¹ China Tobacco Guangxi Industrial CO.,LTD., Nanning, Guangxi, China, Nanning, Guangxi, China
² School of Computer Electronics and Information, Guangxi Key Laboratory of Multimedia Communication and Network Technology, Guangxi University, Nanning530004, China, Nanning, China, China
³ School of Computer and Electronic Information, Guangxi University, Nanning, 530000, China, Nanning, China, China

Ying Lu
Roles: Funding Acquisition, Project Administration, Resources, Validation

Yan Chen
Roles: Conceptualization, Writing – Original Draft Preparation

Yingxiong Nong
Roles: Supervision, Validation, Visualization, Writing – Original Draft Preparation

Jianqin Luo
Roles: Methodology

Zhibin Chen
Roles: Data Curation

Cong Huang
Roles: Software

Zihao Huang
Roles: Formal Analysis

Yingying Jiang
Roles: Validation

Hong Wu
Roles: Investigation

Global climate change is reshaping the landscape of agricultural production in unprecedented ways. The increasing frequency and intensity of extreme weather events, in particular, pose a fundamental threat to global food security and the stable supply of high-value cash crops.^1–3 Tobacco, as a key economic crop in China, has an industry chain whose stability is directly linked to regional economic development and the implementation of rural revitalization strategies.⁴ However, in core production areas like Guangxi, tobacco cultivation faces severe challenges arising from a combination of complex geography and a variable climate.^5,6 For instance, the prevalent acidic red soil in the Hechi region (average pH 5.3) not only limits nutrient availability but also introduces the potential risk of aluminum toxicity.^7,8 Meanwhile, in the Baise region, the annual evaporation rate far exceeds precipitation, causing seasonal drought to become a chronic production bottleneck.^8,9

The traditional cultivation management paradigm relies primarily on fixed farming schedules and universal technical guidelines based on historical experience.⁹ This “one-size-fits-all” management model inherently ignores the micro-domain environmental variability that exists both within and between plots.¹⁰ Consequently, when confronted with sudden, regional meteorological disasters, this paradigm—lacking fine-grained sensing and rapid response capabilities—often leads to delayed decision-making, inappropriate measures, and ultimately, significant economic losses. Statistical data show that in 2022 alone, droughts and floods caused an average yield reduction rate as high as 18% for tobacco in Guangxi. This harsh reality highlights the core challenge facing the tobacco industry today: an urgent need for a strategic transformation from a reactive, “post-event remediation” model to a proactive, “pre-event precision prediction and intra-event dynamic intervention” defense model.

To address the aforementioned challenges, precision agriculture technologies have emerged, leading to two mainstream research pathways.¹¹ The first approach is represented by crop growth mechanistic models (e.g., DSSAT, APSIM).^12,13 Although these models can simulate the growth processes of crops under specific environments based on biological principles, their rigid model assumptions exhibit inherent limitations in capturing the complex, non-linear responses caused by multiple stress factors (such as the concurrent onset of drought and high temperatures).¹⁴ More importantly, they often struggle to integrate real-time, multi-source field data, thus failing to provide dynamic, instantaneous decision support.^15,16

The second approach involves purely data-driven machine learning methods.¹⁷ While these methods have achieved significant success in specific tasks like image recognition and yield prediction,^18,19 their application in agricultural decision support still faces three major bottlenecks: First, a lack of interpretability; their “black-box” nature makes the decision-making process difficult for agronomists to understand and trust.²⁰ Second, strong data dependency; model performance is highly reliant on large-scale, high-quality labeled data, yet data acquisition in agricultural settings is costly.²¹ Third, insufficient generalization ability; the reliability of these models is difficult to guarantee, especially when dealing with small-sample, long-tail (i.e., infrequent but high-impact) disaster events.²²

Consequently, a clear and critical Research Gap has emerged: there is currently no holistic solution that can organically combine the predictive power of mechanistic models with the adaptability of data-driven methods, deeply integrate multi-dimensional perceptual data (such as visual, environmental, and time-series data), and embed this comprehensive intelligence into a practical and trustworthy on-farm decision-making framework. This study aims to directly confront this challenge and bridge the gap between theoretical models and field practice.

