Remote Sensing, Machine Learning, and Smart Farming: An Umbrella Review of Review-Level Evidence for Crop Monitoring, Disease Detection, and Yield Forecasting

Remote sensing and machine learning have become prominent components of smart farming research, generating an expanding body of review literature across agricultural monitoring and prediction applications. This umbrella review synthesises review-level evidence concerning crop monitoring, disease detection, yield forecasting, and decision support while examining the methodological conditions that shape evidence strength and transferability. Peer-reviewed systematic reviews, scoping reviews, structured reviews, surveys, meta-reviews, and other eligible review-level articles were identified through searches of major academic databases and publisher platforms. Evidence was organised by application domain, sensing modality, analytical method, validation design, generalisation, uncertainty, reproducibility, and decision-support relevance. Crop-monitoring reviews describe extensive use of satellite, unmanned aerial vehicle, multispectral, hyperspectral, thermal, and multimodal data for crop mapping, growth assessment, phenotyping, stress monitoring, and related applications. Disease-detection reviews report rapid expansion of image-based machine learning and deep learning, but repeatedly identify data scarcity, class imbalance, curated-image dependence, and limited cross-context validation. Yield-forecasting reviews identify recurring integration of remote sensing, weather, soil, crop-model, and management information, although forecast horizons, spatial and temporal validation, uncertainty communication, and decision relevance remain inconsistent. Across domains, reported technical promise is moderated by heterogeneous datasets, benchmark inconsistency, review overlap, incomplete reproducibility, and limited evidence of transfer across crops, regions, seasons, and acquisition conditions. Review-level evidence supports a substantial research opportunity for remote sensing and machine learning in smart farming, but it does not independently establish field readiness, diagnostic reliability, operational forecasting performance, or deployment readiness. Stronger review reporting, external validation, transparent benchmarking, uncertainty assessment, reproducible workflows, and responsible decision-support evaluation are required.

Review literature portrays smart farming as a data-intensive agricultural transformation in which machine learning and deep learning increasingly mediate heterogeneous sensor, image, and management data, while the evidentiary maturity of applications remains uneven [1–3]. Within this transformation, smart farming refers not merely to the use of isolated digital tools but to interconnected processes through which agricultural observations are acquired, transferred, analysed, interpreted, and potentially incorporated into management decisions. Machine-learning applications reviewed across agriculture include crop, soil, disease, weed, livestock, and yield-related tasks, whereas deep-learning surveys document the growing use of representation-learning methods for imagery and other high-dimensional data. These broad reviews establish the technical diversity of the field while also revealing substantial differences in datasets, analytical objectives, evaluation practices, and application contexts.

Remote sensing constitutes one of the most extensive evidence domains within this wider literature. A meta-review of agricultural remote sensing identified a broad range of sensing platforms, retrieved crop and environmental variables, processing approaches, and intended applications, demonstrating that the literature encompasses far more than a single technology or monitoring task [4]. Satellite, airborne, unmanned aerial vehicle, proximal, multispectral, hyperspectral, thermal, and structural sensing approaches have been reviewed in relation to crop identification, growth assessment, stress detection, phenotyping, resource monitoring, and yield-related applications. Nevertheless, the breadth of reviewed applications should not be interpreted as proof that the underlying systems produce reliable decisions across farms, crops, seasons, or regions. A meta-review maps recurring evidence patterns across earlier reviews; it does not independently verify the validity of every included primary study or establish operational reliability.

An umbrella-review perspective is therefore needed because the evidence base is increasingly mediated by multiple reviews with overlapping but non-identical scopes. Some reviews organise technologies according to sensor platform, whereas others focus on machine-learning families, crop traits, disease recognition, yield prediction, data infrastructures, or decision-support systems. Reviews also differ in whether they examine controlled-image classification, field-scale remote sensing, temporal forecasting, crop-model assimilation, socioeconomic conditions, or responsible deployment. Consequently, similar terms may refer to materially different research tasks, and apparently consistent conclusions may be based on overlapping primary studies, distinct validation designs, or unequal levels of methodological scrutiny. Treating all review conclusions as interchangeable would obscure these differences and could exaggerate the density or maturity of the evidence.

