Computer vision of beans literature review

This working literature review covers image-based crop yield estimation, canopy measurements, plant phenotyping, and model validation, with emphasis on snap beans. It includes direct bean studies, related crop applications, and original references for the computer vision and statistical methods used in this area.

The collection contains 55 references and was updated on 12 November 2026. It is a targeted literature review, not a systematic or exhaustive review. Papers are included according to the questions they address and the evidence they provide.

• Snap bean yield estimation using image-derived canopy information
• RGB imaging and comparisons with multispectral, hyperspectral, and LiDAR approaches
• Plant-level measurements and repeated observations across the growing season
• Pod detection, occlusion, and the distinction between visible pods and harvested yield
• Complementary canopy features and learned visual representations
• Original sources for detection, segmentation, feature extraction, regression, and software
• Validation accounting for repeated observations, spatial structure, and limited sample sizes

Each entry gives a bibliographic citation, a DOI or publisher link, a summary of the approach and its relevance, limitations, and the material reviewed. Abstract-level or partial-text review supports broad descriptions; detailed numerical comparisons require the full methods and results.

Fresh pod mass, dry seed yield, whole-plant biomass, leaf counts, visible-pod detection, and maturity classification are different targets. Per-plant, per-pot, and per-plot results also differ. Correlation, R², adjusted R², detection average precision, and custom accuracy percentages should not be treated as interchangeable.

The references form a working collection. Inclusion here does not mean that every paper must be cited in a manuscript.

Hassanzadeh, A., van Aardt, J., Murphy, S.P., Pethybridge, S.J., 2020. Yield modeling of snap bean based on hyperspectral sensing: a greenhouse study. Journal of Applied Remote Sensing 14(2), 024519.

DOI / article

Approach and relevance: Repeated canopy spectroscopy, measured canopy width, plant height and leaf number were evaluated with partial least squares regression for individual-plant snap bean pod weight. This provides direct precedent for both the prediction unit and the combination of structural and spectral information.

Limitations: The 48 plants were divided into two harvest groups, with each plant harvested once at its assigned stage. The target was not cumulative seasonal harvest, and the inputs were hyperspectral and physical measurements rather than ordinary RGB photographs. Leave-one-out validation within this experiment does not establish independent-site transfer.

Review status: Relevant experimental design, measurement, regression and results sections inspected in an author-deposited published PDF.

Hassanzadeh, A., Zhang, F., van Aardt, J., Murphy, S.P., Pethybridge, S.J., 2021. Broadacre Crop Yield Estimation Using Imaging Spectroscopy from Unmanned Aerial Systems (UAS): A Field-Based Case Study with Snap Bean. Remote Sensing 13(16), 3241.

DOI / article

Approach and relevance: UAS hyperspectral imagery was used to model field snap bean pod yield and seed length. The study considered observation timing and spectral selection.

Limitations: The endpoint is plot-scale pod yield from hyperspectral imagery. Calibration and leave-one-out cross-validation results must be distinguished; fitted R² values should not be described as independent prediction accuracy. The study does not establish transfer to a new site or year. Its use of SAM refers to Spectral Angle Mapper, not Segment Anything.

Review status: Relevant publisher full-text experimental, feature-selection, validation and results sections inspected.

Zhang, F., Hassanzadeh, A., Letendre, P., Kikkert, J., Pethybridge, S., van Aardt, J., 2025. Enhancing snap bean yield prediction through synergistic integration of UAS-Based LiDAR and multispectral imagery. Computers and Electronics in Agriculture 230, 109923.

DOI / article

Approach and relevance: Two years of snap bean trials used UAS-based LiDAR and multispectral imagery. Combining structural and spectral information improved yield prediction relative to the sensor-specific and vegetation-index benchmarks tested.

Limitations: Feature fusion is already established prior work. The detailed splitting and preprocessing procedures still need a full-method check before numerical comparison with another study; adjusted R² is not interchangeable with correlation or an external-test score.

Review status: Publisher abstract and metadata inspected; detailed methods not audited.

Xu, Z., Gill, T., Huang, Q., Wang, Y., 2026. A comprehensive comparison of multispectral and hyperspectral imagery for plot-level crop yield prediction of kidney beans, snap beans, and potatoes. Smart Agricultural Technology 14, 102234.

DOI / article

Approach and relevance: UAV multispectral imagery and low-altitude aircraft hyperspectral imagery were compared for plot-level yield prediction in kidney beans, snap beans and potatoes. The abstract reports comparable multispectral and hyperspectral performance for snap beans, and examines observation timing and band selection.

Limitations: Ordinary RGB-only imaging was not tested. Single-date versus multiple-date findings need crop-specific checking before they are attributed to snap beans.

Review status: Publisher abstract and publisher-deposited bibliographic metadata inspected; full methods not accessed.

Köksal, E.S., Kara, T., Apan, M., Üstün, H., İlbeyi, A., 2008. Estimation of green bean yield, water deficiency and productivity using spectral indexes during the growing season. Irrigation and Drainage Systems 22(3-4), 209-223.

DOI / article

Approach and relevance: Green-bean spectral indices were related to yield, water deficiency, and productivity in field irrigation experiments. This provides historical context for linking canopy spectra to bean production.

