fusing multispectral, lidar, and thermal infrared data, significantly improving recognition accuracy compared to single-spectral recognition. The multimodal mineral feature enhancement module further amplifies the feature differences between similar minerals, and the model's self-iterative update mechanism ensures high recognition accuracy across different geological regions. The mineral purity inversion unit provides accurate data for reserve estimation, effectively avoiding the misjudgment problems of traditional manual recognition and single-model recognition, and improving the reliability of exploration results.

[0018] In terms of environmental adaptability, the harsh environment protection unit of the UAV's adaptive module, through its waterproof, dustproof, temperature-controlled, and wind-resistant design, enables the system to operate stably in complex environments such as sandstorms, low temperatures, and gusts of wind, breaking through the environmental limitations of traditional equipment and expanding the exploration range. The energy consumption optimization unit extends the UAV's endurance and improves the coverage capability of a single mission by dynamically adjusting flight strategies and sensor modes, making it particularly suitable for remote and resupply-difficult field areas.

[0019] In terms of data security and ecological protection, the exploration data anonymization unit employs differential privacy and k-anonymity technologies to protect sensitive information while preserving data validity, thus avoiding disputes arising from information leaks. The exploration ecological impact assessment module generates scientific ecological protection recommendations through vegetation index calculation and sensitive ecological zone identification, reducing the interference of exploration activities on vegetation and sensitive ecological zones. This achieves synergy between exploration and ecological protection, aligning with the concept of green exploration and balancing resource development with sustainable ecological development.

[0020] In addition, the exploration-mining linkage visualization module combines mineral identification results with 3D terrain and mining technology to generate mining scope and reserve estimates. It supports multi-dimensional analysis and BIM format output, providing a direct basis for subsequent mining planning, achieving seamless connection between exploration and mining, reducing intermediate links, further reducing resource development costs, and improving overall economic benefits. Attached Figure Description

[0021] Figure 1 This is a schematic block diagram of the intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification proposed in this invention; Figure 2 This is a schematic diagram comparing the exploration cycles of traditional manual exploration and the intelligent exploration system based on UAV multispectral remote sensing and AI mineral recognition under different exploration area sizes. Figure 3 This diagram illustrates a comparison of the accuracy of single-spectral recognition for the same mineral type and cross-modal AI recognition within the intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification proposed in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0024] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0025] Reference Figures 1 to 3 This invention discloses an intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification, applicable to efficient field exploration of metallic, non-metallic, and energy minerals. Metallic minerals include iron, copper, and gold; non-metallic minerals include limestone and graphite; and energy minerals include coal seams and shale gas. This system addresses the problems of low efficiency, high cost, poor terrain adaptability, and reliance on experience for mineral identification inherent in traditional manual exploration methods. It includes the following modules: The UAV adaptive payload module utilizes a hexacopter folding platform with a maximum takeoff weight of 15kg and an endurance of 60-80 minutes. This module is equipped with a magnetorheological damping mount, capable of supporting a minimum load of 8kg, a vibration attenuation rate of at least 95%, and a response time of no more than 20ms. It also features a multi-payload compatible interface, supporting simultaneous mounting of multispectral cameras, LiDAR, and thermal infrared cameras. The module integrates differential RTK-GPS with a positioning accuracy of ±0.3m and a sampling frequency of 15Hz. It also integrates millimeter-wave obstacle avoidance radar with a detection range of 80m, capable of identifying obstacles with a diameter of at least 0.5m. The module features terrain following capability, with an adaptive altitude adjustment range of 50-200m. When the slope of the exploration area exceeds 30°, it automatically increases its altitude by 10m. It can operate stably in environments ranging from -20℃ to 50℃, with a data acquisition interruption rate of no more than 3%.

[0026] Dynamic multispectral acquisition module: Includes a 20-band tunable multispectral camera, a high-resolution thermal infrared camera, and a pulsed lidar. The 20-band tunable multispectral camera has a spectral range of 350-1100nm, with a bandwidth adjustable between 5-15nm, and a spatial resolution of 0.3m at a height of 100m; the high-resolution thermal infrared camera has a resolution of 1280×1024, a temperature measurement range of -30℃ to 200℃, and a temperature measurement accuracy of ±1℃; the pulsed lidar has a point cloud density of 100 points / m². 2 The ranging accuracy is ±3cm. This module is equipped with a spectral feature sensing unit, which can dynamically adjust the camera band combination according to a preset spectral library of target minerals. The preset spectral library contains 32 common mineral characteristic bands; for example, when detecting hematite, it will focus on the 450-550nm and 850-950nm bands. The module achieves spatiotemporal alignment of multi-sensor data through a 1PPS synchronization clock, with a synchronization error of no more than 500ns. Data storage uses a 2TB NVMe SSD, with a read / write speed of no less than 1GB / s.

[0027] The time-series data preprocessing module performs three levels of correction on the raw data: radiometric correction, atmospheric correction, and geometric correction. Radiometric correction is based on real-time dark current monitoring, with a correction accuracy of no more than 1.5%. Atmospheric correction integrates the MODTRAN model with real-time meteorological data to eliminate the dynamic effects of water vapor and aerosols. Geometric correction combines lidar point clouds and ground control points, with a planar error of no more than 0.8m. This module adds a time-series data fusion function, using weighted average fusion of multispectral data from 3-5 time periods in the same area. The weights are allocated according to the data signal-to-noise ratio (SNR), and the weight is no less than 0.6 when the SNR exceeds 40dB. Noise filtering uses adaptive median filtering, with the filtering window dynamically adjustable within the range of 2×2-5×5, improving the SNR by no less than 20dB. The final output is a standardized time-series multispectral dataset.