The objective of this umbrella review is to synthesise review-level evidence on remote sensing, machine learning, and smart farming across crop monitoring, disease detection, yield forecasting, and decision-support applications. Its contribution is threefold: first, it maps convergent and divergent findings across review domains; second, it evaluates recurring methodological limitations involving data quality, benchmarking, validation, generalisation, uncertainty, reproducibility, and review overlap; and third, it develops a bounded future research agenda connecting technical evaluation with decision-support translation and responsible deployment. The synthesis is deliberately interpretive rather than prescriptive. It assesses what the review literature supports, where evidence remains fragmented, and which claims require direct primary-study validation before field-readiness or deployment-readiness conclusions can be considered.

Review-level evidence presents smart farming and digital agriculture as sociotechnical developments rather than exclusively computational ones. Social-science reviews show that the meaning and consequences of digital agriculture depend on the roles of farmers, advisers, technology firms, data intermediaries, public institutions, and other actors, as well as on changes in labour, skills, knowledge, power, and governance [5]. This perspective qualifies purely technical accounts of remote sensing or predictive modelling. Even where sensing and analytical capabilities appear promising, their decision value may be influenced by data access, institutional arrangements, infrastructure, accountability, and the capacity of users to evaluate or act upon model outputs. Review-level synthesis must therefore distinguish evidence of technical capability from evidence concerning organisational integration, acceptance, equitable access, or practical use.

Crop-data-management reviews offer a complementary systems perspective by connecting sensing, data transmission, storage, analytics, interoperability, and management interfaces. Within this literature, data generated by field sensors, machinery, imagery, farm records, and external information services can be treated as components of an agricultural information architecture rather than independent inputs [6]. Such architectures may support monitoring and decision processes, but the review literature also identifies recurring challenges concerning heterogeneous formats, data ownership, connectivity, platform compatibility, temporal synchronisation, and the conversion of analytical outputs into interpretable actions. Consequently, the existence of multiple agricultural data streams does not itself establish effective data integration or useful decision support.

Updated machine-learning reviews indicate continued expansion in the number and variety of agricultural applications, including crop management, soil assessment, disease and pest recognition, weed detection, water management, livestock monitoring, and yield prediction [7]. The reviewed literature includes conventional statistical learning, random forests, support vector machines, neural networks, convolutional architectures, and other data-driven approaches. However, the apparent expansion of application categories is accompanied by substantial heterogeneity in sample construction, sensor conditions, preprocessing, feature selection, model tuning, comparison procedures, and validation. Review-level evidence can therefore identify recurring analytical approaches and application concentrations, but it rarely supports a context-independent ranking of algorithms or a general conclusion that one method is universally superior.

Precision-agriculture reviews similarly map machine-learning applications across soil, crop, disease, weed, irrigation, and yield domains while exposing uneven benchmarking and evaluation practices [8]. Differences in target variables, input resolutions, class definitions, spatial units, prediction horizons, and performance metrics limit direct comparison across reviewed studies. Review articles may also combine laboratory imagery, experimental plots, research farms, satellite products, and operational datasets within the same application category, despite their distinct validity conditions. The central review-level pattern is thus simultaneous growth and fragmentation: smart farming research is expanding across sensing and analytical domains, but methodological diversity, limited external validation, and inconsistent reporting constrain evidence integration. An umbrella synthesis is needed to preserve domain-specific distinctions while identifying limitations that recur across otherwise separate review literatures.

