Deep learning in agriculture: A survey

Motivation

• Smart farming addresses agricultural challenges involving productivity, environmental impact, food security, sustainability, nutritional quality, and ecosystem protection.These challenges require continuously monitoring, measuring, and analyzing complex, multivariate, and unpredictable agricultural ecosystems.Source: §1 · p. 2

Agricultural imaging

• Remote sensing provides systematic, non-destructive information over large agricultural areas, while images capture agricultural environments and support identification, classification, and anomaly detection.Remote sensing uses satellites, airplanes, and UAVs; imaging analysis is therefore presented as an important agricultural research area.Source: §1 · p. 2

Deep learning

• Deep learning uses deeper convolutional neural networks to form hierarchical data representations, enabling larger learning capabilities and higher performance and precision than conventional approaches.The introduction places deep learning within machine learning and describes it as similar to artificial neural networks.Source: §1 · p. 3

Survey contribution

• The survey is motivated by deep learning’s recent growth and promise in agriculture, with at least 40 research efforts reporting very good results across agricultural problems.The authors state that, to their knowledge, this is the first survey dedicated to deep learning in agriculture, whereas existing surveys are more general and cover other domains, (Wan, et al., 2014), (Najafabadi, et al., 2015).Source: §1 · p. 3

Literature selection

• The search covered IEEE Xplore, ScienceDirect, Web of Science, and Google Scholar using the query ["deep learning"] AND ["agriculture" OR ”farming"].The query filtered out papers referring to deep learning without applying it to agriculture.Source: §2 · p. 4
• 47 papers were initially identified, then reduced to 40 after restricting the corpus to studies with appropriate deep-learning applications and meaningful findings.Papers claiming deep learning without results or performance metrics indicating success were excluded.Source: §2 · p. 4

Study analysis

• Each selected paper was analyzed for its agricultural problem, deep-learning models, data sources and types, labels, preprocessing, augmentation, performance, datasets, and comparisons with other techniques.The analysis explicitly examined whether models were tested on different datasets and how performance differed from other approaches.Source: §2 · p. 4

Reporting

• The review’s main findings were presented in Section 4, while detailed information for each paper was listed in Appendix II.The methodology therefore separated synthesized findings from per-paper documentation.Source: §2 · pp. 4–5

3. Deep Learning

• Deep learning automatically extracts higher-level features by composing lower-level representations, reducing reliance on costly, time-consuming feature engineering.Feature engineering depends on expert knowledge, must change with datasets or problems, and generalizes poorly, whereas deep learning locates important features during training.Source: §3 · pp. 5, 7
• Deep learning models can improve classification accuracy or reduce regression error when adequately large datasets describe the problem.Their hierarchical structure and learning capacity support flexible classification and prediction across complex data-analysis challenges.Source: §3 · p. 5
• CNNs were the most common surveyed technique, appearing in 17 papers (42%).Source: §3 · p. 6
• CNN convolutional layers extract spatially context-aware features, pooling reduces dimensionality, and fully connected layers classify images or produce numerical predictions.In plant-disease identification, disease-revealing image elements become more evident through processing, especially at the final Pool5 step.Source: §3 · p. 6
• Deep learning generally requires longer training but offers faster testing than other machine-learning methods, while small or mismatched datasets, optimization complexity and hardware restrictions remain disadvantages.Source: §3 · p. 7

3.1 Available Architectures, Datasets and Tools

• Available architectures include AlexNet, CaffeNet, VGG, GoogleNet and Inception-ResNet, with pretrained weights and datasets such as ImageNet and PASCAL VOC supporting model development.Tools including Theano, TensorFlow, Keras, Caffe, PyTorch, TFLearn, Pylearn2 and Matlab Toolbox enable experimentation with deep learning.Source: §3.1 · p. 8

4. Deep Learning Applications in Agriculture

• The reviewed works catalogued agricultural problems, deep-learning architectures, data classes and labels, preprocessing or augmentation, performance metrics, and comparisons with other techniques.The 40 identified works are listed in Appendix II with these characteristics.Source: §4 · pp. 8–9

