Foundation Models in Agriculture: A Comprehensive Review

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript “Foundation Models in Agriculture: A Comprehensive Review” gives an overview of foundation models (FMs) applied to agriculture, discussing their development, technical foundations, potential applications, and challenges. The topic is relevant and aligns with the journal’s scope. However, several limitations need to be addressed.

The paper presents FMs in agriculture but does not clearly explain how they advance beyond previous AI reviews. Clarify how this review contributes beyond previous AI-in-agriculture studies. While it describes the evolution of language models and vision transformers, it does not demonstrate their advantages over traditional machine learning methods. Additionally, the discussion on self-supervised learning and generative models would be better with real-world examples addressing data scarcity in agriculture. The paper also does not clearly explain how agricultural datasets are processed and validated.

The introduction covers FMs but does not focus enough on agricultural constraints. It lists data sources such as satellites, drones, and sensors but lacks details on data fragmentation and reliability. References to “multimodal data” and “real-time monitoring” should include more studies on UAV cost-effectiveness and sensor reliability.

In Section 3 (FMs in Agriculture), the paper lacks discussion on the differences between smallholder and industrial farms or the cost-effectiveness of these models. The section on Agricultural Knowledge Question Answering (AKQA) and robotics mentions these technologies but does not address adoption challenges in farms with low connectivity.

In Section 4 (Challenges of AFMs), the discussion on real-world adoption is brief. Large AI models require significant computing power, but the paper does not address how small farms can access them. There is also no analysis of scalability and costs. Farmers need clear AI-driven recommendations, but the paper does not discuss trust in automated decisions or how FMs integrate with existing farm management systems.

In Section 5 (Development Directions for AFMs), the paper correctly links FMs to agricultural challenges but does not provide a roadmap for real-world applications, particularly in resource-limited areas. Define short-, medium-, and long-term adoption plans. Address data governance, costs, and skill gaps for farmers using AI. Suggest best practices for managing data shifts across different crops and regions.

 

Specific Comments
Line 182
In Section “2.2.3. Training and Optimization”, the discussion lacks concrete examples of performance metrics and studies. Include details on cross-validation, sensitivity, specificity, and confusion matrices for agricultural tasks.

Line 396
Regarding the statement: “Recent advancements include multispectral and hyperspectral imaging (MSI/HSI), which enable real-time monitoring of plant health and early disease detection [120,121].”, given the increasing importance of hyperspectral imaging in agriculture and the role of algorithms in processing these data, this section should be expanded. Consider including recent studies in agricultural applications, such as “Application of near-infrared hyperspectral imaging for variety identification of coated maize kernels with deep learning”, “Identification of Geographical Origin of Chinese Chestnuts Using Hyperspectral Imaging with 1D-CNN Algorithm” or “Tracing Pistachio Nuts' Origin and Irrigation Practices through Hyperspectral Imaging”.

Line 549
References are needed for Table 5.

Line 639
in the statement: “AFMs can integrate data from diverse sources.”, how are different field data types (e.g., soil sensors, climate logs, drone images) unified into a single model? Provide clarification on data integration methods.

Comments on the Quality of English Language

It is necessary to revise the English language (there are some mistakes such as “precise argiculture” or “smart argiculture”).

Author Response

REVIEWER 1:

1) Clarify how this review contributes beyond previous AI-in-agriculture studies.

Authors’ response:
Thank you for this suggestion. To clarify our novel contributions:

1. We systematically introduce the development of general-purpose foundation models in computer science, including their technical evolution and core architectures (see Figures 1 and Tables 1-2), providing essential background for non-computer science researchers to understand these transformative AI technologies.
2. We comprehensively review existing AFMs and analyze their agricultural applications in knowledge Q&A, disease detection, and decision support (Table 3), offering practical insights for domain experts.
3. We identify unique challenges in developing AFMs, including agricultural data heterogeneity, temporal shifts in field conditions, and deployment constraints for smallholder farms.
4. We propose future directions emphasizing multimodal integration and intelligent decision systems to bridge AI innovation with agricultural needs.

By synthesizing insights from a large body of research, this work serves as both a technical reference for interdisciplinary researchers and an implementation guide for agricultural practitioners.