[0028] The cross-modal AI mineral recognition module employs a multispectral + LiDAR + thermal infrared cross-modal fusion architecture. Its base network is an improved ResNet-101, taking a 512×512 pixel multimodal image as input. A cross-modal attention mechanism is incorporated into the network, dynamically assigning weights to multispectral reflectance, LiDAR elevation, and thermal infrared temperature features, with a weight error not exceeding 3%. The module includes a built-in mineral feature transfer learning library containing 1 million labeled samples, supporting incremental learning for new mineral types, with a model update time of no more than 1.5 hours. For low-confidence areas with a recognition probability below 60%, the module initiates semi-supervised learning, ensuring a utilization rate of at least 40% for unlabeled samples. Simultaneously, it optimizes recognition results by incorporating mineral thermal radiation characteristics; for example, sulfide minerals are 2-5℃ warmer than the surrounding rock. The overall recognition accuracy of the module is no less than 94%.

[0029] The exploration-mining linkage visualization module constructs a high-precision 3D terrain model based on LiDAR point clouds, with an elevation accuracy of ±0.05m. This module overlays mineral identification results onto the 3D terrain model to generate a dynamic heatmap with a color gradation interval of 3% and supports real-time updates. The module adds a mining simulation function, allowing input of mining process parameters, such as an open-pit bench height of 10-15m and an underground tunnel spacing of 20-30m. After inputting these parameters, the module automatically generates the mining area and reserve estimates. The reserve estimate uses Kriging interpolation with an error not exceeding 6%. The module supports multi-dimensional analysis, including horizontal slicing, vertical profiles, and reserve trend prediction. Analysis results can be exported in both GIS and BIM formats for convenient subsequent mining planning.

[0030] Digital Twin Control and Feedback Module: Centered on an industrial-grade touchscreen terminal with a 10-inch screen and a minimum battery life of 10 hours, this module constructs a digital twin model of the drone and the exploration area, enabling real-time mapping of drone location, sensor status, and mineral distribution. The module achieves real-time data transmission and command issuance via 5G / satellite dual-mode transmission, with a transmission rate of at least 20Mbps and satellite latency of no more than 500ms in remote areas. The module features intelligent task replanning, automatically detouring around no-fly zones with a detour path deviation of no more than 5 meters. It also includes equipment health monitoring, monitoring parameters such as remaining battery power and sensor temperature, with an anomaly warning response time of no more than 2 seconds. Furthermore, it features automatic exploration report generation, producing reports including mineral distribution, reserve estimates, and mining recommendations, with a report generation time of no more than 10 minutes.

[0031] This invention also includes: Multimodal mineral feature enhancement module: This module enhances mineral distinguishability by fusing the integral features of multispectral reflectance and thermal infrared radiation values. The calculation formula is as follows: Where G is the enhanced multimodal eigenvalue; λ1 and λ2 are the characteristic wavelength ranges of the target mineral, in nm, for example, the characteristic wavelength range of copper ore is 500-600 nm; R(λ) is the multispectral reflectance at wavelength λ, ranging from 0 to 1; T(λ) is the normalized value of thermal infrared radiation temperature at wavelength λ, ranging from 0 to 1, calculated from the measured temperature of the thermal infrared camera; z1 and z2 are the reference depths from the surface to the subsurface of the exploration area, in m, usually 0-5 m; μ(λ,z) ​​is the spectral attenuation coefficient at wavelength λ and depth z, in m⁻ 1 The data can be obtained from geological databases. This calculation can amplify the differences in characteristics between similar minerals such as calcite and dolomite, improving the AI's recognition accuracy by at least 12%.

[0032] Dynamic trajectory optimization module: This module adjusts the flight path based on real-time mineral identification results. The optimized trajectory curvature is calculated using the following formula: Where κ(t) is the track curvature at time t, in m⁻ 1 The smaller the curvature, the smoother the flight path; x(t) and y(t) are the planar coordinates of the UAV at time t, in meters; dx(t) / dt and dy(t) / dt are the velocities of the UAV in the x and y directions at time t, in meters per second; d 2 x(t) / dt 2 d 2 y(t) / dt 2 Let t be the acceleration of the UAV in the x and y directions, in m / s². 2When an area with a mineral probability of at least 80% is identified, the module reduces the track curvature to ensure that κ(t) does not exceed 0.005m⁻ 1 At the same time, the flight strip spacing was reduced from 50m to 30m to increase the data acquisition density in the area, while avoiding equipment vibration caused by sudden changes in flight path, and the data acquisition accuracy was improved by no less than 15%.

[0033] In this invention, the cross-modal AI mineral identification module further includes a mineral purity inversion unit. The specific implementation process of this unit is as follows: First, mineral samples of known purity are collected, with purity levels covering 30%-100% at 10% intervals. The number of samples collected for each purity level is no less than 50, and the weight of a single sample is no less than 100g. The actual purity of the samples is determined using an X-ray fluorescence spectrometer with a detection accuracy of 0.1%. Next, the average multispectral reflectance of the samples in characteristic bands is extracted. For example, the characteristic band for gold ore is 650-750nm. Based on the extracted average reflectance and the measured purity, a reflectance-purity calibration curve is established. The calibration curve is fitted using a quadratic polynomial, with a goodness of fit R0. 2 The initial purity is not less than 0.95. Then, for unknown areas, the reflectance of the same characteristic bands is extracted and substituted into the calibration curve to calculate the initial purity. Then, the texture features of the lidar point cloud are combined for correction. Specifically, the elevation standard deviation and slope within a 5×5 window are calculated. The elevation standard deviation is roughness, in meters (m), and the slope is in degrees (°). When the roughness exceeds 0.5m or the slope exceeds 20°, a correction coefficient k is introduced. The formula for k is k = 1 - 0.02 × (roughness + slope / 10). The value range of k is 0.8-1.0. The final purity is the product of the initial purity and k. Through this process, the mineral purity inversion accuracy is kept below 5%, providing accurate mineral content data for reserve estimation.