4.1 Areas of Use

• Deep learning applications span 16 agricultural areas, led by weed identification (5 papers), land-cover classification (4), plant recognition (4), fruit counting (4), and crop-type classification (4).Most studies target crops (20 papers), followed by weed detection (8), land cover (4), livestock agriculture (3), obstacle detection (3), and weather prediction (1).Source: §4.1 · p. 9

4.2 Data Sources

• Large image datasets, sometimes containing thousands of images, were common, while land-cover, crop-type, yield-estimation, and some weed studies used only tens of UAV, airborne, or satellite images.Other inputs included X-ray tomography images and text from repositories or field sensors; more complex problems generally required more training data.Source: §4.2 · pp. 9–10

4.3 Data Variation

• Nineteen papers (47%) revealed some variation among classes; this low variation affected classification accuracy significantly, i.e. more than 5%, except from (Minh, et al., 2017).Reported weed-identification issues included lighting, resolution, and soil.Source: §4.3 · pp. 11–12

4.4 Data Pre-Processing

• 36 papers (90%) used image preprocessing; resizing was most common (16 papers), followed by segmentation (12 papers).Operations also included background or non-green pixel removal, bounding boxes, grayscale or HSV conversion, and extracted image features.Source: §4.4 · p. 12
• Preprocessing varied by data source, including histograms, PCA filters, wavelet transformations, GLCM features, orthorectification, calibration, terrain correction, and atmospheric correction.Satellite and aerial imagery particularly required combinations of geometric, calibration, terrain, and atmospheric corrections.Source: §4.4 · p. 13

4.5 Data Augmentation

• Data augmentation appeared in 15 papers (37%), artificially enlarging training-image datasets to improve learning procedure, performance, and generalization.Label-preserving transformations included rotations (12 papers), partitioning/cropping (3 papers), scaling (3 papers), and transposing.Source: §4.5 · p. 13

4.6 Technical Details

• The studies employed custom CNNs, DRNNs, LSTMs, DBNs, and a PCA–auto-encoder hybrid, with Caffe the most popular framework in 13 papers.Source: §4.6 · p. 14

4.7 Outputs

• Classification studies used 2 to 1,000 classes, while other studies predicted fruit counts, fruit regions, animal growth, weather, crop yield, or soil moisture.Outputs for classification assigned probabilities to classes and selected the highest-probability class; non-classification outputs included scalar values or bounding boxes.Source: §4.7 · p. 16

4.8 Performance Metrics

• Performance evaluation used multiple task-specific metrics, with Table 1 listing their definitions and survey symbols; unspecified uses of “accuracy” were treated as classification accuracy (CA).The survey defines “DL performance” as a score on one of the metrics listed in Table 1.Source: §4.8 · p. 16, Table 1 · p. 17
• CA was the most common metric, used in 24 papers (60%), followed by F1 in 10 papers (25%); IoU, RMSE, and RFC were also reported.IoU thresholds determined positive predictions in bounding-box tasks, with observed thresholds of 20%, 40%, and 50%; metric trade-offs can occur, such as prioritizing weed recall over precision.Source: No. Performance · p. 18

4.9 Overall Performance

• Across the reviewed papers, 19 of 24 papers using CA reported accuracy above 90%, while 5 of 10 using F1 exceeded 0.90, with highest F1 values above 0.99.CA values reached 98% or more in several works. Cross-paper comparison is difficult because different metrics are used.Source: §4.9 · pp. 18–19

4.10 Generalizations on Different Datasets

• Among 40 papers, only 8 (20%) tested their solutions on datasets different from those used for training, including simulated-to-real, later-growth-stage, and different-field evaluations.Two approaches trained on simulated data and tested on real data, while two tested 2–4 weeks later at a more advanced plant growth stage.Source: §4.10 · pp. 19–20

