A Uav-based multisensor framework for legal industrial Cannabis monitoring and open-access dataset development
Department of Electronics and Telecommunications, Faculty of Information Technology, Polytechnic University of Tirana, Albania
Department of Computer Engineering, Faculty of Information Technology, Polytechnic University of Tirana, Albania
⁎Corresponding author. genta.rexha@fti.edu.al@genta_rexhaAbstract
Industrial hemp cultivation is expanding and requires reliable monitoring for legal compliance and agricultural management. This paper presents a standardized UAV-based multisensor framework designed for Cannabis sativa L. It integrates RGB, multispectral, and thermal imaging as core modules, with hyperspectral and LiDAR as optional extensions. The framework sets protocols for sensor integration, flight planning, field measurements, and annotation, ensuring datasets that meet EU altitude limits (≤120 m AGL). Multi-altitude and multi-time-of-day acquisitions are proposed to capture spatial and diurnal variability. These data improve model robustness for phenotyping, stress detection, and THC compliance verification. Potential applications include precision agriculture, breeding, regulatory monitoring, environmental assessment, and illicit crop detection. Open-access datasets generated through this framework will support reproducibility, machine learning development, and collaboration among researchers, farmers, and regulators.
Graphical abstract
1Introduction
In recent years, the demand for accurate monitoring, identification, and classification of Cannabis sativa L., particularly industrial hemp and medical cannabis, has grown due to its expanding applications in medicine, textiles, dietary supplements, and bio-based industries [[1], [2], [3]]. The progressive legalization of hemp cultivation in many regions, alongside the increasing use of medical cannabis, has raised the need for efficient and reliable monitoring systems. Strict legal frameworks in the USA, Canada, and Europe, such as limits on Δ⁹-tetrahydrocannabinol (THC) content and licensing requirements, have reinforced the necessity of non-destructive, scalable methods to verify crop compliance and detect illicit production [[4], [5], [6], [7], [8], [9], [10], [11]].
Cultivation systems for Cannabis sativa vary considerably according to their intended end-use and legal context. Industrial hemp is predominantly grown outdoors in extensive fields to maximize biomass, fiber, or seed yields, making it particularly well suited for remote sensing and UAV-based monitoring [12]. In contrast, medical and recreational cannabis is often cultivated indoors or in controlled-environment greenhouses to optimize cannabinoid and terpene profiles, ensure product uniformity, and protect crops from environmental variability [[13], [14], [15]]. These production differences directly influence sensing strategies: UAV-based aerial imaging is optimal for large-scale outdoor hemp, while indoor facilities require proximal sensing systems or ground-based robotics to achieve equivalent precision.
Multispectral remote sensing has proven valuable for cannabis crop monitoring, particularly in phenotyping, disease detection, and biochemical trait estimation. Building on UAV-based applications, Impollonia et al. successfully inverted the PROSAIL model using multispectral UAV imagery to estimate key hemp traits, achieving fine-scale crop characterization with notable accuracy [12]. Si Ahmed et al. developed a system based on convolutional neural networks (CNN) and transfer learning techniques, trained on multispectral images, capable of detecting early stages of parasitic stress in cannabis plants [16].
Hyperspectral imaging (HSI) offers even greater potential due to its ability to capture the unique spectral signatures of cannabis plants. Azaria et al. used the AISA Eagle hyperspectral detector (400–1000 nm) to detect distinct cannabis-specific reflectance features, particularly in the 500–750 nm range [17]. Houmi et al. leveraged EO-1 Hyperion hyperspectral imagery to extract spectral profiles from known cultivation sites in Morocco, applying the Spectral Angle Mapper (SAM) algorithm to detect plantations in inaccessible areas [4]. Pereira et al. advanced this field by integrating near-infrared hyperspectral imaging with machine learning (ML) and sparse Principal Component Analysis (PCA) to isolate the most informative spectral bands for cannabis detection, demonstrating that reliable detection can be achieved using only four NIR bands [5]. Similarly, Gambardella et al. employed the CASI 1500 airborne hyperspectral sensor to survey Albania between 2012 and 2020, identifying over 5,000 suspected cannabis plots using feature extraction and PCA [6].
In the legal cultivation context, Lu et al. developed a non-destructive method to differentiate between five CBD-rich hemp cultivars and their growth stages using hyperspectral reflectance imaging, achieving classification accuracies above 96 % [14]. In a follow-up study, the same authors estimated cannabinoid content (CBD, THC, CBG) via linear discriminant models, surpassing 97 % accuracy in determining THC compliance [15]. Manggala et al. explored the feasibility of using mobile phone cameras as simplified hyperspectral devices to estimate chlorophyll content in cannabis leaves, demonstrating promising potential for accessible monitoring systems despite challenges in model generalizability [18].