This paper proposes and implements a “Perception-Cognition-Decision-Execution” closed-loop intelligent defense system. Through technological integration and innovation, this system aims to provide a systematic and proactive solution for the disaster management of high-value crops. The main contributions of this study can be summarized in the following four points:

Proposing a novel proactive defense framework: This study is the first in the field of tobacco cultivation to establish a full-cycle dynamic emergency framework encompassing “pre-disaster warning, intra-disaster intervention, and post-disaster assessment.” This elevates the system’s role from a traditional “production support tool” to a “proactive risk defense hub,” achieving a fundamental shift in management philosophy.

Constructing a cognitive engine with a Multimodal Large Model at its core: By performing end-to-end fusion of multi-source, heterogeneous data—including soil data, meteorology, historical disaster records, and UAV hyperspectral remote sensing imagery—this study has created a cognitive model capable of deeply understanding the complex crop-environment interaction system. This model serves as the technological cornerstone for achieving “hyper-local precision” decision-making.

Designing an innovative Human-in-the-Loop (HITL) decision-making model: The system can automatically trigger the intervention of agricultural experts in high-risk decision scenarios based on assessed risk levels. This model ingeniously combines the data-processing advantages of artificial intelligence with the experiential wisdom of human experts, effectively addressing the trust challenges posed by the lack of interpretability in pure-AI applications and ensuring the reliability and practical applicability of critical decisions.

Completing a comprehensive, multi-dimensional empirical validation: Through large-scale field trials in two core production areas of Guangxi, system robustness stress tests based on historical disaster data, and detailed techno-economic benefit analyses, this study has systematically validated the system’s significant practical value in reducing disaster losses, improving management efficiency, and increasing economic returns.

2.1.1. Selection of the study area

The field trials for this study were located in three core tobacco-producing areas of the Guangxi Zhuang Autonomous Region: Hechi City (107°04′E, 24°41′N), Baise City (106°36′E, 23°54′N), and Hezhou City (111°32′E, 24°24′N). The selection of these three areas was intended to establish a gradient of differentiated ecological environments to comprehensively evaluate the model’s environmental adaptability.

Specifically, Hechi City is dominated by acidic red soils, which pose significant edaphic limitations to crop growth.²³ Baise City is situated in a typical dry-hot valley, where seasonal water stress is the primary constraint on agricultural production.²⁴ In contrast to the other two, Hezhou City is located in a subtropical monsoon climate zone with superior hydrothermal conditions and abundant annual rainfall, but it faces the unique challenge of a high-humidity environment²⁵ and possesses diverse soil types, including both red soils and limestone soils.²⁶ Therefore, these three experimental sites, with their significant heterogeneity in soil and climatic conditions, provided an ideal research platform for comprehensively testing the environmental adaptability and model generalization capability of the system proposed in this study.

2.1.2. Construction of the multi-source, heterogeneous dataset

To drive the model’s precision decision-making, this study constructed a multi-source, heterogeneous, and cross-spatiotemporal scale dataset. The composition of this dataset is detailed in Table 1, covering five key dimensions ranging from soil physicochemical properties to high-altitude remote sensing.

From the acquired multispectral imagery, key vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Red Edge Index (NDRE) were calculated to quantify crop health and stress. These indices are formulated as:

where NIR , Red , and RedEdge represent the reflectance values in the near-infrared, red, and red-edge bands of the spectrum, respectively. NDVI is a general indicator of vegetation vigor, while NDRE is more sensitive to chlorophyll content and can detect stress at earlier stages.

2.1.3. Data preprocessing

All raw data underwent a rigorous preprocessing pipeline before being input into the model to ensure data quality and the stability of model training.²⁷ This process primarily included the following steps:

where μ is the mean of the feature across all samples, and σ is its standard deviation. This transforms the features to have a mean of 0 and a standard deviation of 1.

The system framework designed in this study (Fig. 1) features a Multi-modal Large Model (MLM), deeply fine-tuned with agricultural domain knowledge, as its “cognitive hub.”²⁸ It is designed to achieve full-cycle, closed-loop management of cultivation risks and disaster events.²⁹ The core operational logic of the framework follows the dynamic emergency response principle of “pre-disaster warning and defense, intra-disaster monitoring and intervention, and post-disaster assessment and recovery,” effectively integrating data flow, analysis flow, and decision flow.³⁰

2.3.1. Model selection and secondary development

The core cognitive engine of this study—the MLM—was secondarily developed with domain-specific knowledge injection based on the architecture of a leading open-source multimodal large model, Qwen2.5-VL-32B. This model was chosen as the foundation primarily due to its powerful visual-language feature alignment and contextual understanding capabilities, which provide a solid foundation for fusing complex agricultural data.