The umbrella review was organised around six review-level questions addressing the overall evidence landscape, crop monitoring, disease and related detection tasks, yield forecasting and decision support, recurring methodological limitations, and future research priorities. An overview of reviews requires an explicit review-level question, a clearly bounded objective, and a defined unit of synthesis so that conclusions concern the findings of eligible reviews rather than an unstructured mixture of reviews and primary studies [9]. Accordingly, the unit of evidence was the peer-reviewed review article. Eligible sources included systematic reviews, scoping reviews, meta-reviews, structured critical reviews, surveys, integrative reviews, and bibliometric reviews only where they contributed substantive evidence mapping rather than publication counts alone.

Candidate sources were identified through targeted searches of ScienceDirect, SpringerLink and BMC journals, IEEE Xplore, Wiley Online Library, Taylor & Francis Online, the MDPI journal platform, and PubMed. Search concepts combined terms for umbrella reviews and review-level synthesis with smart farming, precision agriculture, digital agriculture, remote sensing, machine learning, deep learning, crop monitoring, plant disease detection, pest and weed detection, yield prediction, yield estimation, forecasting, crop models, data fusion, validation, generalisation, uncertainty, reproducibility, and decision support. Overview methods should be chosen transparently in relation to the purpose and structure of the evidence base [10]. The present synthesis therefore used structured qualitative comparison rather than a meta-analysis of primary-study effect sizes because the included reviews differed substantially in questions, application domains, study designs, outcomes, and reporting formats.

Eligibility assessment considered publication type, peer-review status, publication period, DOI availability, review design, agricultural relevance, and capacity to support at least one predefined umbrella-review domain. Primary model-development studies, field experiments, conference-only papers, preprints, books, theses, product materials, software documentation, vendor pages, standalone datasets, reports, standards, and other non-journal sources were excluded. Bibliometric articles without substantive evidence synthesis were not used as application evidence. Evidence extraction covered review aims, agricultural contexts, sensor modalities, analytical methods, application domains, reported evidence patterns, validation limitations, generalisation concerns, uncertainty, reproducibility, decision-support relevance, and stated research gaps. Because risk of bias and methodological limitations can affect conclusions at both review and primary-study levels, findings were not aggregated as though every included review provided evidence of equivalent strength [11].

Screening proceeded through bibliographic checking, title and abstract assessment, full-text eligibility review, confirmation of review type, evidence-domain mapping, and readiness for structured extraction. Particular attention was given to primary-study overlap because repeated inclusion of the same underlying studies across reviews can make an evidence domain appear denser without adding independent evidence [12]. Overlap awareness was incorporated qualitatively by comparing review scope, application focus, publication period, and frequently recurring evidence themes; no numerical overlap estimate or formal quality score is claimed. The resulting synthesis maps review-level patterns and limitations but does not independently validate primary datasets, sensor performance, model accuracy, diagnostic reliability, forecast reliability, farmer acceptance, or deployment outcomes. Table 1 summarises the umbrella review scope, review questions, eligibility criteria, evidence sources, and review-level synthesis domains.

Review-level evidence maps satellite, airborne, and UAV remote sensing to crop-status assessment, stress surveillance, input-management support, and high-resolution field observation, but repeatedly conditions these opportunities on sensor calibration, preprocessing, scale, and validation design [13–15]. Across these reviews, remote sensing is used as an umbrella term for observations differing substantially in platform, spatial resolution, spectral resolution, temporal frequency, viewing geometry, atmospheric exposure, and processing requirements. Satellite systems provide repeated regional or landscape observations, whereas UAV platforms can provide finer spatial detail and flexible acquisition timing over smaller areas. Multispectral, hyperspectral, RGB, and thermal sensors are reviewed in relation to crop mapping, growth assessment, canopy condition, soil variability, water stress, disease or weed surveillance, and yield-related traits. These uses indicate a broad research base, but differences in acquisition and processing prevent a simple conclusion that one platform is uniformly preferable.