ML and computer vision have further enhanced automation in cannabis monitoring. Sriram et al. developed a CNN using TensorFlow and Keras to detect visible signs of plant disease [19], while Chumchu and Patil compiled a seed classification dataset of over 3,400 high-resolution images across 17 cannabis seed types [20]. Çay compared deep learning (DL) and traditional models for cannabinoid prediction using terpene profiles, achieving accurate THC content estimation [21].
Satellite-based monitoring has also contributed significantly to cannabis detection. Ferreira et al. developed a data-driven ensemble approach for detecting illicit cultivation sites using high-resolution satellite imagery [9]. Bicakli et al. applied PlanetScope NDVI analysis to differentiate cannabis from crops such as wheat, corn, and alfalfa [7], while Sujud et al. integrated optical Sentinel-2 and radar Sentinel-1 data in Google Earth Engine, applying Random Forest (RF), Gradient Boosting (GBT), Classification and Regression Tree (CART), and Support Vector Machine (SVM) classifiers to detect cannabis fields in Lebanon [10].
Advanced image analysis techniques have further improved detection accuracy. Tanasa et al. implemented threshold-based segmentation to process thousands of UAV-acquired images for illegal field detection [11]. Lisita et al. validated object-based image analysis (OBIA) from SPOT-5 imagery using field surveys [8]. Hyperspectral studies by Holmes et al. and Schober et al. demonstrated plant organ identification [22] and nutrient quantification (N, P, K) [13], respectively, while Matros et al. distinguished male from female plants via spectral signatures [23].
Despite these advances, to the best of our knowledge, no open-access UAV-based multisensor dataset currently exists for legally cultivated hemp, annotated across genotypes, phenotypes, growth stages, and environmental conditions. Such a dataset would enable reproducible research, accelerate algorithm development, and facilitate regulatory oversight.
Table 1 summarizes the cannabis-related sensing studies discussed above. As shown, existing approaches span phenotyping, pest and disease monitoring, illicit cultivation detection, and dataset development, but they remain fragmented in terms of sensor types, platforms, and methodological protocols. None of these studies provide a standardized UAV-based multisensor framework for legal hemp cultivation. This gap underlines the need for the framework proposed in this paper.Research Focus Reference Description Phenotyping and Growth Characterization Impollonia et al. [12] PROSAIL inversion with multispectral UAV imagery for hemp phenotyping, achieving high accuracy. Lu et al. [14] Hyperspectral imaging for differentiating five CBD-rich cultivars and growth stages, >96 % accuracy. Lu et al. [15] Non-destructive determination of cannabinoids (CBD, THC, CBG) with linear discriminant models, >97 % accuracy for THC compliance. Manggala et al. [18] Use of mobile phone cameras as simplified hyperspectral devices for chlorophyll estimation. Schober et al. [13] Non-destructive method to quantify nutritional status of Cannabis using in situ hyperspectral imaging with chemometrics. Holmes et al. [22] Plant organ identification (flowers, stems, leaves) with NIR hyperspectral imaging and ML. Matros et al. [23] Non-invasive differentiation of cultivars and plant sex (male/female) via hyperspectral signatures. Pest and Disease Monitoring Si Ahmed et al. [16] Early parasitic stress detection using multispectral imaging with CNN and transfer learning. Sriram et al. [19] CNN developed with TensorFlow & Keras for detecting visible Cannabis plant diseases. Illicit Cultivation Detection Houmi et al. [4] EO-1 Hyperion hyperspectral imagery + SAM algorithm for detecting plantations in Morocco. Pereira et al. [5] Near-infrared hyperspectral imaging + ML + sparse PCA; reliable detection with only four NIR bands. Gambardella et al. [6] Airborne CASI 1500 hyperspectral surveys; over 5,000 suspected plots identified in Albania. Azaria et al. [17] AISA Eagle hyperspectral sensor (400–1000 nm); cannabis-specific reflectance features. Ferreira et al. [9] High-resolution satellite imagery with different CNN architectures for detecting illicit cultivation. Bicakli et al. [7] PlanetScope NDVI analysis to differentiate Cannabis from wheat, corn, and alfalfa. Sujud et al. [10] Optical Sentinel-2 and radar Sentinel-1 fusion with RF, GBT, CART, SVM classifiers in GEE for Lebanese sites. Tanasa et al. [11] Threshold-based segmentation of UAV images for illegal field detection. Lisita et al. [8] OBIA applied to SPOT-5 imagery, validated with field surveys. Datasets and Product Classification Chumchu & Patil [20] Dataset of 3,400 high-resolution images across 17 cannabis seed types. Çay [21] Comparison of DL and traditional models for THC prediction using terpene profiles.