 

2) The paper also does not clearly explain how agricultural datasets are processed and validated.
Authors’ response:
We expanded Section 3 with:

 

The validation and processing of agricultural data begins with collecting multi-source texts (such as journal articles and extension materials) and converting them into standardized formats. This is followed by data cleaning to remove noise including garbled characters, non-UTF-8 characters, and the standardization of punctuation and capitalization. When constructing question-answering datasets, documents are filtered according to paragraph length criteria to divide into training, validation and test sets. Annotation tools are then employed to generate QA pairs where answers must precisely match source text segments without paraphrasing, while ensuring diverse question types. Quality control is implemented through phased supervision by domain experts, including collaborative annotation with regular sampling checks during the annotation phase, and consistency validation during evaluation. The inter-annotator agreement is quantified using both Cohen's Kappa and F1 scores, with the latter (reaching 0.86 in AgXQA1.1) proving more reliable for scenarios with multiple potential answers, thereby ensuring high data quality and task suitability[110].

 

110. Kpodo, J.; Kordjamshidi, P.; Nejadhashemi, A.P. AgXQA: A benchmark for advanced Agricultural Extension question answering. 1013

Computers and Electronics in Agriculture 2024, 225, 109349.

 

3)In Section 3 (FMs in Agriculture), the paper lacks discussion on the differences between smallholder and industrial farms or the cost-effectiveness of these models.

Authors’ response:
We expanded Section 3 with:

 

The implementation of AFMs must address the fundamental dichotomy between smallholder and industrial-scale farming systems. Current research reveals that most AFM architectures follow distinct technological pathways for these operational scales due to divergent infrastructure requirements and economic realities. Smallholder systems, representing over 80% of global farms, require specialized models emphasizing three critical characteristics: offline functionality for connectivity-challenged environments, minimal hardware dependencies, and extreme cost efficiency. These constraints have driven innovations like WhatsApp-based advisory platforms and TinyML implementations capable of operating on low-power edge devices, which achieve 70-80% accuracy at radically reduced deployment costs[106][107]. By contrast, industrial agricultural operations employ high-accuracy (>90%) cloud-based AFMs that process terabytes of multispectral drone and IoT sensor data through sophisticated transformer architectures. This approach requires substantial infrastructure investments in 5G networks and computing resources. The technological bifurcation reflects underlying socioeconomic realities - where smallholder systems prioritize accessibility and resilience, industrial operations optimize for precision at scale through capital-intensive solutions[108].

 

106. Touch, V.; Tan, D.K.; Cook, B.R.; Li Liu, D.; Cross, R.; Tran, T.A.; Utomo, A.; Yous, S.; Grunbuhel, C.; Cowie, A. Smallholder 1004

farmers’ challenges and opportunities: Implications for agricultural production, environment and food security. Journal of 1005

Environmental Management 2024, 370, 122536. 1006

107. Ray, P.P. A review on TinyML: State-of-the-art and prospects. Journal of King Saud University-Computer and Information Sciences 1007

2022, 34, 1595–1623. 1008

108. Klerkx, L.; Jakku, E.; Labarthe, P. A review of social science on digital agriculture, smart farming and agriculture 4.0: New 1009

contributions and a future research agenda. NJAS-Wageningen journal of life sciences 2019, 90, 100315.

4)In Section 4 (Challenges of AFMs), the discussion on real-world adoption is brief.

Authors’ response:
A new subsection 4.5 has been added to discuss：

 

subsection{Practical Deployment Challenges}

 

The operational deployment of AFMs faces fundamental technical constraints that manifest differently across farm scales. While current automation systems like Kenya's FarmBot demonstrate basic inclusivity for smallholders through compressed vision models[183], the integration of full-fledged FMs in low-connectivity environments remains constrained by substantial computational demands. Recent breakthroughs in extreme model compression show promise - binary transformers achieve 99\% size reduction while retaining 80\% original accuracy in controlled settings, potentially enabling localized FM deployment[184]. However, even distilled models confront memory limitations, as ViT-base architectures require >1.7GB memory, while edge devices typically offer <512MB RAM, forcing difficult tradeoffs between functionality and accessibility[132].