Limitations: The accessible abstract does not establish the exact sensor hardware, sampling unit, or independent validation design. This is not ordinary RGB photography.

Review status: Author-institution abstract and metadata inspected; full text not accessed.

El-baki, M.S.A., Ibrahim, M.M., Elsayed, S., Yaseen, Z.M., El-Fattah, N.G.A., 2025. Water status and plant traits of dry bean assessment using integrated spectral reflectance and RGB image indices with artificial intelligence. Scientific Reports 15(1), 16808.

DOI / article

Approach and relevance: Smartphone nadir RGB canopy images, spectral reflectance indices and neural networks were used to assess dry-bean traits and seed yield. Separate and combined index models provide relevant prior work on accessible RGB measurements in beans.

Limitations: The target is dry seed yield aggregated from sampled plants and replicates and expressed per unit area, rather than individual-plant fresh-pod mass. The ANN analysis used a random 70/30 sample split and five-fold training cross-validation. Plant-, row- and year-independent evaluation is not established, and the mapping between aggregated yield labels and modelling samples needs care before numerical comparison.

Review status: Relevant full-text image-acquisition, harvest-measurement, ANN validation and results sections inspected through the published PMC article.

Hassanzadeh, A., Murphy, S.P., Pethybridge, S.J., van Aardt, J., 2020. Growth Stage Classification and Harvest Scheduling of Snap Bean Using Hyperspectral Sensing: A Greenhouse Study. Remote Sensing 12(22), 3809.

DOI / article

Approach and relevance: Canopy spectroscopy was used to classify snap bean growth stages and assess harvest readiness. The study is relevant to the timing of sensing and harvest decisions.

Limitations: Growth-stage classification is a different task from harvested-mass regression. Classification accuracy is not a yield-prediction benchmark.

Review status: Publisher abstract, metadata and relevant classification-validation methods inspected.

Sánchez, N.E., Garzón, J., Londoño, D.F., 2025. Non-Destructive Yield Prediction in Common Bean Using UAV-Based Spectral and Structural Metrics: Implications for Sustainable Crop Management. Sustainability 17(22), 10066.

DOI / article

Approach and relevance: UAV spectral indices and canopy area, ground-based TRAC structural measurements and phenological variables were used to model individual dry-grain yield from 101 common-bean plants. PCA, regularized regression and parsimonious predictor sets were evaluated.

Limitations: The target is individual dry grain, standardized to 14% moisture, rather than seasonal fresh-pod mass. Inputs include multispectral and ground measurements. Internal validation and a random holdout are reported; independent-site or independent-year transfer is not established.

Review status: Publisher full-text design, sensors, harvest target and modelling sections inspected.

Sankaran, S., Zhou, J., Khot, L.R., Trapp, J.J., Mndolwa, E., Miklas, P.N., 2018. High-throughput field phenotyping in dry bean using small unmanned aerial vehicle based multispectral imagery. Computers and Electronics in Agriculture 151, 84–92.

DOI / article

Approach and relevance: Multispectral UAV canopy area and green normalized difference vegetation index were related to dry-bean seed yield and other traits across drought and nitrogen-stress trials over two seasons. Observation stage influenced the relationships.

Limitations: Plot-level dry seed yield and multispectral sensing differ from individual fresh-pod mass and handheld RGB. Reported correlations do not establish independent transfer of a yield model.

Review status: Publisher metadata and the author-institution USDA technical abstract inspected.

Lipovac, A., Bezdan, A., Moravčević, D., Djurović, N., Ćosić, M., Benka, P., Stričević, R., 2022. Correlation between Ground Measurements and UAV Sensed Vegetation Indices for Yield Prediction of Common Bean Grown under Different Irrigation Treatments and Sowing Periods. Water 14(22), 3786.

DOI / article

Approach and relevance: Multispectral UAV vegetation indices were related to canopy cover, leaf area index, water status and common-bean yield under different irrigation treatments and sowing periods. The study documents cases where pod damage was not reflected in canopy indices.

Limitations: This is multispectral field monitoring, not individual-plant RGB fresh-pod prediction. Relationships varied with sowing period and pest damage; they should not be treated as universally transferable yield equations.

Review status: Publisher abstract, discussion, conclusions and metadata inspected; detailed validation design not fully audited.

Saravia, D., Valqui-Valqui, L., Salazar, W., Quille-Mamani, J., Barboza, E., Porras-Jorge, R., Injante, P., Arbizu, C.I., 2023. Yield Prediction of Four Bean (Phaseolus vulgaris) Cultivars Using Vegetation Indices Based on Multispectral Images from UAV in an Arid Zone of Peru. Drones 7(5), 325.

DOI / article

Approach and relevance: Repeated multispectral UAV observations and vegetation indices were used in yield models for four bean cultivars across 48 plots. The study provides same-species context for acquisition timing and canopy-index relationships.

Limitations: Yield was dry grain harvested from plots, standardized to 12–14% moisture and expressed per area. The imagery used ten spectral bands. Independent-site or independent-year model validation was not established in the inspected sections.

Review status: Publisher full-text experimental design, sensor and harvest methods inspected.

Gomes Campos, D., Nogueira Martins, R., 2024. Bean yield estimation using unmanned aerial vehicle imagery. Revista Brasileira de Engenharia de Biossistemas 18, 1219.