[0034] In this invention, the UAV adaptive mounting module also includes a harsh environment protection unit, the specific design of which is as follows: the camera compartment adopts an IP67 waterproof design, capable of being immersed in 1m of water for 30 minutes without leakage; the camera compartment has a built-in positive pressure dust removal system, with an operating frequency of 5Hz and a flow rate of 2L / min, maintaining the air pressure inside the compartment at 105-110kPa. The system achieves dust filtration through a HEPA filter, with a filtration efficiency of not less than 98%, and can withstand PM10 concentrations exceeding 500μg / m³. 3The system keeps the lens clean in dusty environments; the battery compartment is equipped with a 20W constant temperature heating element, which automatically activates when the temperature inside the battery compartment drops below 0℃, with a heating rate of 5℃ / min, maintaining the temperature inside the compartment between 5-25℃. In an environment of -10℃, this can increase battery life by at least 20%; the propeller adopts a carbon fiber-glass fiber composite structure with a thickness of 2mm and a tensile strength of at least 300MPa. Wind tunnel testing has verified that it can withstand gusts of 12m / s. When the wind speed fluctuates within ±3m / s, the drone's flight attitude deviation does not exceed 5°. These designs ensure that the completion rate of exploration missions in complex environments is no less than 90%.

[0035] In this invention, the time-series data preprocessing module also includes an exploration data desensitization unit. This unit employs a combination of differential privacy and k-anonymity techniques to achieve data desensitization. Specifically, Gaussian noise is added to the coordinate data, with a mean of 0 and a standard deviation between 3 and 5 meters. This noise is dynamically adjusted based on regional sensitivity to meet a privacy protection budget of ε=1.5, ensuring that the probability of reconstructing the true coordinates is less than 1%. K-anonymization is applied to the ground control point information, with a k value of at least 5, meaning each obfuscated location corresponds to at least 5 real control points. After processing, only the regional description is retained, such as "Eastern Section of XX Mountain Range," without displaying specific latitude and longitude. Spatial aggregation is performed on the mineral distribution data, merging data into a 100m × 100m grid, with the average mineral content error within the grid not exceeding 8%. The desensitized data still maintains the relative positional relationships and trend characteristics of mineral distribution, making it safe for multi-team collaborative analysis while avoiding the risk to exploration area rights due to geographic information leakage.

[0036] In this invention, the cross-modal AI mineral recognition module further includes a model self-iterative update unit, the workflow of which is as follows: The system completes 100km... 2The exploration task automatically triggers a sample update process, selecting new samples from the identification results. New samples include high-confidence samples and low-confidence validation samples. High-confidence samples have an identification probability of no less than 90% and a spectral similarity of less than 3% with surrounding samples, and their number is no less than 1000. Low-confidence validation samples have an identification probability between 40% and 60%, their number is no less than 500, and they require manual field verification. Based on the new samples, the original model is fine-tuned using the Adam optimizer with a learning rate of 1e-5, momentum of 0.9, batch size of 32, and 50 iterations. After fine-tuning, the model's accuracy on the validation set improves by at least 2%. If manual verification reveals that the model's recognition error exceeds 8%, meaning that more than 40 out of 500 validation samples are misclassified, then full training is initiated. Full training is based on an updated database of over 1 million samples, with a training set to validation set ratio of 8:2. The training employs a cosine annealing learning rate strategy with an initial learning rate of 1e-4 and a training time not exceeding 4 hours. Through this iterative update process, the model's recognition accuracy in different geological regions is ensured to remain at least 92%, avoiding performance degradation due to geological differences.

[0037] In this invention, the exploration-mining linkage visualization module further includes a three-dimensional fracture analysis unit. The working steps of this unit are as follows: First, the lidar point cloud data is preprocessed, including noise point removal and downsampling. Noise point removal uses statistical filtering with a distance threshold of 1.5 times the standard deviation. Downsampling retains key feature points, reducing the point cloud density to 50 points / m. 2 Then, an improved Hough transform was used to detect the linear characteristics of the fractures. The accumulator threshold of the improved Hough transform was no less than 20, the angle step size was 0.5°, the minimum length of the linear segment was no less than 0.5m, and the gap tolerance was no more than 0.2m. Next, the fracture density was calculated using a 10m×10m sliding window. The fracture density was the ratio of the sum of all fracture lengths within the window to the window area, in m / m. 2 Then, the correlation between fracture density and mineral abundance was analyzed using the Pearson correlation coefficient. A strong correlation was defined as an absolute value of the correlation coefficient not less than 0.7. Finally, high-potential exploration areas were highlighted in red in the 3D model. High-potential areas must meet the requirement of a fracture density exceeding 5 m / m². 2 Furthermore, the mineral abundance exceeds 60%, and combined with mining process parameters, such as a borehole diameter of 150mm, the optimal borehole location is automatically generated. The optimal borehole location is no more than 5m away from the center of the high-potential area and avoids the intersection of large fractures. Through these steps, the number of invalid boreholes is reduced by no less than 25%.