 

In real-world operating conditions, these challenges become more acute. Agricultural data streams exhibit inherent asynchrony (e.g., 15-300ms latency between visual and sensor inputs), which significantly degrades decision accuracy without temporal alignment modules - a particular concern for transformer-based multimodal agricultural foundation models where self-attention mechanisms are temporally sensitive. For smallholder farmers, this complexity is addressed through accessible "Model-as-a-Service" platforms rather than local deployment. Farmers can simply upload field photos via mobile apps to access cloud-based FM APIs. These solutions feature carefully designed interfaces with voice input and visual output to build trust in AI recommendations, while employing progressive deployment strategies - starting with basic tasks  before advancing to complex decision support. This approach democratizes access to advanced FM capabilities without requiring expensive infrastructure, though it introduces 40-60\% higher cloud dependency costs compared to industrial on-premise solutions[185].

 

 

182. Murdyantoro, B.; Atmaja, D.S.E.; Rachmat, H. Application design of farmbot based on Internet of Things (IoT). International 1158

Journal on Advanced Science, Engineering and Information Technology 2019, 9, 1163–1170. 1159

183. Fan, A.; Stock, P.; Graham, B.; Grave, E.; Gribonval, R.; Jegou, H.; Joulin, A. Training with quantization noise for extreme model 1160

compression. arXiv preprint arXiv:2004.07320 2020. 1161

184. Muslim, A.B.; Nordemann, F.; Tönjes, R. Analysis of methods for prioritizing critical data transmissions in agricultural vehicular 1162

networks. In Proceedings of the 2020 16th International Conference on Wireless and Mobile Computing, Networking and 1163

Communications (WiMob). IEEE, 2020, pp. 80–85.

132. Barman, U.; Sarma, P.; Rahman, M.; Deka, V.; Lahkar, S.; Sharma, V.; Saikia, M.J. Vit-SmartAgri: vision transformer and 1058

smartphone-based plant disease detection for smart agriculture. Agronomy 2024, 14, 327. 1059

5) In Section 5 (Development Directions for AFMs), the paper correctly links FMs to agricultural challenges but does not provide a roadmap for real-world applications, particularly in resource-limited areas.

Authors’ response:
We expanded Section 5.3:

 

Future development of AFM-based decision systems will follow a phased technical roadmap: initial focus on lightweight mobile applications delivering basic agricultural guidance through model compression; intermediate development of distributed learning architectures for cross-regional knowledge sharing and cost optimization; ultimately evolving into self-adaptive intelligent decision systems. Implementation will incorporate data governance protocols, tiered service models, and farmer-centric interface design to ensure equitable access across farm scales, progressively bridging the digital divide in agricultural intelligence adoption.

 

 

Specific Comments:

Line 182 (Training & Optimization):

 

Regarding this comment, we have provided more detailed and comprehensive explanations about the mentioned large models in Section 3：

AgriBERT was pre-trained from scratch using a standard Masked Language Modeling (MLM) objective, where 15\% of tokens in input sentences were randomly masked, and the model learned to predict them based on contextual information. The training corpus consisted of 46,446 food and agriculture-related journal articles (311 million tokens, 2.39 million words) rigorously cleaned to remove URLs, emails, and non-ASCII characters, supplemented by WikiText-103 and Penn Treebank datasets for linguistic diversity.

The model achieves this through an innovative dual-encoder architecture, featuring a fixed feature extractor and trainable model that employs contrastive learning to optimize sentence representations. By augmenting this framework with Bi-LSTM layers for label-chain modeling and incorporating semi-supervised techniques like proxy labeling and consistency regularization, Chains-BERT effectively leverages both labeled and unlabeled agricultural data.Experimental results demonstrate its superior performance, achieving an 86.5 Micro-F1 score on the CAIL2018 dataset—a 5.5-point improvement over vanilla BERT. The model also exhibits strong generalization capabilities, outperforming baseline BERT by 3.25 F1 points on SQuAD 2.0, validating its effectiveness beyond agricultural domains

 

Line 396 (Hyperspectral Imaging):

 

We thank the reviewer for their valuable suggestions. After reviewing the recommended literature, we agree these studies provide excellent examples of hyperspectral imaging applications in agriculture. We have incorporated them as references to strengthen our discussion of MSI/HSI technology and its practical implementations. These additions enrich the manuscript by offering more comprehensive case studies and technical details.

 

Line 549 (Table 5):

 

We thank the reviewer for pointing out this omission. We have now added the appropriate references to Table 5 to properly support the data presented.

