Deep Learning in Farming: A Systematic Evidence-Weighted Review of Applications, Validation Gaps, and Emerging Frontiers

1. Introduction

1.1. The Global Imperative for Agricultural Innovation

Agriculture is a central pillar of the worldwide food supply, yet it faces mounting, compounding pressures: population growth, climatic instability, resource scarcity, labour constraints, and the need to lower environmental externalities. Recent international assessments stress that agrifood systems must become more productive, resilient, inclusive and sustainable, so as to meet food demand while cutting social and environmental costs [1,2,3]. Rural depopulation and an ageing farming workforce compound these pressures further, making labour shortages especially acute for critical, time-sensitive operations such as harvesting, disease inspection, animal monitoring, and irrigation management [3,4,5]. DL-based automation, decision support, and robotic perception respond directly to this combination of demands: raising productivity while sustaining farm operational capacity under conditions of reduced and ageing rural labour.

1.2. Agriculture 4.0 and Deep Learning as Its Foundational Technology

To address these challenges, agriculture is undergoing a technological transformation often described as Agriculture 4.0 or Smart Agriculture, characterised by the integration of IoT, robotics, big-data analytics, and AI across the agricultural value chain [6,7,8]. Large-scale sensor deployments across fields, machinery, drones, and satellites generate heterogeneous data streams, including imagery, soil measurements, moisture data, and weather records [9,10], with the goal not merely of automating existing practices but of using these streams to support more precise interventions, better resource allocation, higher productivity, and a lower environmental impact [6,7].

Deep Learning (DL), a branch of machine learning grounded in multi-layer neural networks, has become the main technology enabling this shift [6,11], owing to its capacity to learn hierarchical representations directly from large amounts of raw, often unstructured data [12,13], precisely the kind of heterogeneous, multi-scale data that cameras, drones, satellites, animal sensors, and soil probes produce [9,11]. Unlike classical machine-learning pipelines, which rely heavily on hand-crafted features, deep neural networks learn task-relevant representations from images, time series, and multimodal records with reduced manual feature engineering [14,15]. This capacity is particularly valuable in farming, where the same biological target appears differently across crops, cultivars, growth stages, lighting conditions, soil backgrounds, disease severities, and management regimes.

High benchmark performance should not, however, be read as automatic readiness for farm deployment: models trained on curated datasets routinely overfit to narrow imaging conditions or local management practices. For this reason, the present review treats DL not as a catalogue of architectures, but as a family of methods whose value rests on validation realism, transferability, interpretability, and operational feasibility, evaluated jointly with the agronomic and socio-technical conditions in which it is meant to operate.

1.3. Scope and Structure of the Review

This review provides a systematic and analytical synthesis of deep learning in agriculture, with an emphasis on literature published between January 2020 and September 2025. The term farming is used here to delimit the review to the production-oriented part of agriculture, including crop production, livestock production, soil and water management, and autonomous field or barn operations. Broader agri-food applications such as plant breeding, biotechnology, microbiology, seed technology, industrial food processing, post-harvest packaging, retail logistics, and traceability are acknowledged as adjacent areas but are not treated as the principal object of this review, so as to preserve analytical depth and avoid merging operational farming problems with biologically and commercially distinct stages of the agri-food chain. Its objectives are to identify the main areas of progress, highlight methodological and deployment-related weaknesses, and outline the most promising directions for future research. Section 2 presents the review methodology. Section 3 synthesises the literature across four application domains and highlights cross-domain methodological patterns. Section 4 discusses emergent technical directions that may help address current limitations. Section 5 summarises the main conclusions, and Section 6 provides a roadmap for future research and practice.