Building on these prior efforts, the next section outlines the role of UAVs and multisensor platforms in precision agriculture, which provides the technological basis for designing a standardized monitoring framework tailored to legal industrial hemp.
This perspective is needed because, despite rapid progress in cannabis sensing, the field lacks a unified methodological direction. Existing studies remain fragmented across sensors, platforms, and regulatory contexts, making it difficult for researchers and regulators to compare results or build reproducible pipelines. By outlining a standardized UAV-based multisensor framework, this perspective fills a critical gap and provides a forward-looking roadmap for developing open, interoperable, and regulation-aligned datasets for legal industrial hemp monitoring.
2Unmanned Aerial Vehicles (UAVs) and Sensors in Precision Agriculture
UAVs have transformed precision agriculture by enabling rapid, high-resolution, and cost-effective data collection for crop monitoring, phenotyping, and management. UAV-based remote sensing platforms (UAV-RSP) overcome the spatial, temporal, and atmospheric limitations of satellite imagery by offering flexible deployment, adjustable flight altitudes, and compatibility with diverse sensor types [24]. Early studies, such as Sankaran et al. [25], anticipated that UAV adoption would accelerate the development of advanced aerial sensing systems for field-based phenotyping, providing plant breeders and researchers with powerful decision-support tools.
Recent advances have demonstrated the versatility of UAV platforms in integrating RGB, multispectral, hyperspectral, and thermal sensors for diverse agricultural tasks. Abdulsalam et al. [26] proposed a monocular vision-based approach using a fused-YOLO deep neural network architecture to autonomously detect multiple weed species, while Al-Najadi et al. [27] applied UAV-mounted thermal cameras to distinguish plant treatments via temperature mapping. Al-Obeidat and Li [28] further combined UAV imagery with state-of-the-art DL-YOLOv8 for aerial stress detection and Data-Efficient Image Transformers (DeiT) for leaf disease classification, creating a scalable, accurate, and computationally efficient pipeline.
Multispectral UAV imaging has also supported quantitative agronomic trait estimation. Li et al. [29] achieved high-accuracy rice leaf area index (LAI) estimation using CNN on five-band reflectance data, while Dong et al. [30] monitored physiological responses to water stress, such as decreased stomatal conductance and increased canopy temperature, at fine temporal resolution. UAV thermal imaging, as reviewed by Ndlovu et al. [31] and Sharma et al. [32], has proven effective for detecting water stress, with multisensor fusion approaches integrating thermal, multispectral, and hyperspectral data to improve accuracy and enable real-time analysis.
The adaptability of UAVs for object detection and classification has also been demonstrated across crop types. Praveen Kumar and Naveen Kumar [33] enhanced apple detection using a YOLOv7 architecture with multi-head attention mechanism for improved depth estimation, while Castellano et al. [34] employed lightweight Vision Transformers for semantic segmentation of crops and weeds in UAV-acquired multispectral imagery. Rehman et al. [35] addressed large-scale weed identification challenges by optimizing deep neural network architectures for UAV imagery, underscoring the importance of annotated datasets for real-world performance.
These developments highlight the critical role of UAV-based sensing in precision agriculture, offering high-resolution imagery, flexible deployment schedules, and the ability to capture crop-specific physiological and biochemical information. However, despite their proven success in other crops, targeted applications for legal industrial cannabis cultivation remain underdeveloped. Given the strict regulatory requirements on cannabinoid content and the economic significance of industrial hemp, UAV-based multispectral, hyperspectral, and thermal sensing can provide unprecedented capabilities for cannabis phenotyping, growth stage monitoring, stress detection, and cannabinoid content estimation.
Yet, the lack of open-access, labeled UAV datasets for cannabis under controlled, legal cultivation conditions poses a significant barrier to advancing research in this field. Without such datasets, capturing variations in genotype, phenotype, growth stage, and environmental conditions, model development for cannabis monitoring will remain fragmented. Therefore, the strategic integration of UAV and sensor technologies into cannabis research, coupled with the creation of a standardized UAV-based dataset, is essential to unlock the full potential of automated, high-accuracy cannabis monitoring systems.