[0038] In this invention, the digital twin control and feedback module further includes an energy consumption optimization unit, which is implemented as follows: An energy consumption prediction model is constructed. The model's input parameters include terrain complexity, flight speed, and sensor operating mode. Terrain complexity is divided into three categories: flat, gentle slope, and steep slope, with corresponding coefficients of 1.0, 1.3, and 1.6, respectively. The flight speed range is 5-10 m / s, with energy consumption increasing by 5% for every 1 m / s increase in speed. The sensor operating mode is divided into full-band and simplified band, with an energy consumption ratio of 1:0.7 between the two modes. The energy consumption is then adjusted according to the size of the exploration area, for example, 50 km. 2 Based on an energy consumption prediction model, an improved A* algorithm is used to automatically plan the optimal path, shortening the total path length by at least 10%. The sensor's operating mode is dynamically adjusted; for non-target areas with a mineral probability below 30%, the sensor is switched to a simplified band mode, retaining 8 key bands, which reduces energy consumption by 30%. The remaining battery power is monitored in real time at a sampling frequency of 1Hz. When the battery level drops below 20%, data acquisition is prioritized for identified high-potential areas with a mineral probability of at least 70%, automatically abandoning low-potential areas to ensure a key data integrity rate of at least 95%. With the same battery capacity, this energy consumption optimization unit can increase the exploration area by at least 20%, extending the coverage of a single mission.

[0039] This invention also includes: Exploration Ecological Impact Assessment Module: The working process of this module is as follows: First, calculate NDVI (Normalized Difference Vegetation Index). The formula for NDVI is NDVI = (NIR - R) / (NIR + R), where NIR is the near-infrared reflectance (800-900nm) and R is the red reflectance (600-700nm). Areas with an NDVI exceeding 0.6 are identified as high vegetation cover areas. Then, sensitive ecological areas are identified through multispectral data. Sensitive ecological areas include wetlands and rare plant areas. Wetlands are identified based on a 560nm reflectance exceeding 20% ​​and an 860nm reflectance below 15%. Rare plant areas are identified by matching with a preset spectral library, with a matching degree of not less than 85%. Next, construct... An ecological impact index E is constructed, where E = 0.4 × (1 - vegetation cover) + 0.3 × sensitive area distance factor + 0.3 × terrain damage risk. Vegetation cover is the percentage of areas with an NDVI exceeding 0.3. The sensitive area distance factor is 1 - distance / 1000 and is effective when the distance is less than 1000m. The terrain damage risk is the product of the percentage of areas with a slope exceeding 25° and 0.5. Finally, ecological protection recommendations are generated, suggesting avoiding high-impact areas with an E exceeding 0.6, establishing buffer zones at least 1km away from sensitive areas, and using a low-noise flight mode within the buffer zone, reducing propeller speed by 20% in low-noise flight mode. This module achieves synergy between exploration and ecological protection, meeting green exploration requirements and reducing exploration damage to the ecological environment.

[0040] Specific implementation of an intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification: Example 1: Intelligent field exploration in a copper mine area in Yunnan (complex mountainous terrain) This example focuses on a copper mine exploration project in Yunnan Province, covering an area of ​​approximately 80 km². 2 The terrain is mainly mountainous and canyonous, with slopes reaching 35° in some areas. There is a diurnal temperature range of -5°C to 35°C, and occasional sandstorms. The target minerals are chalcopyrite (characteristic wavelength 500-600nm) and associated pyrite (characteristic wavelength 850-950nm). High-precision mineral identification and reserve estimation are required, while ensuring stable operation of the equipment under complex terrain and climate conditions. The specific implementation process is as follows: 1. System Deployment and Drone Preparation The drone's adaptive mounting module utilizes the DJI M350RTK hexacopter foldable platform, with a maximum takeoff weight of 15kg and a flight time of 65 minutes. Three core checks are performed before takeoff: First, battery compartment preheating: when the ambient temperature is below 0℃, a 20W constant-temperature heater is activated, stopping once the compartment temperature reaches 10℃ to ensure battery discharge efficiency. Second, magnetorheological damping calibration: 5-500Hz sinusoidal vibration is input through a vibration test bench, and the viscosity of the magnetorheological fluid is adjusted to stabilize the vibration attenuation rate at 96%. Third, differential RTK-GPS calibration: signals from three nearby ground reference stations are received to complete static positioning calibration, achieving a positioning accuracy of ±0.25m. The millimeter-wave obstacle avoidance radar is in omnidirectional detection mode, with a detection distance set to 80m and an obstacle recognition threshold set to a diameter of 0.5m. The terrain following function parameters are set as follows: initial altitude 80m, automatic 10m increase when slope > 30°, and triggering a slow descent warning when slope > 35°.

[0041] The dynamic multispectral acquisition module is equipped with a Headwall Photonics Hyperspec VI20 band tunable multispectral camera with a spectral range of 350-1100nm. For the characteristic bands of copper ore, the bandwidth of the 500-600nm and 850-950nm bands is set to 5nm (to improve feature resolution), while the bandwidth of the remaining bands is set to 15nm (to balance data volume). The high-resolution thermal infrared camera is a FLIRVue Pro R1280×1024 with a temperature range of -30℃ to 200℃, an accuracy of ±1℃, and a sampling frequency of 10Hz. The pulsed lidar is a Velodyne VLP-16 with a point cloud density of 100 points / m². 2 The ranging accuracy is ±3cm. Multi-sensor synchronization is achieved through a 1PPS synchronization clock module (model TrimbleBD982). The synchronization signal is output from the UAV flight control system, triggering the camera and LiDAR respectively. The measured synchronization error is 350ns. Data storage uses a 2TB Samsung 990 Pro NVMe SSD, formatted and allocated as follows: 1.5TB for raw data storage and 0.5TB for preprocessing temporary data. The measured read / write speed reaches 1.2GB / s.