4.11 Performance Comparison with Other Approaches

• In almost all comparisons between techniques evaluated in the same paper, dataset, and metric, deep-learning models outperformed alternatives; CNN achieved 1–8% higher CA than SVM and 41% higher CA than ANN.Source: §4.11 · p. 21
• CNNs achieved 3-11% higher CA than unsupervised feature learning, 2-44% improved CA over local shape/color features, and 2% better CA or 18% less RMSE than multilayer perceptrons.These comparisons were reported by Luus et al. (2015), Dyrmann et al. (2016), Sørensen et al. (2017), Kussul et al. (2017), and Song et al. (2016).Source: §4.11 · p. 21
• LSTM models achieved 1% higher CA than RF and SVM, 44% improved CA than SVM, and 7-9% better CA than RF and SVM.These results were reported in Ienco, Gaetano, Dupaquier, & Maurel (2017), Rußwurm & Körner (2017), and Minh et al. (2017).Source: §4.11 · p. 22
• Deep learning performed worse in only one reported comparison: a CNN using local descriptors with KNN had 20% worse LC.The comparison involved local descriptors for image representation and KNN classification (Reyes, Caicedo, & Koike, 2015).Source: §4.11 · p. 22

5. Discussion

• Deep learning outperformed traditional approaches in the vast majority of valid within-paper comparisons, including SVM, RF, ANN, and LMC classifiers.Only 28 of 40 papers (70%) performed direct, valid comparisons under the same experimental conditions.Source: §5 · pp. 22–23
• CNN+LSTM achieved 25% higher CA than an LSTM using handcrafted feature descriptors, supporting the effectiveness of automatic feature extraction.The survey attributes this advantage to learned features being more effective than manually engineered descriptors such as SIFT, GLCM, histogram, texture, color, and shape features.Source: §5 · p. 23
• Deep learning was also applied to field sensory data and environmental variables, indicating potential across agricultural problems beyond image-based tasks.The survey identifies five related works using field sensory data or combinations of static and dynamic environmental variables.Source: §5 · p. 23
• Leaf and plant disease detection, plant recognition, and fruit counting reached CA > 95%, F1 > 0.92, or RFC > 0.9, whereas land-cover, crop-type, phenology, and weed tasks showed CA < 87% or F1 < 0.8.The stronger results are associated with available datasets and distinct visual characteristics, while weaker results reflect occlusion, noisy satellite imagery, low crop variation, and gradual phenological change.Source: §5 · pp. 23–24

5.1 Advanced Deep Learning Applications

• RNN-based models captured temporal information unavailable to simple CNNs, producing a 16% CA improvement over CNNs in the reported comparisons.LSTM-based approaches also performed well for low-represented and difficult classes, while temporal models benefited from information across time series.Source: §5.1 · pp. 25–26, Table 2 · p. 26
• Advanced models addressed localization, temporal prediction, and real-time deployment, with a lightweight implementation sacrificing only 3.90% CA while processing 40.6 times more pixels per second.Applications included fruit and weed bounding-box detection, recurrent modeling of agricultural time series, and teacher–student networks for faster inference.Source: §5.1 · pp. 25–26

5.2 Advantages of Deep Learning

• Deep learning automates feature extraction, reducing the considerable effort required to hand-engineer features for agricultural prediction and classification.Manual feature extraction was infeasible for crop-yield estimation (Kuwata & Shibasaki, 2015) and soil-moisture estimation (Song, et al., 2016).Source: §5.2 · pp. 26–27
• Deep learning models generalized across dataset changes and difficult visual conditions, including illumination, occlusion, scale, orientation, variation, and complex backgrounds.Fruit-counting models transferred across circular fruit types, while DeepAnomaly detected distant, heavily occluded, and unknown objects (Christiansen, Nielsen, Steen, Jørgensen, & Karstoft, 2016).Source: §5.2 · p. 27
• Across eight studies using different testing datasets, performance generally remained high with only small reductions relative to same-dataset testing.Although training could take longer than SVM or KNN, testing was faster after training (Christiansen, Nielsen, Steen, Jørgensen, & Karstoft, 2016).Source: §5.2 · p. 27
• Simulated datasets can be designed for real-world problems, enabling weed–maize discrimination under overlapping conditions (Dyrmann, Mortensen, Midtiby, & Jørgensen, 2016).The simulation used top-down images of overlapping plants on soil backgrounds.Source: §5.2 · pp. 27–28