3Proposed Conceptual Framework
The proposed UAV-based monitoring framework (Fig. 1) provides a standardized, evidence-based model for high-resolution monitoring of Cannabis sativa L. and other crops requiring detailed phenotypic assessment. It aligns operational parameters with current regulatory limits and best practices from remote sensing research, ensuring both reproducibility and adaptability across cultivation systems. In the European Union, UAV operations in the open category are restricted to a maximum altitude of 120 m above ground level (AGL) under Regulations (EU) 2019/947 and (EU) 2019/945 [36], a limit that balances operational safety, field coverage, and ground sampling distance (GSD).
In the reviewed literature, most UAV-based agricultural monitoring campaigns operated at altitudes between 100–120 m, producing multispectral imagery with a spatial resolution of 4–8 cm/pixel. Flight altitude plays a decisive role in determining the spatial resolution of image sensors and the accuracy of elevation measurements for ground targets. Fixed-altitude operations, as frequently reported, are well suited for observations over flat terrain, whereas variable-altitude strategies are more appropriate for heterogeneous landscapes [37]. Low-altitude unmanned aerial system (UAS) remote sensing technology offers unique advantages in capturing crop imagery at multiple scales, making it an essential tool for agricultural information monitoring. The integration of ML and DL methods further amplifies the analytical power of UAS datasets, enhancing their efficiency, scalability, and applicability to precision agriculture [38].
To ensure measurement consistency across campaigns, the framework incorporates standard practices that support stable sensor performance and accurate geospatial referencing. Basic radiometric and geometric calibration procedures are included to minimize variability caused by illumination changes, platform motion, or environmental fluctuations, enabling the imagery to maintain reliable spectral and spatial characteristics over time. Equally important is the use of a clear and well-distributed GCP layout, which improves orthomosaic alignment and enhances cross-campaign comparability. This spatial consistency is particularly valuable for ML workflows that rely on temporal analysis, multi-altitude integration, or the fusion of complementary datasets. Together, these calibration and georeferencing steps strengthen the reproducibility of the monitoring process and increase the robustness of models trained on the resulting data.The framework we propose specifies the use of three sensor modalities, RGB, multispectral, and thermal, each deployed at three distinct altitudes (low, medium, and high) to capture multi-scale spatial detail. Data acquisition is conducted at two-week intervals over a four-month crop cycle, yielding eight campaigns that collectively document both vegetative and flowering phases [14]. To enhance temporal robustness, the framework extends the recommendation of Al-Najadi et al. [27], originally applied to thermal imaging, to all three sensor types, capturing imagery at morning, midday, and evening within each campaign. This multi-timepoint approach accounts for diurnal variability in lighting, temperature, and canopy reflectance, improving model stability for classification, phenotyping, and stress detection. Each flight is required to produce a minimum of 500 georeferenced images, a threshold that ensures adequate spatial coverage for statistical reliability and ML model training.
Table 2 summarizes the operational specifications of the proposed framework, ensuring standardization across campaigns.Parameter Specification Growth Stage Coverage Entire 16-week growth cycle (vegetative + flowering phases) Campaign Frequency Biweekly campaigns (8 total per season) Sensor Types RGB, Multispectral, Thermal Altitudes per Sensor Low / Medium / High (≤120 m AGL) Time of Day Acquisitions Morning / Midday / Evening Minimum Images per Flight ≥500
In addition to imagery, each campaign incorporates the collection of core environmental and crop metadata, including air temperature, relative humidity, solar radiation, soil moisture, phenological stage, and management records, along with laboratory analyses of cannabinoid content from biweekly plant subsamples where feasible. This combined spatial–temporal–spectral–context dataset maximizes the analytical value of UAV acquisitions and facilitates reproducibility, interoperability, and cross-study comparisons.
4Potential Applications
The adoption of a standardized, UAV-based multisensor framework for hemp monitoring has the potential to benefit multiple domains, including agriculture, research, and regulatory compliance.
4.1Precision agriculture
The integration of RGB, multispectral, and thermal imagery enables real-time assessment of crop vigor, biomass accumulation, and water status, supporting site-specific management practices and optimizing inputs such as irrigation and fertilization.
4.2Phenotyping and breeding
High-resolution, multisensor datasets facilitate rapid, non-destructive evaluation of genotype performance under varying environmental conditions, aiding in the selection of cultivars with desirable traits such as high biomass yield, cannabinoid composition, and stress tolerance.