Field-based phenotyping illustrates both the analytical potential and the interpretive limits of high-resolution remote sensing. UAV reviews describe extraction of canopy cover, plant height, vegetation indices, temperature-related traits, biomass proxies, lodging characteristics, flowering information, and other phenotypic features [16]. Such measurements may increase temporal and spatial detail relative to manual observation, particularly where repeated acquisition is feasible. However, the reviewed evidence also identifies limitations involving sensor calibration, illumination, wind, geometric correction, plot delineation, occlusion, trait-reference measurements, varietal diversity, and scaling from research plots to heterogeneous fields. Reported associations between image-derived features and crop traits are therefore context-dependent. Phenotyping evidence supports continued evaluation of sensing and feature-extraction methods but does not establish that a model or trait-retrieval procedure will transfer across crop varieties, growth stages, environments, sensors, or seasons.

Reviews of remote-sensing contributions to crop models show a complementary relationship between spatial observation and process-based representation. Remote sensing can provide information on crop extent, vegetation condition, canopy development, leaf area, biomass-related variables, soil moisture, or other states that may initialise, calibrate, constrain, or update crop models [17]. Data assimilation and related integration methods can potentially reconcile observations with simulated crop processes, thereby supporting monitoring or yield estimation at spatial scales that field measurements alone may not cover. Nevertheless, the literature repeatedly identifies mismatches among observation timing, model state variables, spatial support, parameterisation, crop calendars, cloud contamination, and measurement uncertainty. Integration is therefore not inherently superior to either component used separately; its value depends on whether observation and model uncertainties are characterised and whether assimilation is evaluated under conditions distinct from those used for calibration.

Multimodal crop-monitoring reviews extend this logic by examining combinations of satellite, UAV, ground, spectral, thermal, structural, weather, and other data streams. Combining modalities may compensate for limitations of a single sensor, improve temporal coverage, or represent complementary crop properties, but the evidence remains methodologically heterogeneous [18]. Fusion may occur at raw-data, feature, decision, or model levels, with materially different requirements for registration, temporal alignment, missing-data treatment, scale harmonisation, and computational design. Review-level evidence suggests that multimodal monitoring is an important research direction, yet comparison is constrained by limited benchmark transparency and by inconsistent descriptions of preprocessing and fusion architectures. More inputs may also introduce additional sources of error, bias, or redundancy. Accordingly, multimodal complexity should not be treated as evidence of improved generalisation unless evaluated against transparent single-modality baselines and independent contexts.

Hyperspectral evidence illustrates the distinction between sensitive signal detection and validated field interpretation. Reviews report that narrow, contiguous spectral bands can support research on early plant stress and disease discrimination, including feature selection and classification of spectral responses that may precede clearly visible symptoms [19]. Yet spectral signatures can vary with crop species, cultivar, disease stage, water status, nutrient condition, canopy structure, illumination, background, atmosphere, sensor configuration, and acquisition geometry. Similar stressors may also produce overlapping spectral responses, making causal or diagnostic interpretation difficult without appropriate reference measurements. Across remote-sensing domains, repeated limitations include small or context-specific datasets, inconsistent ground truth, spatial and temporal dependence between training and testing data, incomplete uncertainty reporting, and limited external validation. The umbrella synthesis therefore treats remote sensing as a substantial crop-monitoring evidence base while maintaining a strict boundary between research capability and field readiness. Table 2 summarises review-level evidence for remote sensing and crop monitoring in smart farming.

Across reviews of image-based plant disease recognition, convolutional deep-learning pipelines dominate the reported literature, yet performance is strongly shaped by dataset construction, background conditions, class balance, and validation choices [20–22]. The evidence base is concentrated in RGB photographs of leaves or plant organs, frequently collected under controlled or semi-controlled conditions and organised as classification tasks. Convolutional neural networks and transfer-learning approaches are prominent because they can learn image features directly, reducing dependence on manually designed descriptors. However, reviews also describe substantial variation in disease classes, image acquisition, preprocessing, augmentation, train–test partitioning, and reported metrics. Consequently, apparently high within-dataset performance cannot be interpreted as equivalent evidence of reliable disease recognition under field conditions, where symptoms may be subtle, overlapping, partially occluded, or confounded by nutrient, water, weather, and pest stresses.