[0042] 2. Data Acquisition and Preprocessing Mission planning is completed via an industrial-grade touchscreen terminal (10-inch, 10-hour battery life) of the digital twin control and feedback module, employing polygon area drawing functionality to cover an area of ​​80km. 2 The exploration area is divided into 8 10km zones. 2In the sub-region, the initial flight path spacing is set to 50m, and the flight speed is 8m / s. After takeoff, the spectral feature sensing unit calls the preset copper ore spectral library in real time and dynamically adjusts the camera band gain: the gain in the 500-600nm and 850-950nm bands is increased by 20% to enhance the feature signal. When flying to sub-region 2 (slope 32°), the dynamic trajectory optimization module detects that the probability of chalcopyrite in the preliminary AI identification results reaches 82%, and immediately calculates the trajectory curvature: At this time, t = 300s, x = 2400m, y = 1800m, dx / dt = 8m / s, dy / dt = 6m / s, d 2 x / dt 2 =0.3m / s 2 d 2 y / dt 2 =0.2m / s 2 Substituting this into the equation, we get κ(t) = |0.3×6 - 8×0.2| / (8 2 +6 2 )^(3 / 2)=|1.8-1.6| / 1000=0.0002m⁻ 1 <0.005m⁻ 1 This satisfies the requirement for smoothness, while reducing the flight strip spacing to 30m and increasing data collection density by 67%.

[0043] The data preprocessing module follows a three-level process: Radiation correction collects dark current data every 10 minutes (lens cap closed, exposure time consistent with working conditions), eliminates the influence of dark current through linear fitting, and the measured reflectance error after correction is 1.2%; Atmospheric correction integrates the MODTRAN 6.0 model with on-site meteorological station data (temperature 25℃, humidity 60%, aerosol optical thickness 0.2), selects a continental aerosol model, and adopts a mid-latitude summer mode for water vapor profile, improving the atmospheric transmittance at 500nm from 0.7 to 0.98 after correction; Geometric correction selects 10 ground control points (measured via RTK, accuracy ±0.1m), combines them with lidar point clouds to generate a digital elevation model (DEM), and completes geometric fine correction using bilinear interpolation, with a measured plane error of 0.7m. The time-series data fusion targets three time periods (9:00, 12:00, and 15:00) in sub-region 3, assigning weights based on signal-to-noise ratio (SNR): 9:00 (SNR 38dB, weight 0.2), 12:00 (SNR 45dB, weight 0.6), and 15:00 (SNR 35dB, weight 0.2). After fusion, the SNR is improved to 48dB. Noise filtering employs adaptive median filtering, with the window dynamically adjusted according to noise density. The window is expanded to 5×5 in the canyon shadow area and reduced to 2×2 in the open area, ultimately outputting a standardized time-series multispectral dataset of 256×256 pixels.

[0044] 3. AI-powered mineral identification and result visualization An improved ResNet-101 network is used for the cross-modal AI mineral recognition module. The input is multimodal data consisting of "multispectral (256×256×20) + lidar elevation (256×256×1) + thermal infrared temperature (256×256×1)". The cross-modal attention mechanism allocates weights using a Softmax function: multispectral reflectance weight 0.6, lidar elevation weight 0.2, and thermal infrared temperature weight 0.2 (the temperature of chalcopyrite is 3℃ higher than the surrounding rock, thus enhancing the temperature feature). Before recognition, feature values ​​are calculated using a multimodal mineral feature enhancement module. Where λ1=500nm, λ2=600nm, R(λ) is the average reflectance in the 500-600nm band (0.35), T(λ) is the normalized value of the corresponding thermal infrared temperature (0.6, measured chalcopyrite temperature 28℃, surrounding rock 25℃), z1=0m, z2=5m, μ(λ,z) ​​is the spectral attenuation coefficient (0.02m⁻) in the 500-600nm band and 0-5m depth of the copper ore region. 1 (From the Yunnan Provincial Geological Database), substituting the values, we get G=∫500^6000.35×0.6×e^(-∫0^50.02dz)dλ=0.21×e^(-0.1)×100≈18.9. This value is significantly higher than that of calcite (G≈12.5), effectively distinguishing similar minerals.

[0045] The model was pre-trained on 1 million labeled samples, with copper ore samples accounting for 30% (including copper ore data from Yunnan, Jiangxi, and other regions). An additional 500 field validation samples from the local area (purity determined by X-ray fluorescence spectrometry) were added for incremental learning. The model update time was 1.2 hours, and the final recognition accuracy reached 95.2%. The mineral purity inversion unit collected 50 chalcopyrite samples each with a purity ranging from 30% to 100%, and established a reflectance-purity calibration curve (R²). 2 =0.96), extract the reflectivity of 500-600nm from the unknown area to 0.38, calculate the initial purity to be 75%, combine the lidar texture features (roughness 0.4m, slope 18°), the correction coefficient k=1-0.02×(0.4+18 / 10)=0.944, the final purity=75%×0.944≈70.8%, the deviation from the actual sampling purity of 71.2% is only 0.4%.