5.3 Disadvantages and Limitations of Deep Learning

• Deep learning requires large, diverse datasets, with at least some hundreds of images often needed despite data augmentation, while class similarity and noisy or occluded inputs hinder training.Agricultural datasets may contain low-resolution imagery, sensor inaccuracies, crop occlusions, overlapping plants, and clustering (Song, et al., 2016).Source: §5.3 · p. 28
• Annotation is complex, often requires scarce expertise, and can contain labeling errors, especially for fruit counting (Chen, et al., 2017), and weed identification (Sørensen, Rasmussen, Nielsen, & Jørgensen, 2017), (Dyrmann, Jørgensen, & Midtiby, 2017).Banana pathology also faces limited worldwide resources and expertise (Amara, Bouaziz, & Algergawy, 2017).Source: §5.3 · p. 28
• Models cannot generalize beyond their datasets’ expressiveness, failing when real-world disease views, fruit colors, object sizes, distance, or occlusion differ from training conditions.Examples include failure to count green fruits, missed small objects (Sa, et al., 2016), and difficulty with distant or heavily occluded objects (Christiansen, Nielsen, Steen, Jørgensen, & Karstoft, 2016).Source: §5.3 · pp. 28–29

5.4 Future of Deep Learning in Agriculture

• Future research could apply deep learning to underexplored agricultural problems such as irrigation, water stress, pest detection, contaminants, crop damage, and greenhouse monitoring.Other proposed applications include drone-based seeding assessment, grape-maturity monitoring, and animal-welfare or disease monitoring.Source: §5.4 · pp. 29–30
• Publicly available datasets could support architecture testing, pre-training, and adaptation to specific agricultural challenges.Remote-sensing data with multi-temporal, multi-spectral, and multi-source imagery could support land- and crop-cover classification.Source: §5.4 · p. 30
• Future LSTM and other RNN models are expected to exploit the time dimension for higher-performance classification or prediction, including estimating growth, yield, water needs, and disease avoidance.Source: §5.4 · pp. 30–31
• More complex systems may combine deep-learning models, classifiers, and hand-crafted with automatically extracted features to improve overall outcomes.Researchers are also expected to evaluate generalization on more realistic datasets and adopt combined performance metrics for comparison.Source: §5.4 · p. 31
• Researchers should make their datasets publicly available so they can also be used by the general public.Source: §5.4 · p. 31
• The DRNN model predicted daily pig or chicken feed intake quite accurately across the growing period, helping farmers select growth curves while potentially supporting leg health and animal welfare (Demmers T. G., et al., 2010; Demmers T. G., Cao, Parsons, Gauss,).Maintaining final weight and total feed-intake requirements is presented as compatible with following beneficial growth patterns.Source: §5.4 · p. 32

• The survey identified 40 agricultural deep learning papers and examined their problems, models, data sources, preprocessing, augmentation, and reported performance metrics.It also compared deep learning with existing techniques in terms of performance.Source: §6 · p. 32
• Deep learning offers better performance than other popular image-processing techniques across the surveyed agricultural research.The comparison evaluated performance against existing techniques.Source: §6 · p. 32
• Future work will apply the survey’s deep learning concepts and best practices to agricultural areas where the technique remains inadequately used.Some target areas were identified in the discussion section.Source: §6 · p. 32
• The survey aims to motivate research applying deep learning to agricultural classification, prediction, computer vision, image analysis, and broader data analysis.The authors associate these applications with smarter, more sustainable farming and more secure food production.Source: §6 · p. 32

• Agricultural remote sensing uses satellite and airborne hyperspectral imaging, satellite multispectral imaging, and synthetic aperture radar.Source: Appendix I · p. 37
• Popular analysis techniques include image fusion, SVM, decision trees, regression, ANN, NDVI, polarization features, and distance-based classification.Source: Appendix I · p. 37
• Remote-sensing applications include soil and vegetation/crop mapping using airborne hyperspectral and multispectral imaging.Source: Appendix I · p. 37
• Leaf-area-index and crop-canopy applications use linear regression analysis and NDVI.Source: Appendix I · p. 37