The disease-and-pest detection literature extends beyond image-level classification to object detection, lesion localisation, segmentation, severity estimation, and canopy-level surveillance. Deep-learning reviews identify growing methodological diversity, including region-based detectors, one-stage detectors, semantic segmentation, transfer learning, and combinations of learned and handcrafted features [23]. This expansion is important because practical monitoring may require locating affected tissue or plants rather than assigning one label to a pre-cropped image. Nevertheless, task expansion does not remove underlying validity concerns. Small or duplicated datasets, inconsistent symptom definitions, non-independent images from the same plant or acquisition event, and limited reporting of negative or mixed-stress examples can inflate apparent model performance. Cross-crop, cross-disease, cross-region, and cross-season transfer are discussed more often than they are demonstrated through independent testing.

Automated insect-pest detection and monitoring reviews describe heterogeneous systems based on trap images, close-range photographs, field imagery, and other sensing arrangements [24]. The reviewed approaches include conventional image processing, feature-based classification, convolutional networks, and object-detection methods intended to distinguish species, count individuals, or support monitoring. Their reported strengths lie in the possibility of repeated observation and reduction of manual image inspection. Yet field comparison is difficult because pest appearance varies by life stage, posture, damage, image scale, and environmental background, while rare species and low-prevalence classes complicate evaluation. Annotation quality, overlapping insects, debris, illumination, trap design, and inconsistent performance metrics further restrict comparability. Review-level evidence therefore supports continued automated-monitoring research but does not establish autonomous pest diagnosis or pest-management decisions.

Weed-detection reviews show similarly strong computer-vision activity across ground-based and aerial RGB or multispectral imagery [25]. Classification, object detection, and semantic segmentation are used to separate weeds from crops or background, with convolutional models increasingly replacing or supplementing handcrafted spectral, colour, texture, and shape features. Reported research strengths include fine-scale localisation and the potential to characterise spatial weed occurrence. Repeated limitations include crop–weed visual similarity, changing growth stages, canopy overlap, severe class imbalance, inconsistent weed taxonomies, and narrow field or regional coverage. Dataset and task definitions also differ according to whether the objective is species recognition, crop–weed separation, patch mapping, or plant-level localisation. Detection evidence alone cannot establish safe actuation, herbicide selection, or autonomous field intervention because downstream decision consequences, error tolerances, and failure handling are rarely evaluated by the reviewed literature.

Across disease, pest, and weed reviews, dataset properties remain a central determinant of reported performance. Disease-recognition outcomes are sensitive to sample size, class distribution, symptom similarity, image background, acquisition quality, and the relation between training and testing samples [26]. Augmentation and transfer learning may reduce some computational and sample constraints, but synthetic variation does not necessarily reproduce new cultivars, field backgrounds, weather conditions, sensor settings, disease stages, or mixed stresses. Stronger evidence would require provenance-rich datasets, per-class reporting, geographically and temporally separated validation, and explicit testing under realistic field variability. The umbrella synthesis therefore characterises these systems as recognition and monitoring research. It does not infer diagnostic readiness, cross-context generalisation, farmer approval, operational reliability, or deployment readiness from review-level model results.

Review-level evidence on yield prediction and nitrogen-status estimation shows sustained use of machine learning to relate crop outcomes to weather, soil, management, spectral, and crop-state variables [27]. Random forests, support vector machines, artificial neural networks, regression methods, and related models recur across the reviewed literature, but the analytical target is not always defined consistently. Some studies estimate yield near harvest from observed crop traits, whereas others attempt forecasts at earlier decision-relevant stages. This distinction matters because information available close to harvest may support accurate estimation without demonstrating useful forecasting at longer lead times. Differences in crop type, spatial unit, season, input availability, forecast horizon, and evaluation metric also constrain cross-study comparison. Review-level evidence consequently supports a broad methodological opportunity but not a universal claim of reliable yield forecasting.