1.4. Positioning with Respect to Existing Surveys and Systematic Reviews

Multiple surveys have investigated the role of artificial intelligence and deep learning in agriculture, from varied conceptual frameworks. Ref. [6] produced one of the earliest structured syntheses of deep learning in agriculture, concentrating on convolutional architectures and vision tasks documented through approximately 2017. Ref. [11] addressed farm automation using both machine learning and deep learning, emphasising robotic and sensor-equipped infrastructure, while Ref. [7] reviewed machine-learning methods for precision agriculture more broadly, positioning deep learning as one algorithmic family among others. Ref. [16] focused specifically on deep learning and agricultural advancement, identifying constraints and potential developments. Adjacent contributions have examined complementary parts of the agri-food digitalisation workflow: Ref. [17] surveyed machine-learning applications spanning yield modelling, disease identification, and supply-chain logistics; Ref. [18] examined how machine learning and, more narrowly, deep learning support sustainable supply-chain performance; and specialised reviews have covered deep learning for dense visual scene interpretation [19], satellite and airborne image classification [9,20], and mask-based segmentation for farming imagery [10]. Recently, there have also been a number of survey papers that emerged: Ref. [21] provided a survey of 129 papers in five domains of DL application in agriculture but without the realism constraint of deployment; Ref. [22] conducted a systematic review of machine learning and computer vision in precision agriculture focusing on innovative methods; Ref. [23] employed a bibliometric and keyword network approach to deep learning in crop monitoring; and Ref. [24] provided a systematic review of deep learning for lettuce farming alone. The table below contrasts this review with the aforementioned reviews.

The present review departs from these earlier contributions along four lines: it centres on the 2020–September 2025 window, during which agricultural DL shifted from mostly CNN-based classification toward object detection, segmentation, transformer architectures, self-supervised learning, edge inference, federated learning, and hybrid PBM–DL systems; it follows a PRISMA 2020-oriented protocol with explicit search blocks, screening counts, and a common critical-appraisal framework (Section 2); it tightens the object of analysis onto farming specifically, rather than the whole agri-food chain; and it weights the evidence explicitly, reading reported model performance through dataset representativeness, validation protocol, benchmarking transparency, deployment realism, and reproducibility, so as to separate studies showing algorithmic promise from studies giving firmer evidence of field relevance.

1.5. Main Contributions of This Review

This review makes four main contributions. First, it applies a transparent review methodology that allows readers to assess the strength of evidence of individual studies rather than treating all reported results as equivalent. Second, it synthesises the literature across four agricultural domains using common comparative dimensions. Third, it explicitly addresses the laboratory-to-field gap together with benchmarking, socio-economic, and governance challenges. Fourth, it provides a near-term research agenda, explicitly ordered by the strength of evidence currently available for each direction (Section 5.2), focused on realistic benchmarks, multimodal and self-supervised learning, system-level evaluation, and techno-economic assessment.

1.6. Analytical Framework and Originality of the Review

The originality of this review does not rest on a claim that agricultural deep learning has remained unreviewed. Section 1.4 and Table 1 show clearly that it has not. Its contribution is rather to re-read the recent literature through an evidence-weighted and deployment-oriented framework, interpreting the empirical studies discussed along five analytical dimensions. Data realism points to the degree to which the dataset mirrors field variability.Validation strength points to the presence of external, temporal or cross-site testing. Model–task suitability asks whether the chosen architecture fits the structure of the agricultural problem. Deployment readiness looks for evidence of latency, edge feasibility, robustness and system integration. Practical value weighs agronomic interpretability, economic feasibility and usefulness to farmers or agricultural professionals. These five dimensions are not merely descriptive: Section 2 operationalises them directly as the risk-of-bias rubric in Table 2, so the same criteria introduced here are the ones actually applied to every included study. Through this framework, the review steps past a descriptive list of applications, separating studies showing narrow benchmark success from those offering sturdier evidence of operational relevance, and judging the field’s recent growth in CNN-based vision, object detection, segmentation, transformers, self-supervised learning, federated learning, edge AI, and process-informed hybrid models against one practical question: are the reported systems reproducible, robust, interpretable, and economically plausible inside real farming environments?

Table 1. Positioning of the present review relative to representative prior surveys of deep learning in agriculture.

Table 2. Risk-of-bias/critical-appraisal rubric applied to the 56 included studies, operationalising the five analytical dimensions of Section 1.6. Authors’ synthesis.

2. Materials and Methods

A systematic literature search, structured around a PRISMA 2020-oriented protocol [25], was used to identify recent empirical studies and relevant contextual reviews concerning deep learning in farming. Searches covered Scopus, Web of Science, and IEEE Xplore; Google Scholar supported backward and forward citation searching. The principal publication window was 1 January 2020 to 30 September 2025. Earlier studies were consulted only when they provided foundational technical or historical context and were not treated as part of the recent empirical evidence base. Farming and agricultural-production terms were combined with descriptors of deep learning methods. Production-stage terms such as “sowing”, “harvesting”, “robotic harvesting”, “poultry”, “precision livestock farming”, and “animal welfare” were included to reduce vocabulary bias. A representative Scopus query is:

TITLE (“deep learning” OR “CNN” OR “RNN” OR “LSTM” OR “Transformer” OR “GNN” OR “self-supervised learning” OR “neural network”) AND TITLE-ABS-KEY (“agriculture” OR “crop” OR “farm” OR “livestock” OR “irrigation” OR “phenotyping” OR “robotics”) AND PUBYEAR > 2019 AND PUBYEAR < 2026 AND (LIMIT-TO (SUBJAREA, “COMP”) OR LIMIT-TO (SUBJAREA, “ENGI”) OR LIMIT-TO (SUBJAREA, “AGRI”)) AND (LIMIT-TO (DOCTYPE, “ar”) OR LIMIT-TO (DOCTYPE, “cp”))

The query was adapted to the syntax and indexing fields of each database while preserving three conceptual blocks: (i) deep-learning terminology, (ii) farming and agricultural-production terminology, and (iii) task-specific terminology. Google Scholar was not used as a primary database because its ranking and coverage are difficult to reproduce; it was instead used only for backward and forward snowballing from already-eligible papers. The final search update was performed in September 2025, and online-first material was considered only when its peer-reviewed publication status could be verified by that date. The completed PRISMA 2020 Checklist is provided in the Supplementary Materials. We intentionally opted for using traditional bibliographic databases rather than a scientific search or summarisation based on LLMs (Elicit, Consensus, and others). There were three main reasons behind our choice. First, reproducibility: Boolean search in Scopus, Web of Science, and IEEE Xplore can be repeated independently by any other researcher and will produce consistent results, while an LLM search is highly dependent on specific wording of prompts, model version used, and heuristic ranking/summarisation algorithm of the tool itself, which is all non-reproducible. Second, transparent coverage: bibliographic databases provide an explicit count of included/excluded articles at each screening step (Section 2.2), and that is what provides audibility of the PRISMA flowchart in Figure 1; no general-purpose search tool powered by LLM can provide the same kind of transparent accounting of the search process. Third, reliability: LLM-driven summarisation of scientific text comes with known risk of generating false citations or misattribution of findings - exactly the kind of mistake that a critical evidence synthesis should avoid. However, these tools are a truly valuable addition to the traditional screening process, and we will further discuss them, as well as the application of large language models in agricultural decision-making in Section 4.4 and Section 6.

Figure 1. PRISMA 2020 flow diagram summarising the identification, screening, eligibility assessment, and inclusion of studies (2020–2025) [25]. Diagram authored natively in TikZ by the authors from the underlying screening counts, so that it can be revised as plain text rather than as a raster image.

Candidate publications were checked by title, abstract and, where necessary, full text. Bibliographic metadata were verified against the cited publication rather than inferred from citation-key names. For the evidence-weighted synthesis, the extraction fields were application domain, task, sensing modality, dataset type, architecture family, comparison models, evaluation metrics, validation protocol, code or data availability, deployment status, and stated limitations.

2.1. Inclusion and Exclusion Criteria

The core empirical evidence comprised English-language, peer-reviewed journal articles and indexed conference papers published from 2020 through September 2025 that presented and evaluated a deep-learning model for a production-oriented agricultural task. Eligible tasks included crop disease and weed detection, yield or fruit estimation, livestock identification, health and welfare monitoring, soil-property or crop-water assessment, and robotic perception or control. Empirical papers had to specify a task, a dataset or acquisition setting, a model family, and quantitative or operational performance evidence.

Review articles, reports and studies published before 2020 were used only for contextual framing and were not represented as recent empirical evidence. Papers were not used as task-level evidence when they evaluated only non-deep-learning methods, concerned a non-agricultural application, or addressed supply-chain security, industrial processing, breeding or traceability without a direct connection to farming-stage sensing, prediction or action. This distinction is important because contextual sources and primary empirical studies serve different purposes in a critical review. The reasons for full-text exclusions were categorised into three categories, as depicted in Figure 1. Non-eligible content (

$n = 1250$
) includes records that did not actually involve evaluating a deep-learning model upon further checking of the full text (such as opinion pieces, theoretical works only, or classical machine-learning-based papers classified due to erroneous search terms), those that discussed topics not related to agriculture even though they matched the query, or those that were about breeding, processing, or traceability but did not have any farming stage sensing/prediction/action application. Incomplete data (

$n = 254$
) includes records where the dataset information, model information, or results were not sufficiently available to be extracted. Pu