• Preprocessing and feature extraction include wavelet-based filtering, Fourier transforms, NDVI, DISCRIM, and vegetation indices such as ICWSI.Source: §3 · p. 37
• Satellite remote sensing uses LIDAR, hyperspectral and multispectral imaging, SAR, red-edge cameras, and thermal infrared data for crop-phenology analysis.Source: §3 · p. 37
• Analyses estimate crop height, yields, fertilizer effects, and biomass using satellite and imaging data.Source: §3 · p. 37
• The reported analytical methods include linear and exponential regression, ANN, SAR polarization features, and classification and regression tree analysis.Source: §3 · p. 37

• Crop monitoring includes seed identification and species reorganization using remote sensing, cameras, photo-detectors, and hyperspectral imaging.Source: §5 · p. 37
• Soil and leaf nitrogen content, treatment, and salinity detection are addressed with hyperspectral, multispectral, and thermal imaging.Source: §5 · p. 37
• Plants’ water stress and drought conditions are monitored through satellite, hyperspectral, multispectral, radar, thermal, NIR, and red-edge imaging.Source: §5 · p. 37

10. Water erosion assessment

• Water erosion assessment combines hyperspectral, multispectral, microwave, and thermal imagery with regression, statistical analysis, CEM nonlinear signal processing, and NDVI.Source: §5 · pp. 37–38

11. Pest detection and management

• Pest detection and management uses clustering and Bayes classification, FFT and GLCM feature extraction, wavelets, Gabor filtering, genetic algorithms, fuzzy techniques, neural networks, segmentation, regression, and edge and color detection.Source: §5 · p. 38

12. Weed detection

• Remote sensing, optical cameras, and photo-detectors are identified as technologies relevant to weed detection.Source: §12 · p. 38
• Fuzzy techniques and discriminant analysis are identified as methods associated with weed detection.Source: §13 · p. 38

13. Herbicide

• Optical cameras and photo-detectors capture monochrome images under different illuminations for herbicide analysis.Source: §13 · p. 38
• The processing toolkit includes K-means clustering, image fusion, color histogram techniques, SVM-based machine learning, Bayesian methods, Bayes filtering, and linear discriminant analysis.Source: §13 · p. 38

• Methods span 3D vision, invariance and pattern recognition, PCA, K-means clustering, SVM, discriminant analysis, classification trees, nearest neighbors, decision trees, fusion, feature extraction, and pulsed-phase thermography.Source: §14 · p. 38
• Fruit grading targets packaged-food contamination, diseases, defects, and bruise detection.Source: §14 · p. 38
• The data include X-ray or transmitted-light imaging, CCD and thermal cameras, and multispectral, hyperspectral NIR-based, and monochrome images under varied illumination.Source: §14 · p. 38

Appendix II: Applications of deep learning in agriculture.