2.3.2. Domain Fine-tuning dataset construction and training

To adapt the model to the specific scenarios of tobacco cultivation, we constructed a domain-specific instruction fine-tuning dataset containing over 50,000 high-quality image-text pairs.^31,32 This dataset was derived from the historical data in Table 1, with each pair consisting of an image of a tobacco plant or canopy captured by a UAV or from the ground,³³ a corresponding set of structured environmental parameters (e.g., soil moisture, air temperature, growth stage), and a diagnostic conclusion or management recommendation written by an agronomist (e.g., “Leaf yellowing, combined with low soil N content and recent rainfall, indicates nitrogen deficiency. Recommendation: apply XX kg/ha of urea.”).³⁴

The fine-tuning process aims to optimize an objective function, specifically the cross-entropy loss, which mathematically represents maximizing the likelihood of generating the correct text sequence. The objective function L is defined as:

where Lθ is the loss function parameterized by model weights θ ; N is the total number of samples; for each sample i , Ii is the input image, Si represents the structured environmental parameters, and Yi is the target text (the expert’s diagnosis). The training objective is to adjust θ to maximize the conditional probability P .

In terms of technical implementation, the model utilizes its native Vision Encoder to extract deep visual features from the images, while all environmental parameters and expert knowledge are textualized and fed into the language model. Through efficient Cross-modal Attention Fusion within the model’s deeper layers, the model learns the complex, non-linear relationships between image phenotypes, environmental factors, and agronomic diagnoses. This fusion is mathematically realized through a cross-modal attention mechanism, formulated as:

where Qtext is the query matrix derived from textual embeddings, while Kimg (keys) and Vimg (values) are matrices from the visual features extracted by the Vision Encoder. This mechanism allows the model to dynamically weigh the importance of different image regions based on the textual context. dk is the dimension of the keys.

The fine-tuning process was conducted on a 4-card NVIDIA A100 (80G) GPU cluster, using the AdamW optimizer with an initial learning rate of 1e-5 and a cosine annealing schedule, for a total of 20 training epochs.

The learning rate ηt at each epoch t was adjusted according to the cosine annealing schedule, given by:

where ηmax is the initial learning rate (1e-5), ηmin is the minimum learning rate (often close to 0), Tmax is the total number of epochs (20), and Tcur is the current epoch number.

To achieve a transition from plot-level to sub-plot-level fine-grained management, we divided the entire study area into 100 m × 100 m homogeneous management grid cells using geographic information system (GIS) technology. The system’s workflow is illustrated in Figure 2, and its core innovation lies in replacing traditional risk calculation formulas, which are based on linear weighting or thresholds, with the fine-tuned MLM.

Therefore, the MLM functions as a comprehensive risk assessment engine. For any given grid cell g , the Risk Index ( RIg ) is generated by a function fMLM that can be formally expressed as:

where RIg is the output risk score for grid g . The function fMLM , parameterized by the fine-tuned weights θ , takes the multimodal data for that grid as input, including remote sensing data ( Dremoteg ), meteorological data Dmetg , soil data ( Dsoilg ), and historical records ( Dhistg ), mapping these high-dimensional, heterogeneous inputs to a single risk index.

To systematically and rigorously evaluate the performance and value of the system proposed in this study, we designed a four-tiered comparative experiment:

Comprehensive Performance Comparison Test: (Objective: To validate the overall superiority of the system). The complete system proposed in this study (named Ours-Full (MLM)) was compared against two key baselines: (a) the classic crop growth mechanistic model (DSSAT), representing traditional model-driven methods; and (b) a simplified version of the system without grid-based management (Ours-Lite), to highlight the value of fine-grained management.

Core Module Ablation Study: (Objective: To quantify the contribution of the MLM). A comparative group, Ours-Formula, was added to the complete system. In this group, the MLM module was replaced by a traditional weighted risk assessment formula based on expert knowledge, with all other components remaining unchanged. By comparing the performance difference between Ours-Full (MLM) and Ours-Formula, we precisely quantified the performance gain brought by the MLM as the cognitive core.

Validation of Human-in-the-Loop Mechanism Effectiveness: (Objective: To verify the necessity and value of human-AI collaboration). Two operational modes were compared: (a) Ours-Full (Auto), a fully automated decision-making mode where all system-generated decisions are executed directly; and (b) Ours-Full (Human-in-the-Loop), the standard human-AI collaborative mode where high-risk decisions require expert review. This experiment aimed to evaluate the role of human collaboration in enhancing decision accuracy and reliability.