Systematic review evidence identifies weather, soil, management, and remote-sensing variables as recurring inputs to crop-yield prediction, with tree ensembles and neural approaches frequently represented [28]. Weather inputs can include rainfall, temperature, radiation, humidity, or derived agroclimatic indicators; soil inputs may include texture, moisture, fertility, or mapped attributes; and management inputs may include planting date, cultivar, irrigation, or fertilisation. Their presence in a model does not guarantee that they were measured at the relevant spatial depth, resolution, or time, nor that they would be available when a real forecast is issued. Inconsistent datasets, missing management records, static soil layers, differing yield definitions, and non-comparable metrics limit conclusions about algorithm superiority. Spatially or temporally dependent train–test partitions may also exaggerate performance by allowing closely related observations to appear on both sides of the evaluation.

Reviews focused on integrating agrarian factors reinforce the importance of contextual completeness. Yield is shaped by interacting weather, soil, genotype, crop condition, and management processes, making data fusion conceptually attractive but methodologically demanding [29]. Tabular machine learning, ensemble methods, deep architectures, and hybrid approaches can combine these sources, yet missingness, measurement error, proprietary farm records, inconsistent coding, and scale mismatch may reduce transportability. Temporal modelling is particularly vulnerable to leakage when predictors include information that would not have been known at the stated forecast date. Spatial modelling is similarly vulnerable when neighbouring fields, pixels, or administrative units are split randomly. Forecast evaluation should therefore preserve true temporal cut-offs and geographic independence while reporting data provenance and availability at prediction time.

Remote-sensing assimilation into crop models offers a distinct pathway for combining spatial observation with process knowledge. Reviews describe the use of remotely observed crop states to initialise, calibrate, constrain, or update process-based simulations, potentially linking canopy or vegetation information with weather, soil, cultivar, and management representations [30]. Such integration may improve consistency between observed and simulated crop development, but it also introduces uncertainty from sensor retrievals, crop-model structure, parameterisation, spatial support, observation timing, and assimilation design. A more complex hybrid model is not automatically a better forecast. Comparative evidence should distinguish observation uncertainty from process-model uncertainty, evaluate whether added data improve out-of-sample performance, and report sensitivity to assumptions and missing observations.

Deep-learning and remote-sensing yield reviews document expanding use of convolutional, recurrent, and hybrid architectures with satellite, UAV, weather, and temporal crop data [31]. Their research value lies in learning spatial or sequential representations that may be difficult to specify manually. However, reviewed comparisons remain limited by crop- and region-specific datasets, short time series, heterogeneous baselines, incomplete reporting, and sparse external validation across seasons and locations. Point-error metrics also provide little information about forecast calibration or decision risk unless accompanied by intervals, probabilistic scores, or scenario analysis. Decision-support claims therefore require evidence that forecasts are available at a meaningful lead time, interpretable within a farm workflow, and evaluated for uncertainty and downstream consequences. Table 3 maps review-level evidence for machine learning in disease detection, yield forecasting, and decision-support applications.

The integrated evidence base shows that technical research is embedded in a supply and service ecosystem that influences which technologies are developed, whose data are available, and how analytical outputs reach farms. Reviews of the digital agricultural supply side identify technology firms, equipment manufacturers, input providers, platforms, start-ups, advisers, research organisations, and public actors as important drivers of digital development [32]. Their roles shape interoperability, business models, access, maintenance, data control, and the capacity to translate remote-sensing or machine-learning outputs into decisions. Consequently, decision-support value cannot be inferred from model performance alone. It also depends on whether data flows are sustainable, responsibilities are allocated, interfaces fit existing workflows, and users can understand, contest, or override analytical outputs.

Across application domains, a common pattern is that evidence strength declines as claims move from technical discrimination or prediction toward reliable action in open environments. Crop-monitoring reviews show extensive sensing and feature-extraction activity; disease reviews show concentrated image-recognition research; and yield reviews show diverse predictive and assimilative methods. Yet systemic agricultural-AI risk extends beyond average accuracy to include distributional effects, feedback loops, resource use, cybersecurity, market concentration, and harms arising from interactions among models, institutions, environments, and users [33]. The umbrella synthesis therefore interprets field readiness as a system property requiring evidence about data, models, infrastructure, human oversight, maintenance, and consequences rather than as a direct extension of benchmark performance.