 

Line 639 (Data Integration):

 

We appreciate this insightful question regarding data integration methods. In response, we have significantly expanded Section 5.1 to provide detailed explanations of how AFMs unify diverse agricultural data types：

 

Future AFMs should expand beyond traditional text and image modalities to harness emerging technological capabilities.  This multimodal expansion  forms the technical foundation for cross-domain integration across agricultural supply chains. Recent studies demonstrate the untapped potential of multimodal integration in agriculture: video analytics enable real-time crop monitoring during critical growth stages[186][187].

This technological convergence  directly addresses supply chain fragmentation challenges. By fusing complementary data streams—from spectral signatures to temporal growth patterns—multimodal AFMs can synchronize production data with downstream processing parameters (e.g., matching harvest quality metrics to storage facility conditions). The resulting  closed-loop intelligence system  not only optimizes field-level production efficiency but also enhances the resilience of agricultural supply networks through data-driven coordination[191].

 

186. Tzachor, A.; Devare, M.; King, B.; Avin, S.; Ó hÉigeartaigh, S. Responsible artificial intelligence in agriculture requires systemic 1167

understanding of risks and externalities. Nature Machine Intelligence 2022, 4, 104–109. 1168

187. Chandrasiri, C.; Kiridena, S.; Dharmapriya, S.; Kulatunga, A.K. Adoption of Multi-Modal Transportation for Configuring 1169

Sustainable Agri-Food Supply Chains in Constrained Environments. Sustainability 2024, 16, 7601. 1170

191. Ren, Y.; Huang, X.; Aheto, J.H.; Wang, C.; Ernest, B.; Tian, X.; He, P.; Chang, X.; Wang, C. Application of volatile and spectral 1177

profiling together with multimode data fusion strategy for the discrimination of preserved eggs. Food chemistry 2021, 343, 128515. 1178

Comments on the Quality of English Language：

Authors’ response:

We have carefully corrected all spelling errors and conducted a thorough proofreading of the manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Insufficient connection between sections and lack of logical: For example, in Section 5 "Development Directions for AFMs," the subsections (such as 5.1. Leveraging Multimodal FMs in Agriculture and Integrating AFMs Across the Agricultural and Food Sectors) lack logical transitions and do not clearly explain how multimodal data can facilitate cross-domain integration. Add transitional paragraphs to explain the intrinsic connection between multimodal technology and supply chain optimization.
2. Unclear correspondence between figures/tables and the main text: For example, "Chains-BERT" mentioned in Table 3 (Overview of Agricultural Models and Tasks) is only briefly referred to in the main text as a semi-supervised learning method, without detailed explanation of its specific architecture or comparative experiments. Supplement the main text with key model architecture diagrams or performance comparison data.

3. Disconnection between challenges and application scene: For example, Section 4 "Challenges of AFMs" discusses the difficulties of data acquisition but does not directly address the real-time data needs for agricultural robot automation mentioned in Section 3. 

4. Missing details on model training: For example, in Table 3, the "Pre-training Method" column for "AgriBERT" is "MLM trained from scratch," but the main text does not explain its masking strategy or the scale of the training dataset.

5. Lack of comparative analysis in experimental validation: For example, Section 3.1 mentions that "GPT-4 has an accuracy rate of 93% in agricultural certification exams," but no control group (such as traditional methods or performance of other models) is provided. A comparative experiment table could be added to show the performance differences of different models in the same task.

6. General discussion of data privacy issues: For example, Section 4.2 mentions that "Federated Learning can solve data privacy issues," but does not explain the specific implementation difficulties in agriculture.

7. Add case analysis (such as the actual deployment effect of the SAM model in apple harvesting).

8. The paper devotes a large amount of space to describing the basic models, which is not suitable for a paper intended for professionals rather than a textbook. Focus the paper on the current agricultural models in use, the existing problems and challenges, and potential areas for innovation and breakthroughs.

Comments on the Quality of English Language

1. Complex sentence structure affecting readability: For example, "The high computational cost not only restricts the widespread application and promotion of FMs but also affects their efficient deployment and sustainable operation." The sentence structure is complex and can be split into two shorter sentences.
2. Spelling errors in words, such as "agriculture" being written as "argiculture" in keywords.

Author Response

REVIEWER 2:

1) Insufficient connection between sections and lack of logical

Authors’ response:
&nbspWe have revised and appropriately expanded Sections 5.1 and 5.2 to enhance the logical flow and strengthen the connections between these sections, resulting in more natural transitions and improved coherence:

Future AFMs should expand beyond traditional text and image modalities to harness emerging technological capabilities.  This multimodal expansion  forms the technical foundation for cross-domain integration across agricultural supply chains. Recent studies demonstrate the untapped potential of multimodal integration in agriculture: video analytics enable real-time crop monitoring during critical growth stages[186][187].