Cross-regional Generalization Capability Test: (Objective: To examine the model’s transfer application capability). An independent historical disaster dataset from Yuxi City, Yunnan Province, which was not used in model training, was utilized to back-test the disaster prediction accuracy of the pre-trained model, thereby assessing the system’s cross-regional generalization performance.

The quantitative results of all experiments were statistically analyzed using SPSS 26.0 software. For performance comparisons between two groups, a paired-samples t-test was used; for comparisons among multiple groups, a one-way analysis of variance (ANOVA) was conducted, followed by post-hoc tests. The significance level for all statistical tests was set at p < 0.05. The experimental results are presented as the mean ± standard deviation (SD).

To evaluate the overall efficacy of the system, we conducted a comparative analysis of its comprehensive performance against a traditional mechanistic model (DSSAT) and a simplified version of our system (Ours-Lite) under simulated disaster scenarios ( Table 2). The results indicate that the complete system proposed in this study, Ours-Full (MLM), demonstrates a distinct advantage across all key performance indicators. In terms of response efficiency, the average disaster response time for Ours-Full (MLM) was merely 2.8 ± 0.4 hours, a nearly five-fold improvement compared to the 14.2 ± 1.5 hours required by the DSSAT model (p < 0.001).

More critically, in the aspect of yield reduction control, Ours-Full (MLM) successfully maintained the average yield reduction rate at 7.1 ± 0.9%. This represents a significant decrease of 60.6% and 42.3% when compared to the DSSAT baseline (18.0 ± 2.1%) and the Ours-Lite version (12.3 ± 1.4%), which lacks fine-grained grid management. This outcome provides strong evidence that the combination of an MLM-based cognitive engine with a gridded, fine-grained management strategy is crucial for achieving proactive and efficient disaster prevention.

To precisely quantify the contribution of the Multimodal Large Model (MLM) to the system, we designed an ablation study comparing it with a system that employs a traditional weighted risk assessment formula (Ours-Formula) ( Table 3, Fig. 3). The experimental results irrefutably confirm that the MLM is the core driving force behind the qualitative leap in system performance.

In terms of decision accuracy, the MLM-equipped system (Ours-Full (MLM)) achieved an accuracy of 94.5 ± 2.1%, a significant improvement over the 78.2 ± 3.5% from the formula-driven version. From the perspective of practical agricultural applications, a more valuable metric is the false alarm rate. The MLM drastically reduced the system’s false alarm rate from 14.3 ± 2.1% to just 4.6 ± 1.5%, a substantial decrease of 67.8% (Fig. 3B). This implies that the system largely avoids unnecessary agricultural interventions, directly saving users costs in production materials—such as water, fertilizers, and pesticides—and labor.

This experiment was designed to quantify the practical benefits of a “Human-in-the-Loop” closed-loop system for enhancing decision reliability ( Table 4, Fig. 4). The results show that although the fully automated decision mode (Ours-Full (Auto)) already achieved a high intervention success rate of 91.3 ± 2.8%, the introduction of review and confirmation by agricultural experts in the Human-in-the-Loop mode (Ours-Full (Human-in-the-Loop)) elevated this rate to a near-perfect 99.2 ± 0.5%.

Simultaneously, because expert intervention effectively prevented misjudgments and excessive actions in a few complex scenarios, the resource waste rate was significantly reduced from 8.1 ± 1.9% to 2.0 ± 0.7%. This outcome demonstrates that human-AI collaboration is not a simple superposition of functions, but rather an organic coupling of AI’s macro-level data analysis capabilities with the deep domain knowledge of human experts. It stands as a key mechanism for ensuring fail-safe, high-stakes decisions and achieving optimal resource utilization.

To evaluate the model’s potential for transfer application, we conducted back-testing on the system using an independent historical disaster dataset from Yuxi City, Yunnan Province, which was not included in the model’s training data. On this entirely new dataset, the system achieved a prediction accuracy of 91.7% for the two primary types of historical disasters, namely droughts and floods. This result indicates that the knowledge acquired by the system’s underlying MLM concerning the relationship between crop stress and environmental factors possesses excellent universality and transferability, demonstrating the system’s significant potential for widespread application across different agro-ecological zones.

To intuitively demonstrate the system’s end-to-end workflow in a real-world context, we present a representative case of a successful precision intervention for aluminum toxicity risk in the red soil region of Hechi (Fig. 5).