DOI / article

Approach and relevance: Five-band multispectral UAV imagery from six observation dates supplied vegetation variables for common-bean grain-yield estimation. This provides additional evidence on canopy sensing and observation timing.

Limitations: Yield is expressed in kg/ha rather than individual fresh-pod mass. The reported regression fit should not be treated as independently validated transfer performance; detailed validation methods were not audited.

Review status: Publisher abstract, bibliographic metadata and indexed published-PDF front matter inspected.

Mortlock, M.Y., Carey, D., Murray, H., Baker, P.J., Corry, P.G., 2025. Enhancing green bean crop maturity and yield prediction by harnessing the power of statistical analysis, crop records and weather data. PLOS ONE 20(3), e0306266.

DOI / article

Approach and relevance: Commercial crop records, weather and pod fresh-weight observations were analysed to improve green-bean maturity and yield prediction. This supplies agronomic context for harvest timing and fresh-yield assessment.

Limitations: This is not a computer-vision study. It supports the importance of timing and environmental conditions rather than the performance of image-based predictors.

Review status: Publisher metadata, abstract and relevant crop and fresh-yield methods inspected.

Jollet, D., Junker-Frohn, L.V., Steier, A., Meyer-Lüpken, T., Müller-Linow, M., 2023. A new computer vision workflow to assess yield quality traits in bush bean (Phaseolus vulgaris L.). Smart Agricultural Technology 5, 100306.

DOI / article

Approach and relevance: RGB images from a closed imaging box were used to measure the shape and colour of harvested bush-bean pods, with expert reference measurements. Pods were placed separately without touching.

Limitations: The target was postharvest pod quality under controlled imaging conditions, not seasonal yield prediction from standing-plant canopies.

Review status: Published full text inspected through an institutional repository.

Qi, J., Liu, C., Chu, H., 2025. Kidney bean detection in complex agricultural scenarios based on the kidneyB-YOLO model. Frontiers in Plant Science 16, 1672179.

DOI / article

Approach and relevance: KidneyB-YOLO was evaluated for pod detection in open-field trellis scenes. Colour similarity and occlusion were identified as challenges. The annotation protocol excluded targets with approximately more than half of the pod body occluded.

Limitations: Detection scores concern annotated visible pods, not all hidden pods or harvested mass. Images originated from video frames and the paper describes random splitting of augmented images; plant-, video- and site-disjoint evaluation is not established. The trellised crop setting differs from bush snap beans.

Review status: Relevant publisher full-text data collection, annotation and data-splitting methods inspected.

Torres, C., Clément, A., Frison, E., Auperpin, E., Parmentier, P., Feutry, A., 2012. Image analysis for characterising French bean (Phaseolus vulgaris L.) pods. Acta Horticulturae 934, 145–150.

DOI / article

Approach and relevance: Image analysis extracted French-bean pod length, diameter, and curvature and related projected pod surface to weight for grading. The official abstract reports validation on more than 150 batches.

Limitations: This is postharvest grading rather than standing-plant seasonal-yield estimation. A numerical correlation was not available in the inspected abstract.

Review status: Publisher abstract and metadata inspected; full text not accessed.

Zhang, F., Hassanzadeh, A., Kikkert, J., Pethybridge, S.J., van Aardt, J., 2021. Comparison of UAS-Based Structure-from-Motion and LiDAR for Structural Characterization of Short Broadacre Crops. Remote Sensing 13(19), 3975.

DOI / article

Approach and relevance: UAS structure-from-motion and LiDAR were compared for measuring canopy height and row width in short snap bean crops.

Limitations: The study is about structural measurement, not harvested-yield prediction or validation of a two-dimensional pot-scale height method. Its image platform used a multispectral camera.

Review status: Publisher abstract and sensor-method excerpts inspected.

Quille-Mamani, J., Porras-Jorge, R., Saravia-Navarro, D., Herrera, J., Chávez-Galarza, J., Arbizu, C.I., Valqui-Valqui, L., 2022. Assessment of vegetation indices derived from UAV images for predicting biometric variables in bean during ripening stage. Idesia 40(2), 39–45.

DOI / article

Approach and relevance: RGB and multispectral UAV indices were related to bean biometric traits during ripening. Associations between RGB indices and measured height provide relevant evidence for accessible canopy phenotyping.

Limitations: The endpoint is plant morphology rather than harvested yield. Predicting height from index relationships is different from direct geometric height measurement.

Review status: Publisher-hosted SciELO abstract and experimental-method excerpts, and an author-institution record, inspected.

Parker, T.A., Palkovic, A., Gepts, P., 2020. Determining the Genetic Control of Common Bean Early-Growth Rate Using Unmanned Aerial Vehicles. Remote Sensing 12(11), 1748.

DOI / article

Approach and relevance: UAV imaging was used to phenotype early common-bean canopy growth and investigate its genetic control. This provides same-species precedent for repeated canopy measurements.

Limitations: Early-growth phenotyping and genetic mapping are different from harvested-yield regression. The paper should not be cited as evidence of seasonal fresh-pod prediction accuracy.

Review status: Author-institution publication record and primary preprint abstract screened; full published methods not audited.