[0046] The exploration-mining linkage visualization module constructs a 3D terrain model (elevation accuracy ±0.05m) based on lidar point clouds, overlays chalcopyrite identification results to generate a dynamic heat map, and sets the color scale to 0-100% (3% interval), with areas of purity above 70% displayed in dark red; the mining simulation inputs open-pit mining parameters (step height 12m, platform width 8m), automatically generates the mining area, and uses Kriging interpolation to estimate reserves with an error of 5.8%; it supports horizontal slice analysis (each 5m layer) and vertical profile extraction (along the canyon direction), and exports the results in Shapefile (for GIS analysis) and BIM (for the mining simulation software Revit) formats, with an export time of 8 minutes.

[0047] 4. Implementation effect verification By comparing the key indicators of traditional manual exploration with those of this system, the advantages of the system are verified: Table 1: Comparison of Key Exploration Indicators for a Copper Mine in Yunnan

[0048] Table 1 shows that this system has significant advantages in exploration efficiency, accuracy, and cost control. Traditional manual exploration is limited by terrain, covering only 65% ​​of the area, and its 45-day cycle is insufficient for rapid exploration needs. This system, through UAV terrain following and dynamic trajectory optimization, achieves 98% coverage and completes 80km in 3 days. 2 Exploration efficiency is increased by 15 times. Mineral identification accuracy is improved by 17.2%, mainly due to multimodal fusion and feature enhancement, effectively distinguishing chalcopyrite from similar minerals; reserve estimation error is reduced by 9.2%, attributed to accurate calculations of purity inversion and kriging interpolation. Labor costs are only 15% of traditional methods, as the number of field exploration personnel and laboratory testing is reduced, while avoiding safety risks in complex terrain, making it perfectly suited to the high-efficiency exploration needs of mountain copper deposits.

[0049] Example 2: Intelligent exploration of a limestone mining area in Shandong (including wetland sensitive areas) This example focuses on a limestone mine exploration project in Shandong Province, covering an area of ​​100 km². 2 The terrain is mainly plains with some gentle slopes (slope ≤ 15°), and there is a 10km stretch in the southeast. 2 In wetland sensitive areas (provincial-level ecological protection areas), the target mineral is limestone (characteristic wavelength 850-950nm). While ensuring exploration accuracy, it is necessary to minimize the impact on the wetland ecosystem. The specific implementation process is as follows: 1. System Deployment and Ecological Parameter Settings The drone's adaptive mounting module utilizes the DJI Mavic 3 Enterprise RTK hexacopter platform, with a maximum takeoff weight of 12kg and a flight time of 70 minutes. For flat terrain, the magnetorheological damping mount is set to a 6kg load capacity (for only a multispectral camera and LiDAR), with a vibration attenuation rate adjusted to 94%. Differential RTK-GPS calibration receives signals from two ground reference stations, achieving a positioning accuracy of ±0.3m. The millimeter-wave obstacle avoidance radar has a detection range of 50m. Due to the flat terrain, the terrain-following function has a fixed ground clearance of 80m, with only minor adjustments of 5m for slopes greater than 15°. For harsh environment protection, considering the local rainy weather, the camera compartment has undergone IP67 waterproof testing (no leakage after 30 minutes in 1m water), the positive pressure dust removal system operates at 3Hz (to reduce energy consumption), and the battery compartment's constant-temperature heating element has a threshold of 5℃ (to cope with low morning temperatures).

[0050] The dynamic multispectral acquisition module's 20-band camera targets the limestone's characteristic wavelength range of 850-950nm, with a bandwidth of 8nm, and the remaining bands are set to 15nm. The thermal infrared camera is not currently mounted due to the small temperature difference between limestone and the surrounding rock (<1℃). Only the multispectral camera and lidar (Velodyne VLP-16, point cloud density 80 points / m²) are retained. 2 The measured synchronization error of the synchronous clock was 400ns. Data storage used a 2TB WD Black NVMe SSD with a read / write speed of 1GB / s. The exploration ecological impact assessment module pre-imported regional vegetation and wetland data: areas with high NDVI vegetation cover (NDVI > 0.6) were mainly distributed around wetlands, with wetland boundaries marked by satellite maps. These areas were characterized by 560nm reflectance > 20% and 860nm reflectance < 15%.

[0051] 2. Data Acquisition and De-identification Processing The mission plan will cover 100km. 2 The area is divided into 10 10km zones. 2 The sub-region has a flight strip spacing of 50m and a flight speed of 10m / s. For wetland-sensitive areas, a "no-fly buffer zone" (1km from the wetland boundary) is set up. Within the buffer zone, the flight altitude must be reduced to 60m and the propeller speed reduced by 20% (low-noise mode). During data acquisition, the spectral feature sensing unit calls upon the limestone spectral library, increasing the gain in the 850-950nm band by 15%. The dynamic trajectory optimization module detects sub-region 7 (limestone probability 85%) and calculates the trajectory curvature κ(t) = 0.00015m⁻. 1 The flight strip spacing was reduced to 35m, increasing data density.

[0052] The data preprocessing module's radiometric correction, combined with a diffuse reflection reference board (99% reflectivity @ 850-950nm), involves collecting reference board data three times before each takeoff and calculating the sensor response coefficient k = R_ref / R_meas (R_ref = 0.99, R_meas = 0.98). Real-time reflectivity correction is performed with an error ≤ 2.5%. Atmospheric correction utilizes the MODTRAN model and on-site meteorological data (temperature 28℃, humidity 55%), selecting marine aerosol type (near the coast). Geometric correction has a measured plane error of 0.8m. The exploration data desensitization unit, designed for multi-team collaboration (3 units involved in the analysis), adds Gaussian noise (standard deviation 4m, ε = 1.5) to the coordinate data. Ground control points are blurred to "western part of XX town." Mineral distribution is aggregated into a 100m × 100m grid, with an average limestone content error of 7.2% within the grid. Even after desensitization, the data retains the western-to-eastern trend of limestone distribution, ensuring secure data sharing.