On May 10, 2024, the system automatically classified a specific grid cell within a plot in Hechi City as “high-risk” (Risk Index, RI = 0.72) by fusing real-time soil sensor data (pH = 5.1) with early growth stress features detected in UAV multispectral imagery. The system then autonomously generated a decision plan: “Diagnosis: Moderate aluminum toxicity stress. Recommendation: Immediately apply 225 kg/ha of agricultural hydrated lime.” This was accompanied by an explanatory report, generated by the Multimodal Retrieval-Augmented Generation (MM-RAG) module, that included supporting imagery and data evidence. Concurrently, a high-risk review directive was instantly pushed to the mobile terminal of a contracted agronomist.

Within 1 hour and 45 minutes of receiving the alert, the agronomist arrived at the field, confirmed the system’s diagnosis through an on-site inspection, and authorized the execution of the intervention plan. Subsequent growth cycle tracking and final yield measurement revealed that the yield reduction rate for the intervened grid was successfully controlled at 4.0%. In stark contrast, a neighboring control grid cell, which had similar initial conditions but received no intervention, ultimately recorded a yield reduction rate of 16.5%. This case vividly illustrates the system’s full-loop capability, from perception and cognition to decision-making and execution.

Statement The data supporting the findings of this study are not publicly available due to strict commercial confidentiality and proprietary restrictions associated with China Tobacco Guangxi Industrial Co., Ltd. Data access may be granted to qualified researchers upon reasonable request to the corresponding author, subject to the approval of the data owner and the signing of a Non-Disclosure Agreement (NDA).

The authors would like to thank the editorial team and reviewers for their constructive feedback.