Convergence is strongest around the expansion of remote sensing and machine learning, the centrality of data quality, and the need for validation beyond the development dataset. Fragmentation is greatest in benchmark definitions, sensor and crop contexts, disease ground truth, forecast horizons, uncertainty reporting, and links to practical decisions. Responsible digital-agriculture guidance further indicates that transparency, accountability, privacy, fairness, stakeholder participation, and governance should be considered throughout the analytical lifecycle [34]. These requirements connect methodological limitations to future research priorities: better cross-context testing is needed to assess transferability; transparent datasets and workflows are needed for comparison; uncertainty must be communicated at decision-relevant horizons; and user-centred evaluation is needed before advisory value is inferred. Figure 1 synthesises review-level evidence across remote sensing, machine learning, crop monitoring, disease detection, yield forecasting, decision support, evidence limitations, and future research priorities.

Alt-text description: A conceptual umbrella evidence synthesis map showing review-level evidence from remote sensing and machine learning reviews connected to crop monitoring, disease detection, yield forecasting, decision support, evidence strength, methodological limitations, gaps, and future research priorities.

Reproducibility is a cross-cutting condition for interpreting apparent evidence strength. Remote-sensing research can be difficult to reproduce when data versions, preprocessing decisions, code, computational environments, spatial partitions, parameter settings, or analytical workflows are incompletely reported [35]. Similar concerns apply to image-based disease detection and yield modelling, where undocumented augmentation, data cleaning, hyperparameter selection, or feature engineering can alter outcomes. Reproducibility does not prove validity, because a transparent workflow may still contain biased samples or inappropriate validation. It does, however, enable scrutiny, sensitivity analysis, independent reruns, and more credible comparison. Benchmarking should therefore include data provenance, contextual metadata, declared splits, and complete evaluation protocols rather than relying on leaderboard values alone.

Overall, the umbrella synthesis identifies a layered evidence structure. Breadth is greatest for remote-sensing applications and image-based machine learning; methodological concentration is strongest for crop monitoring and controlled-image disease classification; and evidence is more variable for transferable disease recognition, calibrated yield forecasting, and workflow-embedded decision support. Review overlap and scope heterogeneity mean that repeated positive conclusions cannot be counted as independent confirmations without examining their underlying studies [12]. The appropriate interpretation is that remote sensing and machine learning offer a substantial, rapidly developing research base whose decision-support value remains conditional on data provenance, independent validation, uncertainty, reproducibility, explainability, institutional context, and responsible deployment assessment. This synthesis supports comparative evidence mapping and research-priority development, not universal model selection or operational prescription.

The first priority is to improve the reliability of review-level inference itself. Future umbrella reviews should map overlap among included reviews whenever primary-study lists permit, because repeated inclusion can create apparent convergence without independent replication [12]. Review reporting should specify search coverage, review type, eligibility rules, primary-study designs, application definitions, appraisal methods, and limits to synthesis, consistent with the methodological need for explicitly planned overview questions and methods [9]. Benchmark transparency should extend beyond dataset names to versions, provenance, licensing, duplicate detection, class distributions, preprocessing, split logic, and known domain constraints. These advances would make evidence updates and comparisons more defensible, although they would not remove bias or validate the underlying primary studies by themselves.

A second priority is external validation designed around agricultural variation. Cross-region testing should separate training and evaluation geographically; cross-crop testing should hold out entire crop types or provide per-crop transport analyses; and multi-season testing should preserve chronology and include genuinely unseen seasons. Yield reviews particularly support the need for geographically and temporally independent evaluation because current comparisons are often crop-, region-, and dataset-specific [31]. Disease-recognition reviews likewise justify prospective field datasets containing complex backgrounds, cultivar diversity, mixed stressors, symptom stages, and expert-confirmed reference labels [22]. Yield forecasting should state the true information cut-off, forecast horizon, and data availability, and should complement point errors with calibrated intervals, probabilistic scores, or scenario sensitivity. These practices would clarify transportability and risk without implying global or universal generalisation.

