Technical Summary: AI in Agriculture: A Survey of Deep Learning Techniques for Crops, Fisheries and Livestock

Problem Statement

Global food production, anchored by crops, fisheries, and livestock, faces escalating challenges driven by climate variability, resource limitations, and the necessity for sustainable management. While the agricultural sector has evolved through four distinct eras—from manual labor (Agriculture 1.0) to precision farming via IoT and GPS (Agriculture 3.0)—the current era (Agriculture 4.0) demands the integration of Artificial Intelligence (AI) to optimize efficiency and resilience.

Despite the proliferation of AI applications, existing literature suffers from fragmentation. Previous surveys typically focus on specific sectors (e.g., only crops or only fisheries) or are limited to conventional machine learning and early deep learning architectures like Convolutional Neural Networks (CNNs). There is a critical gap in comprehensive reviews that encompass the entire agricultural spectrum (crops, fisheries, and livestock) while addressing the recent emergence of Vision Transformers (ViTs) and Vision-Language Foundation Models (VLMs). Furthermore, the unique challenges of each domain—such as underwater imaging constraints in fisheries, environmental variability in crops, and data scarcity in livestock—require a unified yet nuanced analysis.

Methodology

This paper presents a systematic survey of over 200 research works, categorizing them across three primary domains: Crops, Fisheries, and Livestock. The review spans a methodological evolution from Conventional Machine Learning (CM) to advanced Deep Learning (DL) and Foundation Models (FMs).

The survey structures its analysis around:

1. Task Taxonomy: Defining core computer vision tasks applicable to each domain, including Image Classification (IC), Object Detection (OD), Image Segmentation (IS), Object Tracking (OT), and Anomaly Detection (AD).
2. Methodological Progression:
   • Conventional ML: Analysis of Support Vector Machines (SVMs), k-Nearest Neighbors (kNN), Decision Trees, K-Means clustering, and Case-Based Reasoning (CBR).
   • CNNs: Review of architectures like VGG, ResNet, Mask R-CNN, and YOLO variants tailored for agricultural imagery.
   • Vision Transformers (ViTs): Examination of models utilizing self-attention mechanisms (e.g., ViT, Swin Transformer, DETR variants) for capturing long-range dependencies and complex spatial relationships.
   • Foundation Models: Investigation of Vision-Language Models (VLMs) and generative models (e.g., CLIP, SAM, GPT variants) adapted for agriculture, including domain-specific pre-training and zero-shot capabilities.
3. Dataset and Challenge Analysis: A detailed mapping of publicly available datasets (e.g., PlantVillage, DeepFish, CID) and a critical discussion of domain-specific challenges such as data quality, environmental variability, and infrastructure limitations.

Key Contributions

The paper makes several distinct contributions to the field of agricultural AI:

• Comprehensive Scope: It is the first survey to simultaneously cover crops, fisheries, and livestock, providing a holistic view of AI applications across the entire food production chain.
• Inclusion of Emerging Architectures: Unlike prior works, this survey extensively covers the transition from CNNs to Vision Transformers and, crucially, the integration of Vision-Language Foundation Models (e.g., AgriCLIP, MarineGPT, AgroGPT) which enable multi-modal understanding and zero-shot learning.
• Detailed Taxonomy and Visualization: The authors provide a structured taxonomy (Fig. 3) and statistical distribution of papers (Fig. 1), illustrating the shift in research focus toward transformers and foundation models across different agricultural sectors.
• Domain-Specific Challenge Synthesis: The paper systematically categorizes challenges unique to each sector:
  • Crops: Environmental variability, diverse growth stages, and data annotation costs.
  • Fisheries: Underwater imaging constraints (turbidity, light distortion), data volume complexity, and distinguishing rare events from routine fluctuations.
  • Livestock: Data scarcity, class imbalance, noisy data from moving animals, and the need for real-time inference on edge devices.
• Resource Curation: The authors maintain a publicly accessible GitHub repository (https://github.com/umair1221/AI-in-Agriculture) to track the latest literature, datasets, and trends, serving as a living resource for researchers.

Results and Findings

The survey synthesizes findings from over 200 papers, revealing several key trends and performance benchmarks:

• Performance Evolution: While Conventional ML methods (SVM, kNN) remain viable for small datasets or resource-constrained edge devices, Deep Learning (CNNs) has become the standard for high-accuracy tasks. Recent studies show ViTs and Foundation Models outperforming traditional CNNs in tasks requiring fine-grained discrimination and handling of complex, unstructured data.
• Foundation Model Efficacy: Domain-adapted foundation models demonstrate significant improvements in zero-shot and few-shot scenarios. For instance, AgriCLIP showed a 24.5% accuracy gain over generic CLIP in fisheries classification, and BioCLIP achieved a 17% absolute improvement in zero-shot biological classification by leveraging taxonomic hierarchies.
• Task-Specific Success:
  • Crops: Models like VGG-ICNN and PlantViT have achieved accuracies exceeding 98% in disease detection.
  • Fisheries: AquaticCLIP and MarineGPT have demonstrated robustness in underwater environments, achieving up to 96.8% accuracy in zero-shot classification tasks.
  • Livestock: AnimalFormer and StarFormer have successfully integrated segmentation, tracking, and action recognition, enabling non-invasive monitoring of animal behavior and health.
• Dataset Landscape: The survey highlights a disparity in dataset availability. While crop datasets (e.g., PlantVillage) are abundant, livestock and fisheries datasets often suffer from class imbalance, limited scale, or lack of standardized benchmarks, hindering the development of generalizable models.

Significance and Future Directions

The paper asserts that the rapid evolution from conventional methods to domain-tuned foundation models is re-energizing digital agriculture. Its significance lies in providing a roadmap for researchers to navigate the complexities of multi-domain agricultural AI.

The authors identify five critical future research directions:

1. Unifying Multi-modal Foundation Models: Developing models that natively process heterogeneous data streams (satellite imagery, sensor logs, text, audio) to create transferable embeddings for cross-modal pattern recognition.
2. Continual and Federated Adaptation: Implementing mechanisms for models to learn continuously from edge devices without catastrophic forgetting, and utilizing federated learning to improve shared models while preserving data privacy.
3. Edge-Efficient AI: Optimizing models (e.g., MobileNetV4, MobileVLM) for deployment on low-power devices (drones, underwater robots) to enable real-time decision-making in areas with limited connectivity.
4. Agentic AI for Autonomous Farming: Moving beyond passive feature extraction to proactive problem-solving, where AI agents can formulate hypotheses, plan actions (e.g., scheduling drone surveys), and execute corrective measures autonomously.
5. Integrated Systems: The future of agriculture lies in fully integrated, multi-domain intelligent systems that merge visual, environmental, and textual data to support the entire food production chain, from crop yield prediction to livestock welfare monitoring.

The paper concludes that while the journey toward fully automated and sustainable agriculture is ongoing, the convergence of multimodal AI, edge computing, and domain-adaptable foundation models offers a promising path to address global food security challenges.