• 1.  Change O-C: Intergovernmental panel on climate change. World Meteorological Organization. 2007; 52(1–43): 1. Publisher Full Text
• 2.  Zscheischler J, Martius O, Westra S, et al.: A typology of compound weather and climate events. Nat. Rev. Earth Environ. 1(7): 333–347.
• 3.  Li J, Zhang Y, Bevacqua E, et al.: Future increase in compound soil drought-heat extremes exacerbated by vegetation greening. Nat. Commun. 2024; 15(1): 10875. PubMed Abstract | Publisher Full Text | Free Full Text
• 4.  He B, Mao Y, Zuo Y, et al.: Extraction of tobacco planting information based on UAV high-resolution remote sensing images. Remote Sens. 2024; 16(2): 359. Publisher Full Text
• 5.  Tian Y, Li X, Sun J, et al.: Identification of dry and wet events in karst and non-karst areas of Guangxi based on a novel optimization method involving the antecedent precipitation attenuation coefficient. J. Hydrol. 628: 130522.
• 6.  Wang S, Zhang Q, Yue P, et al.: Effects of evapotranspiration and precipitation on dryness/wetness changes in China. Theor. Appl. Climatol. 2020; 142(3): 1027–1038. Publisher Full Text
• 7.  Liu C, Jiang M, Liang Y, et al.: Root microbiota confers rice resistance to aluminium toxicity and phosphorus deficiency in acidic soils. Nat. Food. 2023; 4(10): 912–924. PubMed Abstract | Publisher Full Text
• 8.  Wang Y, Xiao J, et al.: Strategies for alleviating aluminum toxicity in soils and plants. Plant Soil. 2024; 504(1): 167–190. Publisher Full Text
• 9.  Basso B, Antle J: Digital agriculture to design sustainable agricultural systems. Nat. Sustainability. 2020; 3(4): 254–256. Publisher Full Text
• 10.  Heino M, Guillaume JHA, Müller C, et al.: A multi-model analysis of teleconnected crop yield variability in a range of cropping systems. Earth Syst. Dynam. 2020; 11(1): 113–128. Publisher Full Text
• 11.  Klerkx L, Jakku E, Labarthe P: A review of social science on digital agriculture, smart farming and agriculture 4.0: New contributions and a future research agenda. NJAS – Wageningen Journal of Life Sci. 2019; 1: 90–91. Publisher Full Text
• 12.  Hoogenboom G, Porter CH, Boote KJ, et al.: The DSSAT crop modeling ecosystem. Advances in crop modelling for a sustainable agriculture. Burleigh Dodds Science Publishing; pp. 173–216.
• 13.  Murmu K, Sarkar A, Sarma S, et al.: An Overview of Crop Simulation Models: A Decision Support System for Evaluation of Agronomic Managements in Climate Change Context. Mitigation and Adaptation Strategies Against Climate Change in Natural Systems. 2025; 399–426. Publisher Full Text
• 14.  Qi W, Feng L, Yang H, et al.: Increasing concurrent drought probability in global main crop production countries. Geophys. Res. Lett. 2022; 49(6): e2021GL097060. Publisher Full Text
• 15.  Dlamini L, Crespo O, van Dam J , et al.: A global systematic review of improving crop model estimations by assimilating remote sensing data: implications for small-scale agricultural systems. Remote Sens. 2023; 15(16): 4066. Publisher Full Text
• 16.  Chen Y, Tao F: Potential of remote sensing data-crop model assimilation and seasonal weather forecasts for early-season crop yield forecasting over a large area. Field Crop. Res. 2022; 276: 108398. Publisher Full Text
• 17.  Oikonomidis A, Catal C, Kassahun A: Deep learning for crop yield prediction: a systematic literature review. N. Z. J. Crop. Hortic. Sci. 2023; 51(1): 1–26. Publisher Full Text
• 18.  Khaki S, Wang L: Crop yield prediction using deep neural networks. Front. Plant Sci. 2019; 10: 621. Publisher Full Text
• 19.  Khaki S, Wang L, Archontoulis SV: A CNN-RNN framework for crop yield prediction. Front. Plant Sci. 2020; 10: 1750. Publisher Full Text
• 20.  Lundberg SM, Erion G, Chen H, et al.: From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2020; 2(1): 56–67. PubMed Abstract | Publisher Full Text | Free Full Text
• 21.  Li J, Xum M, Xiang L, et al.: Foundation models in smart agriculture: Basics, opportunities, and challenges. Comput. Electron. Agric. 2024; 222: 109032. Publisher Full Text
• 22.  Ovadia Y, Fertig E, Ren J, et al.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Adv. Neural Inf. Proces. Syst. 2019; 32.
• 23.  Dong Y, Yang J-L, Zhao X-R, et al.: Contribution of different proton sources to the acidification of red soil with maize cropping in subtropical China. Geoderma. 2021; 392: 114995. Publisher Full Text
• 24.  Zhong P, Wei S, Wang Z, et al.: Impact of seasonal drought on soil multifunctionality in dry–hot valley forests. Plant Soil. 2025; 516: 651–671. Publisher Full Text
• 25.  Climate-Data.org: Hezhou Climate: Weather Hezhou & Temperature by Month. Climate-Data.org; 2025. (accessed 2025-09-26).
• 26.  Lin K, Yang Z, Yu T, et al.: Enrichment mechanisms of Mo in soil in the karst region Guangxi. China. Ecotoxicol. Environ. Saf. 2023; 255: 114808. PubMed Abstract | Publisher Full Text
• 27.  Abdulmuttalib H, Al Doori M, Barsi Á, et al.: Remote Sensing Data Quality in the Era of AI. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences. 2024; X-3-2024: 1–11. Publisher Full Text
• 28.  Szwarcman D, Roy S, Fraccaro P, et al.: Prithvi-eo-2.0: A versatile multi-temporal foundation model for earth observation applications [J]. arXiv preprint arXiv:2412.02732. 2024.
• 29.  Purcell W, Neubauer T: Digital Twins in Agriculture: A State-of-the-art review. Smart Agricultural Technology. 2023; 3: 100094. Publisher Full Text
• 30.  Yao M-H, Hsu Y-H, Li T-Y, et al.: Agricultural disaster prevention system: Insights from Taiwan’s adaptation strategies. Atmos. 2024; 15(5): 526. Publisher Full Text
• 31.  Dai W, Li J, Li D, et al.: Instructblip: Towards general-purpose vision-language models with instruction tuning. Adv. Neural Inf. Proces. Syst. 2023; 36: 49250–49267.
• 32.  Liu H, Li C, Wu Q, et al.: Visual instruction tuning. Adv. Neural Inf. Proces. Syst. 2023; 36: 34892–34916.
• 33.  Han W, Chen S, Xiao S, et al.: Large-scale tobacco identification via a very-high-resolution unmanned aerial vehicle benchmark and a ConvFlow Transformer. Int. J. Appl. Earth Obs. Geoinf. 2025; 139: 104549. Publisher Full Text
• 34.  Awais M, Alharthi AHS, Kumar A, et al.: Agrogpt: Efficient agricultural vision-language model with expert tuning [C]. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE; 2025. 5687–5696.

Current Reviewer Status:

AWAITING PEER REVIEW

AWAITING PEER REVIEW

?