• Applications include classifying 32 plant species, distinguishing 13 diseases from healthy leaves, and identifying 14 crop species with 26 diseases.The associated datasets include Flavia’s 1,907 leaf images and PlantVillage’s 54,306 controlled-condition leaf images.Source: Area · pp. 39–40
• Remote-sensing applications classify cultivated land, land cover, and crop types using hyperspectral, Landsat, Sentinel, Pléiades, and UAV imagery, with classes ranging from 2 to 21.Reported cases include cultivated versus non-cultivated land and time-series crop classification.Source: Area · pp. 40–42
• 75.34% CA and 0.7463 F1 on the first dataset, versus 84.61% CA and 0.8441 F1 on the second, exceeded RF/SVM results of 74.20% and 83.82% CA.The corresponding RF/SVM F1 scores were 0.7158 and 0.8274.Source: Area · p. 41
• 92.7% for the multilayer perceptron versus 88% for RF in crop classification demonstrates a reported deep-learning advantage.The appendix also reports 96.30% CA and better results than SVM in another application.Source: Area · pp. 39, 41
• The appendix covers recognition of seven plant views, root and shoot feature identification and localisation, and 23 pixel classes comprising 22 crops plus soil; scans offer higher accuracy.Some forest-land samples were easily confused, with general confusion between maize.Source: Area · pp. 41–43
• Plant and fruit applications used visual classification or regression for 44 plant species, three bean species, root-versus-soil pixels, and sweet pepper, rock melon, apple, almond, and mango detection.The appendix also includes ISO barrel-shaped obstacle recognition and phenological-stage classification across several crops.Source: Area · pp. 43, 44, 46
• Phenotyping applications classified plant growth stages and four Arabidopsis accessions from visual sequences, but intermediate stage durations and gradually changing appearance complicated recognition.The datasets included sequences captured on different days during plant growth.Source: Area · p. 44
• Performance and dataset conditions varied substantially because soil-root contrast, occlusion, uncontrolled illumination, fruit-foliage similarity, camera setup, viewpoint, time, and location affected agricultural imagery.The appendix lists datasets ranging from INTA Argentina leaf images and Turkish agrostations to orchard imagery and Sentinel-1 time series.Source: Area · pp. 44, 46
• Agricultural time-series applications estimated county-level corn yield, mapped winter vegetation quality across five classes, and predicted tomato fruit counts from synthetic imagery.Vegetation-quality classification showed repeated confusion among classes, including misclassification between “low” and “bare.”Source: Area · p. 45
• 0.968 (RFC), 13.8 (L2) for oranges and 0.913 (RFC), 10.5 (L2) for apples were reported, while ZF network scores were 0.892, 0.876, and 0.726 for apples, mangoes, and almonds.These results concern fruit-counting or fruit-detection applications in orchard imagery.Source: Area · p. 46
• 0.807 was obtained with a Conditional Random Field modeling color and visual texture features, while the best texture-based regression model reached 0.682 (RFC).The reported models combine or compare visual cues for agricultural image analysis.Source: Area · p. 46
• The appendix also covered fine-grained agricultural classification, including 91 weed-seed types and 22 early-stage weed and crop species, where similarity, shape, texture, lighting, resolution, and soil differences complicated recognition.The corresponding datasets contained 3,980 images for weed seeds and 10,413 images for early-stage weed and crop species.Source: Area · p. 47
• Deep learning applications targeted weed segmentation for robotic platforms, weed detection under heavy leaf occlusion, and sugar-beet-versus-weed recognition from RGB, NIR, UAV, and robot imagery.Reported challenges included distant or heavily occluded obstacles, overlapping weeds and wheat, and changing field conditions.Source: Area · pp. 48–49
• 94% CA was achieved with 0.71 IoU for crops, 0.70 IoU for weeds, and 0.93 IoU for soil in crop–weed–soil detection.The task involved detecting and classifying weeds and maize in fields.Source: Area · p. 50
• An MLP-MCA model reduced RMSE by 18% for predicting soil moisture content over an irrigated corn field.The application also created soil-moisture maps through geospatial interpolation.Source: Area · p. 50
• Livestock applications addressed practical cattle identification across five races, pig-weight estimation, and broiler-chicken growth control using environmental and production measurements.The cattle dataset contained 1,300 images, while pig and broiler studies measured weight alongside housing, feed, temperature, humidity, or ventilation variables.Source: Area · p. 51
• CNN without extra inputs achieved 89.68%, compared with 90% for a Gaussian Mixture Model in chicken-weight estimation.The task was scalar estimation of chicken weight for broiler-growth control.Source: Area · p. 51
• The supplied Appendix II excerpt is incomplete and only contains a fragment concerning chickens.The fragment reads: “chicken than the ones which.”Source: Area · p. 51

Appendix III: Publicly-available datasets related to agriculture.

• Public datasets include plant imagery and disease resources such as Image-Net, ILSVRC, the University of Arcansas Plants Dataset, Plant Village, Leafsnap, LifeCLEF, MalayaKew, and Flavia.These resources cover varied plants, crop diseases, herbicide injury, leaves, plant identity and distribution, and 32 plant species.Source: Appendix III: Publicly-available… · p. 53
• Other listed datasets support animal recognition, soil mapping, land-use classification, and crop-versus-weed field analysis.Examples include PASCAL Visual Object Classes, AFSIS digital soil maps for sub-Saharan Africa, the UC Merced 21-class land-use dataset, and field images with segmentation masks and crop/weed annotations.Source: Appendix III: Publicly-available… · p. 53
• The catalog also includes datasets for sugar-beet plant classification, localization and mapping, and corn-hybrid analysis linked to weather and soil conditions.The corn data cover 2,267 hybrids in 2,122 locations between 2008 and 2016.Source: Appendix III: Publicly-available… · p. 53