Maimaitijiang, M., Sagan, V., Sidike, P., Hartling, S., Esposito, F., Fritschi, F.B., 2020. Soybean yield prediction from UAV using multimodal data fusion and deep learning. Remote Sensing of Environment 237, 111599.

DOI / article

Approach and relevance: UAV RGB, multispectral, and thermal observations supplied structural, spectral, thermal, and texture information for soybean grain-yield models. Alternative data-fusion approaches were evaluated.

Limitations: Soybean grain yield at field or plot scale differs from individual snap bean fresh-pod mass. The study motivates testing complementary predictors but does not establish that fusion will improve every model.

Review status: Publisher abstract, highlights and author-institution metadata inspected; detailed validation methods not audited.

Tatsumi, K., Igarashi, N., Mengxue, X., 2021. Prediction of plant-level tomato biomass and yield using machine learning with unmanned aerial vehicle imagery. Plant Methods 17, 77.

DOI / article

Approach and relevance: Repeated RGB and multispectral UAV images provided plant-level height, vegetation-index, texture, and growth-rate variables for predicting tomato shoot mass, fruit weight, and fruit number.

Limitations: The crop, acquisition system, and endpoints differ from snap bean seasonal fresh-pod mass. Plant height was not uniformly useful for fruit-yield prediction.

Review status: Relevant sections of the full publisher article inspected.

Bendig, J., Yu, K., Aasen, H., Bolten, A., Bennertz, S., Broscheit, J., Gnyp, M.L., Bareth, G., 2015. Combining UAV-based plant height from crop surface models, visible, and near infrared vegetation indices for biomass monitoring in barley. International Journal of Applied Earth Observation and Geoinformation 39, 79–87.

DOI / article

Approach and relevance: UAV RGB-derived canopy height and visible or near-infrared vegetation indices were evaluated as predictors of barley biomass. Visible-band indices combined with height showed potential during early growth.

Limitations: Biomass is different from harvested pod yield. Some reflectance information came from a ground spectrometer, so the entire experiment should not be described as ordinary RGB-only photography.

Review status: Publisher abstract and author-institution metadata inspected.

Lu, W., Du, R., Niu, P., Xing, G., Luo, H., Deng, Y., Shu, L., 2022. Soybean Yield Preharvest Prediction Based on Bean Pods and Leaves Image Recognition Using Deep Learning Neural Network Combined With GRNN. Frontiers in Plant Science 12, 791256.

DOI / article

Approach and relevance: Image recognition of soybean leaves and pod classes was combined with a generalized regression neural network to estimate seed counts. Measured average seed weights were then used to calculate yield.

Limitations: The target was soybean grain mass at pot level, not individual snap bean seasonal fresh-pod mass. The custom yield-accuracy percentage is not R². The paper describes augmentation before random image splitting, so independent plant or pot evaluation is not established.

Review status: Relevant publisher full-text acquisition, split procedure, target calculation and results sections inspected.

Johansen, K., Morton, M.J.L., Malbeteau, Y., Aragon, B., Al-Mashharawi, S., Ziliani, M.G., Angel, Y., Fiene, G., Negrão, S., Mousa, M.A.A., Tester, M.A., McCabe, M.F., 2020. Predicting Biomass and Yield in a Tomato Phenotyping Experiment Using UAV Imagery and Random Forest. Frontiers in Artificial Intelligence 3, 28.

DOI / article

Approach and relevance: UAV-derived plant shape, vegetation indices and texture were used to predict tomato fresh shoot mass, fruit number and yield mass under control and salinity treatments. This supports indirect estimation using canopy information.

Limitations: The work includes RGB and multispectral analyses and differs from handheld RGB snap-bean imagery. Aggregate performance should not be confused with accuracy for individual plants.

Review status: Published full-text abstract and relevant discussion inspected; detailed validation design not completely audited.

Aguilar-Ariza, A., et al., 2023. UAV-based individual Chinese cabbage weight prediction using multi-temporal data. Scientific Reports 13, 20122.

DOI / article

Approach and relevance: Automatic plant detection and repeated UAV RGB and multispectral observations supplied individual-plant structural, spectral and temporal features for predicting Chinese cabbage harvest weight. Different temporal resolutions and preharvest observation windows were compared.

Limitations: The inputs are not RGB-only, and whole-cabbage harvest weight differs from cumulative fresh-pod yield. The study’s absolute prediction errors should not be compared directly with errors for much smaller bean yields.

Review status: Publisher full-text abstract, feature extraction, data-splitting and results sections inspected.

Riera, L.G., Carroll, M.E., Zhang, Z., Shook, J.M., Ghosal, S., Gao, T., Singh, A., Bhattacharya, S., Ganapathysubramanian, B., Singh, A.K., Sarkar, S., 2021. Deep Multiview Image Fusion for Soybean Yield Estimation in Breeding Applications. Plant Phenomics 2021, 9846470.

DOI / article

Approach and relevance: Multiple soybean RGB views were combined using frozen VGG19 features from a supervised pod detector and a trained regression module. Pod counts were estimated for breeding plots, supporting discussion of multiview imaging, occlusion and learned representations.

Limitations: Pod count is a yield proxy, not measured harvested seed or fresh-pod mass. The feature extractor had supervised pod-detection training, unlike general-purpose self-supervised DINOv2. A plot contains multiple plants.