[0053] 3. AI Identification and Ecological Impact Assessment The cross-modal AI mineral recognition module takes multispectral and lidar data as input. In the ResNet-101 network, the limestone feature band weight is 0.7, and the lidar elevation weight is 0.3 (the limestone region has relatively flat elevations). The model's self-iterative update unit completes 100km of iteration because the limestone characteristics in this region are similar to the pre-training samples (northern limestone mines). 2 After exploration, 1200 high-confidence samples (identification probability ≥ 90%) and 500 low-confidence samples were selected. 38 misclassifications were manually verified (error 7.6% < 8%). Only fine-tuning was performed (Adam optimizer, learning rate 1e-5, 50 iterations). After fine-tuning, the accuracy improved from 94% to 95.5%.

[0054] The exploration ecological impact assessment module calculates NDVI: NIR (850nm reflectance 0.55), R (650nm reflectance 0.25), NDVI=(0.55-0.25) / (0.55+0.25)=0.375, determining the wetland periphery as a medium vegetation cover zone (NDVI 0.3-0.6); the wetland area is identified as having 23% reflectance at 560nm and 13% reflectance at 860nm, consistent with wetland characteristics; an ecological impact index E=0.4×(1-0.45)+0.3×(1-800 / 1000)+0.3×0.08=0.4×0.55+0.3×0.2+0.024=0.22+0.06+0.024=0.304 (low impact), generating ecological recommendations: avoid the wetland buffer zone, detour the flight path around the eastern side of the wetland, and use a low-noise mode in the buffer zone.

[0055] The exploration-mining linkage visualization module generates a 3D heat map of limestone distribution, with areas containing over 85% purity concentrated in the west. Mining simulation inputs underground mining parameters (tunnel spacing 25m, mining depth 50m), with a reserve estimation error of 6%. The 3D fracture analysis unit extracts fractures from the lidar point cloud, with a density of 2.5 m³ / m². 2 The correlation with limestone abundance was |r|=0.65 (moderate correlation), marking high-potential areas (purity > 80% and fracture density < 3 m / m). 2 The recommended drilling locations were 12, and subsequent field drilling verified that 10 of them contained high-purity limestone, resulting in an invalid drilling rate of 16.7%.

[0056] 4. Implementation effect verification Comparison of ecological impact and overall performance between traditional exploration and this system: Table 2: Comparison of Key Indicators in the Exploration of a Limestone Mine in Shandong

[0057] Table 2 data highlights the advantages of this system in ecological protection and overall performance. Traditional manual exploration, due to sampling near wetlands, results in 25% vegetation damage. This system, through buffer zone settings and low-noise mode, reduces ecological interference to 3%, meeting the requirements of green exploration. Data anonymization solves information security issues in multi-team collaboration, avoiding disputes arising from coordinate leakage. The invalid borehole rate is reduced by 23.3% due to accurate 3D fracture analysis and high-potential area marking, reducing blind drilling. The coverage area of ​​a single mission is increased by 7 times, attributed to high-speed flight in plain terrain and energy consumption optimization (flight speed 10m / s, sensor simplified mode reduces energy consumption by 30%). The equipment failure rate is reduced from 18% to 3%, thanks to the waterproof and temperature-controlled design of the harsh environment protection unit, enabling stable operation even in rainy weather, fully adapting to the exploration needs of limestone mines in sensitive ecological areas.

[0058] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification, characterized in that, include: The UAV adaptive mounting module adopts a six-rotor folding platform, equipped with magnetorheological shock absorption mounts, and has a compatible interface to support the simultaneous mounting of multispectral cameras, LiDAR, and thermal infrared cameras; it also integrates millimeter-wave obstacle avoidance radar and terrain following function. Dynamic multispectral acquisition module: includes a 20-band tuned multispectral camera, a thermal infrared camera and a pulsed lidar; equipped with a spectral feature sensing unit, it dynamically adjusts the camera band combination according to the target mineral's preset spectral library, which contains 32 common mineral feature bands; the module achieves spatiotemporal alignment of data from each sensor through a 1PPS synchronization clock. The time-series data preprocessing module performs three levels of correction on the raw data: radiometric correction, atmospheric correction, and geometric correction. Radiometric correction is based on real-time monitoring of dark current. Atmospheric correction integrates the MODTRAN model with real-time meteorological data to eliminate the dynamic effects of water vapor and aerosols. Geometric correction combines lidar point clouds and ground control points. A new time-series data fusion function has been added, which uses weighted average fusion of multispectral data from 3-5 time periods in the same area. The weights are allocated according to the data signal-to-noise ratio, and the final output is a standardized time-series multispectral dataset. Cross-modal AI mineral recognition module: adopts a multispectral + lidar + thermal infrared cross-modal fusion architecture; the basic network is ResNet-101, the input is a 512×512 pixel image, and a cross-modal attention mechanism is added to the network to assign dynamic weights to multispectral reflectance, lidar elevation and thermal infrared temperature features. The module has a built-in mineral feature transfer learning library that supports incremental learning for new mineral types. The exploration-mining linkage visualization module: Based on the lidar point cloud, a 3D terrain model is constructed, and the mineral identification results are superimposed on the 3D terrain model to generate a dynamic heat map that supports real-time updates; the module adds a mining simulation function, which can automatically generate the mining range and reserve estimation after inputting mining process parameters. The reserve estimation uses Kriging interpolation, and the module supports horizontal slicing, vertical profile, and reserve trend prediction analysis. Digital Twin Control and Feedback Module: Builds a digital twin model of the UAV and exploration area, mapping the UAV's location, sensor status, and mineral distribution in real time; the module achieves real-time data transmission and command issuance through 5G / satellite dual-mode transmission; the module is equipped with task replanning and health monitoring functions, and also realizes the automatic generation of exploration reports, which include mineral distribution, reserve estimation, and mining recommendations.

2. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, Also includes: Mineral Feature Enhancement Module: Enhances mineral differentiation by fusing the integral features of multispectral reflectance and thermal infrared radiation values. The calculation formula is as follows: Where G is the enhanced multimodal eigenvalue; λ1 and λ2 are the characteristic band intervals of the target mineral; R(λ) is the multispectral reflectance at wavelength λ; T(λ) is the normalized value of thermal infrared radiation temperature at wavelength λ, which is obtained by converting the measured temperature from the thermal infrared camera; z1 and z2 are the reference depths from the surface to the subsurface in the exploration area; and μ(λ,z) ​​is the spectral attenuation coefficient at wavelength λ and depth z, which is obtained by querying the geological database.

3. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, Also includes: Dynamic trajectory optimization module: Adjusts the flight path based on real-time mineral identification results. The optimized trajectory curvature is calculated as follows: Where κ(t) is the trajectory curvature at time t, the smaller the curvature, the smoother the trajectory; x(t) and y(t) are the planar coordinates of the UAV at time t, dx(t) / dt and dy(t) / dt are the velocities of the UAV in the x and y directions at time t, and d 2 x(t) / dt 2 d 2 y(t) / dt 2 Let t be the accelerations of the UAV in the x and y directions.

4. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, The cross-modal AI mineral identification module also includes a mineral purity inversion unit. First, mineral samples with known purity are collected, and the actual purity of the samples is determined using an X-ray fluorescence spectrometer. Then, the average multispectral reflectance of the samples in the characteristic bands is extracted.

5. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, The drone adaptive mounting module also includes a harsh environment protection unit. The camera compartment is waterproof and has a built-in positive pressure dust removal system. The system uses a HEPA filter to filter dust. The battery compartment is equipped with a 20W constant temperature heating element. When the temperature inside the battery compartment is below 0°C, the heating element automatically starts. The propeller adopts a carbon fiber-glass fiber composite structure.

6. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, The time-series data preprocessing module also includes an exploration data desensitization unit, which uses a combination of differential privacy and k-anonymity to achieve data desensitization. The specific process is as follows: Gaussian noise is added to the coordinate data, with a mean of 0. The noise is dynamically adjusted according to the regional sensitivity to meet the privacy protection budget ε=1.

5. The desensitized data maintains the relative positional relationship and trend characteristics of mineral distribution, and can be safely used for multi-team collaborative analysis, while avoiding the risk to exploration area rights caused by the leakage of geographic information.

7. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, The cross-modal AI mineral recognition module also includes a model self-iterative update unit, which updates the model every 100km. 2 The exploration task automatically triggers the sample update process, which selects new samples from the identification results. The new samples include high-confidence samples and low-confidence verification samples. The identification probability of high-confidence samples is not less than 90% and the spectral similarity with surrounding samples is less than 3%, and the number is not less than 1,000. The identification probability of low-confidence verification samples is between 40% and 60%, and the number needs to be manually verified on-site.

8. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, The exploration-mining linkage visualization module also includes a three-dimensional fracture analysis unit. First, the lidar point cloud data is preprocessed, including noise point removal and downsampling. Noise point removal uses statistical filtering. Then, an improved Hough transform is used to detect the linear characteristics of fractures. Next, the fracture density is calculated using a 10m×10m sliding window. The fracture density is the ratio of the sum of all fracture lengths within the window to the window area. Then, the correlation between fracture density and mineral abundance is analyzed using the Pearson correlation coefficient. When the absolute value of the correlation coefficient is not less than 0.7, it is considered a strong correlation. Finally, the potential exploration area is marked in red in the three-dimensional model.

9. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, The digital twin control and feedback module also includes an energy consumption optimization unit to construct an energy consumption prediction model. The input parameters of the model include terrain complexity, flight speed, and sensor operating mode. Terrain complexity is divided into three categories: flat, gentle slope, and steep slope. Based on the energy consumption prediction model, an algorithm is used to automatically plan the optimal path, dynamically adjust the sensor operating mode, and monitor the remaining battery power in real time.

10. The intelligent exploration system based on UAV multispectral remote sensing and AI mineral identification according to claim 1, characterized in that, Also includes: The exploration ecological impact assessment module first calculates the Normalized Difference Vegetation Index (NDVI), calculated as NDVI = (NIR - R) / (NIR + R), where NIR is the near-infrared reflectance and R is the red reflectance. Areas with an NDVI exceeding 0.6 are classified as high vegetation cover areas. Then, sensitive ecological zones are identified using multispectral data, including wetlands and rare plant areas. Next, an ecological impact index E is constructed: E = 0.4 × (1 - vegetation cover) + 0.3 × sensitive area distance factor + 0.3 × terrain damage risk. Vegetation cover is the percentage of areas with an NDVI exceeding 0.3, the sensitive area distance factor is 1 - distance / 1000 and is effective when the distance is less than 1000m, and the terrain damage risk is the product of the percentage of areas with a slope exceeding 25° and 0.

5. Finally, ecological protection recommendations are generated, suggesting avoiding high-impact areas with an E exceeding 0.6.

Images