This technological convergence  directly addresses supply chain fragmentation challenges. By fusing complementary data streams—from spectral signatures to temporal growth patterns—multimodal AFMs can synchronize production data with downstream processing parameters (e.g., matching harvest quality metrics to storage facility conditions). The resulting  closed-loop intelligence system  not only optimizes field-level production efficiency but also enhances the resilience of agricultural supply networks through data-driven coordination[191].

 

Recent research demonstrates that multimodal fusion techniques are critical for handling heterogeneous agricultural data. For instance, the Rice-Fusion framework combines CNN-processed leaf images (200×200×3 RGB) with MLP-analyzed agro-meteorological data (temperature, humidity, NPK values) through feature-level concatenation. This early fusion approach aligns temporal sensor readings (minute-level) with daily drone images via timestamp synchronization, achieving 95.31\% disease diagnosis accuracy—a 12.8\% improvement over unimodal models. The framework further addresses modality disparities by normalizing sensor data (Min-Max scaling) and image pixels (0-1 range) before fusion, ensuring compatibility between numerical and visual features[193].

 

186. Tzachor, A.; Devare, M.; King, B.; Avin, S.; Ó hÉigeartaigh, S. Responsible artificial intelligence in agriculture requires systemic 1167

understanding of risks and externalities. Nature Machine Intelligence 2022, 4, 104–109. 1168

187. Chandrasiri, C.; Kiridena, S.; Dharmapriya, S.; Kulatunga, A.K. Adoption of Multi-Modal Transportation for Configuring 1169

Sustainable Agri-Food Supply Chains in Constrained Environments. Sustainability 2024, 16, 7601. 1170

191. Ren, Y.; Huang, X.; Aheto, J.H.; Wang, C.; Ernest, B.; Tian, X.; He, P.; Chang, X.; Wang, C. Application of volatile and spectral 1177

profiling together with multimode data fusion strategy for the discrimination of preserved eggs. Food chemistry 2021, 343, 128515. 1178

193. Patil, R.R.; Kumar, S. Rice-fusion: A multimodality data fusion framework for rice disease diagnosis. IEEE access 2022, 1181

10, 5207–5222.

2)Unclear correspondence between figures/tables and the main text: For example, "Chains-BERT"

Authors’ response:
&nbspWe have supplemented detailed explanations of the mentioned examples in their respective sections:

The model achieves this through an innovative dual-encoder architecture, featuring a fixed feature extractor and trainable model that employs contrastive learning to optimize sentence representations. By augmenting this framework with Bi-LSTM layers for label-chain modeling and incorporating semi-supervised techniques like proxy labeling and consistency regularization, Chains-BERT effectively leverages both labeled and unlabeled agricultural data.Experimental results demonstrate its superior performance, achieving an 86.5 Micro-F1 score on the CAIL2018 dataset—a 5.5-point improvement over vanilla BERT. The model also exhibits strong generalization capabilities, outperforming baseline BERT by 3.25 F1 points on SQuAD 2.0, validating its effectiveness beyond agricultural domains[114].

 

114. Huang, Y.; Liu, J.; Lv, C. Chains-BERT: A High-Performance Semi-Supervised and Contrastive Learning-Based Automatic 1021

Question-and-Answering Model for Agricultural Scenarios. Applied Sciences 2023, 13, 2924. 1022

3) Disconnection between challenges and application scene

Authors’ response:
&nbspWe have supplemented the corresponding subsection in Chapter 4 to establish stronger connections with the preceding content:

For agricultural robots performing time-sensitive tasks (e.g., harvesting, pest control), processing delays >200ms directly impact accuracy and yield, exacerbated by foundation models' computational demands [178].

Delay-Tolerant Networks (DTNs) address transmission delays in remote areas, as demonstrated by PotatoScanner's store-carry-forward approach . For robotics, DTN-edge computing hybrids maintain real-time responsiveness in large fields, while model pruning/quantization reduces inference latency for harvesting operations [180].

Time series analysis and dynamic modeling help accommodate agricultural data's temporal dependencies. Lightweight transformer architectures now balance accuracy and efficiency for real-time field automation [182], while continual learning maintains model relevance.