Review status: Published full-text data collection, architecture, training and results sections inspected through PMC.

Tanaka, Y., et al., 2023. Deep Learning Enables Instant and Versatile Estimation of Rice Yield Using Ground-Based RGB Images. Plant Phenomics 5, 0073.

DOI / article

Approach and relevance: A convolutional neural network estimated rice grain yield from ground-based canopy RGB photographs collected across seven countries. The study examined independent-data performance and sensitivity to illumination, imaging angle and observation timing.

Limitations: This is plot-level rice grain yield with a large supervised training collection. It is not individual snap-bean fresh-pod estimation or a frozen foundation-model approach. Its R² should not be compared directly with Pearson correlation.

Review status: Publisher metadata, abstract and relevant primary methods and results excerpts inspected.

Quamer, A., Asgari, N., Pearce, J.M., 2026. Two-stage deep learning model for non-destructive leaf counting in tomato crops. Computers and Electronics in Agriculture 247, 111717.

DOI / article

Approach and relevance: A MobileNetV2 regression stage and a U-Net segmentation stage were used for longitudinal tomato leaf counting under greenhouse and agrivoltaic greenhouse conditions. The study included 36 plants over five weeks and used pseudo-labeled counts where manual counting was infeasible.

Limitations: Leaf counting requires different ground truth from harvested-yield estimation. The workflow is relevant to repeated imaging and complementary image-analysis components; it is not evidence of another pipeline’s accuracy or generalization.

Review status: Publisher abstract, highlights and bibliographic metadata inspected.

Minervini, M., Giuffrida, M.V., Perata, P., Tsaftaris, S.A., 2017. Phenotiki: an open software and hardware platform for affordable and easy image-based phenotyping of rosette-shaped plants. The Plant Journal 90(1), 204–216.

DOI / article

Approach and relevance: Affordable off-the-shelf imaging hardware and openly available analysis software supported repeated measurements of rosette-shaped plant morphology and growth.

Limitations: A controlled rosette-plant phenotyping platform does not establish snap bean yield accuracy or robustness to arbitrary field backgrounds.

Review status: Publisher summary and primary-paper abstract inspected.

Giuffrida, M.V., Doerner, P., Tsaftaris, S.A., 2018. Pheno-Deep Counter: a unified and versatile deep learning architecture for leaf counting. The Plant Journal 96(4), 880–890.

DOI / article

Approach and relevance: A multimodal deep network learned image features and leaf-count regression for top-view rosette-shaped plants across several species and imaging modalities.

Limitations: This is leaf counting rather than yield estimation, and it is not a frozen foundation-model embedding approach. Results apply to the imaging conditions evaluated.

Review status: Publisher summary, introduction and bibliographic citation inspected; detailed experiments not fully audited.

Akiyama, R., et al., 2023. Seasonal pigment fluctuation in diploid and polyploid Arabidopsis revealed by machine learning-based phenotyping method PlantServation. Nature Communications 14, 5792.

DOI / article

Approach and relevance: PlantServation combined outdoor RGB acquisition with automated individual-plant detection and segmentation to track plant colour and estimate anthocyanin content over time. It demonstrates repeated image-based phenotyping under changing field conditions.

Limitations: The endpoint is pigment-related phenotyping, not harvested yield. The study does not establish that an arbitrary RGB greenness index measures chlorophyll or validate frozen-embedding yield regression.

Review status: Publisher full-text excerpts covering acquisition, phenotyping, results and metadata inspected.

Chen, X., Nousi, P., Boss, M., Volpi, M., Roth, L., 2025. Comparative Analysis of Image-Based Deep Learning and Genomic Models for Yield and Protein Content Prediction in Winter Wheat. In: 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 7178–7186.

DOI / article

Approach and relevance: Frozen DINOv2 ViT-S/14 representations of image sequences were combined with a trainable temporal transformer for winter-wheat grain-yield and protein-content prediction. The study also examined adding plant-height sequences.

Limitations: Adding height did not consistently improve prediction; the main yield comparison reported poorer DINOv2 results after height was added. This is a workshop paper and does not study individual-plant pod mass or a simple PCA-plus-ridge pipeline.

Review status: Primary conference metadata and relevant full-paper methods, validation design and main results tables inspected.

Tonu, S., Tonu, I.-A., Zvorișteanu, O., Achirei, Ș.D., 2026. A Vision-Based Approach for Multi-Component Pasture Biomass Estimation. AgriEngineering 8(8), 330.

DOI / article

Approach and relevance: Frozen DINOv2 visual features were paired with ridge regression, tree-based regressors and cross-fitted stacking for multi-component pasture biomass estimation from RGB quadrat images.

Limitations: Pasture dry biomass differs from snap bean fresh-pod yield. Evaluation used repeated shuffled four-fold cross-validation without an independent external test. Canopy height and vegetation indices were not explicit additional predictors.

Review status: Publisher full-text feature extraction, evaluation and limitations sections inspected.

Hao, P., An, J., Cai, Q., Cao, J., Hu, Z., Lin, B., 2025. Self-Supervised and Multi-Task Learning Framework for Rapeseed Above-Ground Biomass Estimation. Agriculture 15(23), 2516.