4.3Disease and pest detection
ML algorithms trained on UAV imagery can identify early signs of biotic stress, enabling timely interventions that reduce yield losses and minimize pesticide use.
4.4Regulatory compliance monitoring
Remote sensing combined with targeted field sampling can serve as a scalable, non-invasive tool to ensure hemp crops remain within legal THC thresholds, reducing the reliance on extensive manual inspections.
4.5Environmental monitoring
UAV-based datasets can be used to study the effects of soil conditions, irrigation regimes, and climatic factors on hemp growth and chemical composition, providing insights into sustainable production practices.
4.6Illicit crop detection and eradication support
While primarily designed for legal industrial hemp, the same UAV-based multisensor framework can be adapted to assist law enforcement agencies in detecting, mapping, and monitoring illegal cannabis cultivation, particularly in remote or inaccessible areas. Integrating cannabis-specific spectral and structural signatures with ML classifiers can improve detection accuracy while reducing false positives in heterogeneous landscapes.
Conclusions
UAV-based multisensor imaging, when standardized and integrated into a unified acquisition and annotation workflow offers unprecedented capabilities for legal hemp monitoring. This study presents a framework, balancing core sensors with optional advanced modalities, and aligning operational parameters with EU UAV regulations. By generating interoperable, annotated datasets, it enables reproducible ML applications in phenotyping, stress detection, cannabinoid compliance, and environmental assessment. Making available an open-access UAV-based cannabis dataset will accelerate research, enable comparative assessment of detection and classification algorithms, and foster collaboration among researchers, industry, and regulatory bodies. Implementing this framework will set a precedent for standardized UAV-based crop monitoring worldwide, with immediate priorities on pilot deployments and dataset publication to support large-scale adoption.
Limitations
Limitations of this approach include potential variability in sensor calibration across UAV platforms, reduced data quality under adverse weather or lighting conditions, and privacy or regulatory restrictions on aerial imaging. Additional challenges may arise from the logistical demands of multi-time-of-day data acquisition, such as operator availability, battery management, and the need for repeated equipment setup within the same day. Furthermore, UAV flight regulations can vary considerably between countries, potentially restricting altitude, flight times, or sensor use, which may limit direct replication of the framework in some regions.
UAV-based cannabis monitoring may be subject to crop-specific legal restrictions. Even in legal cultivation settings, permits for aerial imaging and geospatial data collection may be required, and authorities may limit where and how monitoring can take place. These regulatory constraints could restrict data acquisition or sharing, affecting the replicability of the framework across different jurisdictions.
Future Work
Future work will focus on developing the first UAV-based multisensor dataset for cannabis in Albania, starting in 2026 with the establishment of the country’s first legal cannabis farm. As no open-access dataset currently exists, the authors aim to make all collected data publicly available under an open license once field access is permitted. This plan, however, remains contingent on national regulatory approvals, and legal constraints may limit the scope or level of detail that can be shared. Within these boundaries, the dataset will be designed to support transparency, reproducibility, and cross-study comparability.
Future developments will also include a dedicated dataset with hyperspectral and LiDAR acquisitions, offering expanded capabilities for biochemical profiling, structural characterization, and 3D canopy reconstruction.To support reproducibility, the framework will be accompanied by standardized protocols for flight planning, sensor calibration, metadata acquisition, data annotation, and file structuring (naming, formats, and geospatial referencing). A complete technical guide, including GCP layout templates, camera settings, ML-ready preprocessing steps, and recommended quality-control thresholds, will be included when the dataset is released. These elements ensure that other researchers can replicate both the acquisition workflow and the analytical pipeline across different geographic regions and cultivation systems.
Ethics Statement
The authors confirm that they have read and followed the ethical guidelines for publication in Data in Brief. This study did not involve human or animal subjects, nor did it use data from social media platforms.
Declaration of generative AI and AI-assisted technologies in the writing process
During the preparation of this work the authors used ChatGPT (OpenAI) to improve the clarity, grammar, and style of the text. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
Acknowledgments
This publication was made possible through the financial support of the Albanian National Agency for Scientific Research and Innovation (NASRI), under Grant No. 948/1 Prot., dated 12 June 2025. The content is solely the responsibility of the authors, and the views expressed do not necessarily reflect those of NASRI.
The authors gratefully acknowledge the Albanian National Agency for Cannabis Control for their valuable discussions and insights on the subject.
Declaration of Competing Interest
The authors declare no known financial or personal conflicts that could have influenced this work.