 

 

178. Mahmud, M.S.A.; Abidin, M.S.Z.; Emmanuel, A.A.; Hasan, H.S. Robotics and automation in agriculture: present and future 1151

applications. Applications of Modelling and Simulation 2020, 4, 130–140. 1152

180. Kondo, S.; Yoshimoto, N.; Nakayama, Y. Farm Monitoring System with Drones and Optical Camera Communication. Sensors 1156

2024, 24, 6146. 1157

182. Nabaei, S.H.; Zheng, Z.; Chen, D.; Heydarian, A. Multimodal Data Integration for Sustainable Indoor Gardening: Tracking 1159

Anyplant with Time Series Foundation Model. arXiv preprint arXiv:2503.21932 2025. 1160

4) Missing details on model training:

Authors’ response:
&nbspWe have supplemented the details of relevant concrete examples to enhance clarity and completeness:

AgriBERT was pre-trained from scratch using a standard Masked Language Modeling (MLM) objective, where 15\% of tokens in input sentences were randomly masked, and the model learned to predict them based on contextual information. The training corpus consisted of 46,446 food and agriculture-related journal articles (311 million tokens, 2.39 million words) rigorously cleaned to remove URLs, emails, and non-ASCII characters, supplemented by WikiText-103 and Penn Treebank datasets for linguistic diversity.

5) Lack of comparative analysis in experimental validation

Authors’ response:
&nbspWe have incorporated comparative experimental analyses of relevant case studies to more clearly demonstrate their functional roles and advancements:

For example,Large language models have demonstrated significant potential in agricultural applications, with performance varying substantially based on model architecture and augmentation techniques. In evaluations of the Certified Crop Advisor (CCA) certification exams, GPT-4 achieved a 93\% accuracy rate when enhanced with RAG, representing a 14-point improvement over its baseline performance of 79\% without retrieval augmentation. The comparative analysis revealed consistent performance hierarchies across model types, with GPT-4 outperforming GPT-3.5 (88\% with RAG, 64\% without) and substantially surpassing smaller models like Llama2-70B (81\% with RAG, 55\% without) and Llama2-13B (70\% with RAG, 47\% without) demonstrating how model scale and knowledge retrieval techniques collectively enhance agricultural question-answering capabilities. These results underscore the transformative potential of large language models in agricultural education and decision support, particularly when combining advanced architectures with domain-specific knowledge augmentation.

6)Add case analysis :

Authors’ response:
We have supplemented the analysis of specific case studies in Section 3.

Comments on the Quality of English Language:

Authors’ response:

We have appropriately modified the sentence structures to improve readability, while also conducting spell-checking and standardizing terminology throughout the text.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper systematically introduces FMs and AFMs, which can serve as a reference for researchers in related fields. However, the paper has the following specific issues:

1. Line 22, is it FM or FMs? There are many similar issues. Line 59, foundational models should be FMs.

2. Figure 1 is not reflected in the main text. The font size of the text in Table 2 is too small.

3. Increase the content of bibliometric analysis.

4. Table 1 should include the advantages and disadvantages of different models.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

REVIEWER 3:

1) Line 22

Authors’ response:
&nbspWe have appropriately modified the sentence structures to improve readability, while also conducting spell-checking and standardizing terminology throughout the text.

2) Figure 1 is not reflected in the main text. The font size of the text in Table 2 is too small.

Authors’ response:
&nbspAfter careful consideration, we concluded that Figure 1 contributes minimally to the manuscript's overall value. Therefore, we have decided to remove Figure 1 and have concurrently optimized the font style in Table 2 for better clarity.

3) Increase the content of bibliometric analysis.

Authors’ response:
&nbspWe have incorporated bibliometric analysis in the introduction section, thereby enhancing the rationality and academic rigor of this study.

4) Table 1 should include the advantages and disadvantages of different models.

Authors’ response:
&nbspTables 1 and 2 were originally part of a single table that was split due to space constraints, but the separation caused misalignment issues that could lead to misinterpretations. These tables have now been properly adjusted and aligned. It should be noted that these tables primarily serve to introduce currently popular foundation models as a precursor to discussing agricultural foundation models. Adding detailed advantages and disadvantages for each model would make the tables unnecessarily redundant and less readable.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have made the revisions as suggested. I have no further comments.