DOI / article

Approach and relevance: Top-down smartphone RGB images, self-supervised plant-image pretraining, and multi-task learning were used to estimate fresh and dry rapeseed biomass with relatively limited labeled data.

Limitations: The backbone was fine-tuned rather than used as an unchanged frozen feature extractor. Evaluation used cultivar-grouped five-fold cross-validation with a separate validation fold within the training data. Seedling fresh and dry biomass differ from seasonal harvested pod mass.

Review status: Publisher full-text fine-tuning and cultivar-grouped validation methods inspected, alongside metadata and abstract.

Chen, F., Giuffrida, M.V., Tsaftaris, S.A., 2023. Adapting Vision Foundation Models for Plant Phenotyping. In: 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 604–613.

DOI / article

Approach and relevance: MAE, DINO and DINOv2 representations were evaluated for plant phenotyping using both decoder-only tuning with unchanged frozen backbones and LoRA adapter tuning. The tasks included leaf counting, disease classification and leaf segmentation, with the evaluated models differing by task.

Limitations: Decoder-only tuning is a precedent for unchanged pretrained features, while LoRA changes the representation through trainable adapters. Neither configuration is identical to PCA followed by ridge regression. This workshop paper does not evaluate harvested-yield prediction.

Review status: Primary conference metadata, architecture figure and relevant full-text methods inspected.

Vitali, G., Terzi, F., Arru, M., Pelusi, F., Fragassa, C., Gatti, M., Mattetti, M., 2026. An Image-Based Approach for Grape Yield Estimation During Mechanical Harvesting Using Pre-Trained Visual Embeddings. AgriEngineering 8(9), 402.

DOI / article

Approach and relevance: Images recorded during mechanical grape harvesting were represented using fixed pretrained models, including DINOv2 ViT-B/14. LASSO and sparse-fusion regression related the extracted features to measured yield, providing a close precedent for agricultural embedding-based regression.

Limitations: The target concerns harvested grape flow rather than standing-plant pod mass. Feature selection used the complete dataset, and the authors explicitly acknowledge the lack of independent validation. Reported fitted R² values are exploratory rather than established transfer accuracy.

Review status: Publisher full-text feature extraction, model table, selection procedure and limitations inspected.

Ubbens, J., Cieslak, M., Prusinkiewicz, P., Parkin, I., Ebersbach, J., Stavness, I., 2020. Latent Space Phenotyping: Automatic Image-Based Phenotyping for Treatment Studies. Plant Phenomics 2020, 5801869.

DOI / article

Approach and relevance: Learned image representations were used to characterize treatment-response phenotypes beyond a predefined set of explicit measurements. The study supplies conceptual background for latent image features in plant phenotyping.

Limitations: The representation was trained with treatment supervision. It is not a frozen general-purpose foundation model or a harvested-yield study and does not establish small-sample yield-regression performance.

Review status: Primary metadata and indexed full-text abstract, methods and discussion excerpts inspected.

Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., Zhu, J., Zhang, L., 2025. Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection. In: Computer Vision – ECCV 2024. Lecture Notes in Computer Science 15105, pp. 38–55.

DOI / article

Approach and relevance: Introduces a language-guided open-set object detector that accepts category-name or referring-expression prompts. This is the original method reference for Grounding DINO.

Limitations: Generic detection benchmarks do not establish bean detection accuracy. A crop application must report the checkpoint, prompt, thresholds, and software version and evaluate its own detections.

Review status: Publisher abstract and metadata inspected; full chapter not read.

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.-Y., Dollár, P., Girshick, R., 2023. Segment Anything. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 3992-4003.

DOI / article

Approach and relevance: Introduces the Segment Anything Model (SAM) for promptable image segmentation. This is the original method reference for SAM-derived masks.

Limitations: The original benchmarks do not validate bean canopy masks or occlusion robustness. The selected checkpoint, prompting procedure, and mask quality need study-specific documentation.

Review status: Official CVF abstract and bibliographic record inspected; IEEE DOI resolves to the corresponding paper, but detailed IEEE metadata was not accessible during this update.

Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.-Y., Li, S.-W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P., 2024. DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research.

Published article

Approach and relevance: Introduces self-supervised DINOv2 models and transferable visual representations. The official implementation includes ViT-B/14 checkpoints.

Limitations: General visual transfer does not establish snap bean yield accuracy. Applications should identify the exact checkpoint, frozen or fine-tuned setting, and feature-extraction procedure.

Review status: Journal publication listing, author-deposited published PDF and relevant model sections, and official implementation inspected.

Woebbecke, D.M., Meyer, G.E., Von Bargen, K., Mortensen, D.A., 1995. Color Indices for Weed Identification Under Various Soil, Residue, and Lighting Conditions. Transactions of the ASAE 38(1), 259–269.

DOI / article

Approach and relevance: Evaluates colour indices for separating vegetation from soil and residues under varying conditions. The normalized chromatic-coordinate index 2g − r − b provides a source for excess-green features.

Limitations: An implementation must specify whether it uses normalized r, g, b or raw R, G, B values and document its masking or threshold rules. Vegetation separation is not validation of harvested-yield prediction.

Review status: Publisher abstract and bibliographic metadata inspected.

Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics. Chemometrics and Intelligent Laboratory Systems 58(2), 109–130.

DOI / article

Approach and relevance: Describes partial least squares regression using latent variables that relate predictor and response matrices. It provides methodological background for regression with correlated predictors.

Limitations: PLS is supervised and uses training outcomes. It is different from ordinary least squares. This reference does not determine the appropriate number of components or demonstrate a particular crop model’s accuracy.

Review status: Publisher abstract, section excerpts and original-paper metadata inspected.

Hoerl, A.E., Kennard, R.W., 1970. Ridge Regression: Biased Estimation for Nonorthogonal Problems. Technometrics 12(1), 55–67.

DOI / article

Approach and relevance: Presents a classic formulation of ridge regression for regularized estimation with correlated predictors.

Limitations: The penalty strength and any preceding dimensionality reduction must be specified and justified for the actual dataset. The original method paper does not establish crop-prediction accuracy.

Review status: Publisher metadata and the introduction and formulation in the original published article inspected.

Jolliffe, I.T., Cadima, J., 2016. Principal component analysis: a review and recent developments. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 374(2065), 20150202.

DOI / article

Approach and relevance: Reviews principal component analysis, which constructs uncorrelated components that successively maximize explained predictor variance and can reduce dimensionality.

Limitations: PCA is unsupervised. High predictor variance does not guarantee information about yield, and the paper does not justify a particular component count for a new dataset.

Review status: Publisher abstract and DOI metadata inspected.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, É., 2011. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12, 2825–2830.

Published article

Approach and relevance: Describes scikit-learn, a Python library for machine learning, including regression and preprocessing tools. This is the software citation when the analysis uses scikit-learn.

Limitations: The actual software version, transformations, hyperparameters, and data splits still need documentation. A software citation does not substitute for a reproducible methods description.

Review status: Publisher abstract, author list and official software citation instructions inspected.

He, K., Gkioxari, G., Dollár, P., Girshick, R., 2017. Mask R-CNN. In: IEEE International Conference on Computer Vision (ICCV), pp. 2961–2969.

DOI / article

Approach and relevance: Introduces an instance-segmentation framework that predicts object boxes, classes and masks. It is the original architecture reference for applications using Mask R-CNN to localize or segment pots and other objects.

Limitations: The original paper does not specify later Torchvision ResNet-50-FPN v2 weights or establish accuracy on bean photographs. Applications must identify their implementation, checkpoint, selected classes and filtering procedure.

Review status: Official conference metadata and abstract, and author-deposited abstract, inspected.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L., 2014. Microsoft COCO: Common Objects in Context. In: Computer Vision – ECCV 2014. Lecture Notes in Computer Science 8693, pp. 740–755.

DOI / article

Approach and relevance: Introduces the COCO dataset and its object-instance annotations. It provides the dataset reference when describing COCO-pretrained detection or segmentation models.

Limitations: COCO is not a bean-specific training or evaluation dataset. Pretraining on it does not establish the accuracy of pot selection or crop segmentation in a new experiment.

Review status: Publisher abstract, author list and bibliographic metadata inspected.

Habibi, L.N., Matsui, T., Tanaka, T.S.T., 2024. Critical evaluation of the effects of a cross-validation strategy and machine learning optimization on the prediction accuracy and transferability of a soybean yield prediction model using UAV-based remote sensing. Journal of Agriculture and Food Research 16, 101096.

DOI / article

Approach and relevance: Compares random, spatial and leave-one-field-out cross-validation strategies and independent-field evaluation for soybean yield prediction. The results show why random validation can give optimistic expectations for transfer to another field.

Limitations: Applying these findings to held-out plants or rows is a methodological interpretation that depends on the new study’s design. The paper does not by itself establish the adequacy of a particular plant or row split.

Review status: Author-deposited published full text inspected at the abstract and relevant validation passages; numerical analyses not independently reproduced.

Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera‐Arroita, G., Hauenstein, S., Lahoz‐Monfort, J.J., Schröder, B., Thuiller, W., Warton, D.I., Wintle, B.A., Hartig, F., Dormann, C.F., 2017. Cross‐validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8), 913-929.

DOI / article

Approach and relevance: Discusses cross-validation for temporal, spatial, hierarchical, and phylogenetic dependence. Validation should reflect both observation dependence and the intended prediction objective.

Limitations: Keeping repeated photographs of a plant together and evaluating held-out rows are applications of these principles, not procedures proven sufficient for a particular dataset. Blocking must reflect the intended prediction setting and can change the degree of extrapolation.

Review status: Publisher full-text abstract and relevant methodological sections inspected.

Krstajic, D., Buturovic, L.J., Leahy, D.E., Thomas, S., 2014. Cross-validation pitfalls when selecting and assessing regression and classification models. Journal of Cheminformatics 6(1), 10.

DOI / article

Approach and relevance: Explains pitfalls when selecting and assessing regression or classification models. Feature selection must be repeated within cross-validation, and tuning decisions should be separated from performance assessment through nested or suitably independent evaluation.

Limitations: The article discusses feature selection and model tuning but does not explicitly discuss PCA. A PCA-specific implementation rule should not be attributed to it. Repeating identical leave-one-plant-out partitions does not provide the partition variability obtained from repeated random folds.