A third priority is integration that remains testable, explainable, and reproducible. Weather, soil, management, crop-model, and remote-sensing data should be combined through modular designs that preserve source provenance, describe missing-data handling, and permit ablation of each modality [29]. Process-based and data-driven components should report their separate and propagated uncertainties rather than allowing hybrid complexity to obscure error sources [30]. Explainable machine learning should be evaluated for stability, faithfulness, agronomic coherence, and usefulness to intended users; explanation alone must not substitute for predictive validity, fairness, or accountability [34]. Executable workflows, versioned data and code, declared environments, and independent reruns would strengthen reproducibility and comparison [35]. Open evaluation protocols should include spatial, temporal, imbalance-aware, and external tests, while avoiding the interpretation of leaderboard success as deployment certification.

A final priority is to study translation from analytical output to responsible decision support. Farmer-centred and workflow-embedded evaluations should examine timing, comprehensibility, burden, access, trust calibration, override behaviour, error consequences, and diversity among farm contexts, reflecting the social and institutional dimensions emphasised in digital-agriculture reviews [5]. Supply-side structures, service relationships, ownership arrangements, and platform dependencies should also be examined because they affect who can use a system and under what conditions [32]. Responsible deployment assessment should address privacy, fairness, accountability, security, environmental and market externalities, incident handling, and post-deployment monitoring where ethically appropriate [33]. Future umbrella-review updates should maintain versioned extraction records and reassess overlap, methodological quality, and emerging evidence rather than appending new reviews mechanically. Table 4 presents evidence gaps and future research priorities emerging from the umbrella review.

This umbrella review provides an integrated review-level synthesis of remote sensing, machine learning, and smart-farming evidence for crop monitoring, disease detection, yield forecasting, and decision support. The reviewed literature indicates substantial methodological development across satellite and UAV observation, multispectral and hyperspectral imaging, computer vision, conventional machine learning, deep learning, crop models, and multimodal forecasting. Its strongest contribution is evidence mapping: it identifies where research is concentrated, where findings converge, and where methodological limitations recur. Review-level evidence does not independently establish primary-study validity, reliable diagnosis, operational forecasting, universal model transfer, farmer approval, or deployment readiness. Claims of practical readiness should therefore remain conditional on review quality, primary-study overlap, transparent data provenance, independent spatial and temporal validation, cross-crop and cross-region testing, uncertainty assessment, reproducibility, explainability, workflow fit, and responsible governance. Progress toward decision-support value requires not only better models but also stronger evidence designs and explicit attention to the institutional and human settings in which analytical outputs may be used.

None

None

None

None

Gabriel Costa, Lucas Ribeiro, Ricardo Alves & Mariana Lopes contributed to this work.

Authors and affiliations

Department of Remote Sensing and Machine Learning for Agriculture, Faculty of Agricultural Sciences, Federal University of Minas Gerais, Belo Horizonte, Brazil
Gabriel Costa & Lucas Ribeiro

Department of Crop Monitoring and Disease Detection, Faculty of Agriculture, University of Coimbra, Coimbra, Portugal
Ricardo Alves

Department of Yield Forecasting and Umbrella Review, Faculty of Agriculture, University of São Paulo, São Paulo, Brazil
Mariana Lopes

Corresponding author

Correspondence to Gabriel Costa

Open Access The author(s) retain copyright. This article is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. It may be shared and adapted for non-commercial purposes with appropriate attribution, an indication of changes, and distribution of adaptations under the same license. Third-party material may be subject to separate terms identified in its credit line. View the license at https://creativecommons.org/licenses/by-nc-sa/4.0/.