Review status: Full primary article inspected.

Varoquaux, G., 2018. Cross-validation failure: Small sample sizes lead to large error bars. NeuroImage 180, 68-77.

DOI / article

Approach and relevance: Shows that cross-validation estimates can have substantial uncertainty with small samples and that variability between folds can understate that uncertainty.

Limitations: The examples come from neuroimaging. Numerical error bars, sample-size thresholds, and classification-specific percentages cannot be transferred directly to snap bean yield regression.

Review status: Publisher abstract inspected; full text not audited.

Murphy, K.M., Ludwig, E., Gutierrez, J., Gehan, M.A., 2024. Deep Learning in Image-Based Plant Phenotyping. Annual Review of Plant Biology 75, 771–795.

DOI / article

Approach and relevance: Reviews deep learning for extracting plant traits from images, including applications, evaluation practices, and unresolved challenges. It provides broad context for image-based crop measurement.

Limitations: Original experimental studies are still needed to support specific accuracy results and comparisons with the nearest prior work.

Review status: Publisher abstract and bibliographic metadata inspected.

van Klompenburg, T., Kassahun, A., Catal, C., 2020. Crop yield prediction using machine learning: A systematic literature review. Computers and Electronics in Agriculture 177, 105709.

DOI / article

Approach and relevance: Synthesizes machine-learning algorithms and input features used for crop-yield prediction, including widespread use of weather and soil variables and early deep-learning approaches.

Limitations: Its search predates recent foundation models and cannot establish the current absence of RGB bean-yield studies.

Review status: Publisher abstract, highlights and author-institution metadata inspected.

Tsaftaris, S.A., Minervini, M., Scharr, H., 2016. Machine Learning for Plant Phenotyping Needs Image Processing. Trends in Plant Science 21(12), 989–991.

DOI / article

Approach and relevance: A letter/comment explaining the continuing role of image processing and feature extraction in machine learning for plant phenotyping. It provides conceptual context for converting images into quantitative plant measurements.

Limitations: This is a methodological perspective, not evidence validating a particular segmentation, scaling or yield-prediction pipeline.

Review status: Author-institution metadata, reproduced opening text and an indexed accepted-manuscript excerpt inspected; complete full text not audited.

Minervini, M., Scharr, H., Tsaftaris, S.A., 2015. Image analysis: The new bottleneck in plant phenotyping [Applications Corner]. IEEE Signal Processing Magazine 32(4), 126–131.

DOI / article

Approach and relevance: Discusses image analysis as a constraint on converting plant-image collections into quantitative phenotypes. It provides background for the processing of repeated photographs into measurements.

Limitations: This is conceptual background rather than evidence validating a particular crop-yield model.

Review status: Author-institution abstract and accepted-manuscript metadata inspected.

The Grounding DINO chapter belongs to ECCV 2024, but Springer’s formal chapter citation uses 2025. The entry follows that DOI-linked publication record.

The Segment Anything entry uses IEEE DOI-linked pagination. CVF lists a different page range for the same work.

DINOv2 is cited as the 2024 Transactions on Machine Learning Research article. Its OpenReview link identifies the published record; an arXiv DOI is not substituted for a journal DOI.

Lu et al. is cited as 2022, following the publisher’s recommended citation, although its volume and DOI contain 2021. DOI strings do not necessarily match publication years.

The third author of Tatsumi et al. (2021) is recorded as Mengxue, X. in the publisher’s recommended citation. The entry follows that record.

The Plant Phenomics papers listed here were published when the journal was an AAAS Science Partner Journal. This affiliation should not be confused with publication in Science, Science Advances or Science Robotics. The journal’s legacy homepage states that it ceased publication with AAAS in January 2025.

For software implementations, record the version and checkpoint actually used. The official Torchvision model documentation describes the later Mask R-CNN ResNet-50-FPN v2 implementation. The scikit-learn guidance on common pitfalls explicitly discusses preprocessing leakage, including PCA. These documentation links supplement the 55 papers rather than count as additional research references.

The literature already includes individual-plant snap bean pod-weight models, individual dry-bean grain-yield models, RGB bean phenotyping, canopy feature fusion, and frozen visual representations for crop yield or biomass prediction. A new study should therefore define its exact crop, outcome, prediction unit, image-acquisition conditions, and validation setting before stating its contribution. The following are questions for investigation rather than completed findings:

• Does combining explicit canopy measurements with frozen visual embeddings improve prediction of individual-plant seasonal harvested fresh-pod mass relative to either approach alone?
• How do single-date measurements compare with repeated imaging, growth descriptors, and seasonal aggregation under the same validation procedure?
• How sensitive are canopy measurements and predictions to overlap, viewing angle, background, scaling, and segmentation errors?
• Which images are available at the intended prediction time, and do they precede all harvests included in the target yield?
• How does performance change when holding out entire plants or rows, with all repeated images from each plant kept together?
• Are preprocessing, feature selection, and tuning confined to the training data used in each evaluation split?
• How uncertain are the reported errors given the number of independent plants and growing conditions?
• How well do the models transfer to an independent growing batch, season, cultivar, or site?

Free Appropriate Sustainable Technology Research Group

Justin Jinchuan Zheng