PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning in agriculture: A survey

Andreas Kamilaris, Francesc X. Prenafeta-Boldu

arXiv:1807.11809v1cs.LGcs.CVstat.ML

TL;DR

Agricultural challenges have prompted interest in deep learning, but evidence across applications is dispersed. This survey examines 40 agricultural studies and finds superior performance in most comparisons.

  • Problem

    Agricultural challenges motivate examining deep learning applications in smart farming.

  • Method

    The paper surveys 40 relevant deep learning studies in agriculture and examines their problems, models, data, preprocessing, and performance.

  • Results

    Deep learning offers superior performance in the vast majority of related comparisons.

  • Takeaways & Limitations

    Deep learning outperforms other commonly used image-processing techniques across the surveyed agricultural applications.

  • Takeaways & Limitations

    Deep learning requires large datasets, creating a considerable barrier to its use.

Abstract

from arXiv · show

Deep learning constitutes a recent, modern technique for image processing and data analysis, with promising results and large potential. As deep learning has been successfully applied in various domains, it has recently entered also the domain of agriculture. In this paper, we perform a survey of 40 research efforts that employ deep learning techniques, applied to various agricultural and food production challenges. We examine the particular agricultural problems under study, the specific models and frameworks employed, the sources, nature and pre-processing of data used, and the overall performance achieved according to the metrics used at each work under study. Moreover, we study comparisons of deep learning with other existing popular techniques, in respect to differences in classification or regression performance. Our findings indicate that deep learning provides high accuracy, outperforming existing commonly used image processing techniques.

1. Introduction

Agricultural challenges require continuous analysis of complex ecosystems and large-scale, image-rich observations. This survey presents deep learning as a promising agricultural technique and reviews at least 40 efforts, claiming to be the first survey focused on the domain.

  • Motivation: Smart farming addresses agricultural productivity, environmental impact, food security, sustainability, nutrition, and ecosystem protection.These challenges require continuous monitoring, measurement, and analysis of complex, multivariate, and unpredictable agricultural ecosystems.
  • Agricultural imaging: Remote sensing enables systematic, non-destructive information collection across large geographical areas using satellites, airplanes, and UAVs.It provides wide-view snapshots of agricultural environments.
  • Agricultural imaging: Images form a large subset of remote-sensing data and support agricultural image identification, classification, and anomaly detection.Consequently, imaging analysis is an important agricultural research area.
  • Deep learning: Deep learning uses deeper convolutional neural networks to build hierarchical data representations, enabling greater learning capabilities, performance, and precision.It is presented as a recently gaining machine-learning technique similar to artificial neural networks.
  • Survey contribution: At least 40 research efforts employ deep learning for agricultural problems with very good results, motivating this survey as the authors’ first agricultural-domain survey.The authors distinguish it from more general deep-learning surveys covering other domains.

2. Methodology

The survey used a two-step methodology: systematic literature collection followed by detailed analysis of 40 selected papers on deep learning in agriculture and food production.

  • Literature collection: A keyword-based search covered conference papers and journal articles in IEEE Xplore, ScienceDirect, Web of Science, and Google Scholar.The query combined “deep learning” with “agriculture” or “farming” to exclude work unrelated to agricultural applications.
  • Literature selection: 47 papers were initially identified, then reduced to 40 based on appropriate deep-learning application and meaningful findings.Papers without results or performance metrics indicating technique success were excluded.
  • Paper analysis: The selected papers were reviewed individually to characterize agricultural problems, deep-learning models, data sources and types, classes and labels, preprocessing, and augmentation.The review also examined label variations and differences observed by the authors.
  • Paper analysis: The analysis assessed overall performance by adopted metric, evaluation on different datasets, and comparisons with other techniques.For comparisons, the researchers recorded which techniques were used and the difference in performance.

3. Deep Learning

Deep learning extends machine learning through hierarchical representations and automatic feature learning, enabling flexible classification and prediction across complex agricultural data challenges. Its CNNs, components, advantages, and limitations frame its application in the surveyed studies.

  • Deep Learning: Deep learning adds model depth and hierarchical abstraction while automatically learning features from raw data through compositions of lower-level features.This reduces reliance on manually engineered features that are costly, time-consuming, expert-dependent, and poorly generalizable.
  • Deep Learning: Deep learning models perform classification and prediction across complex challenges, provided adequately large datasets are available.Their complex models also allow massive parallelization, while classification accuracy can increase and regression error can decrease.
  • Convolutional Neural Networks: CNN processing made disease-indicating elements in leaf images increasingly evident, especially at the final Pool5 step.The visualization followed convolution, pooling, and normalization operations in a plant-disease identification problem.
  • Advantages and Limitations: Deep learning generally requires longer training but offers faster testing than other machine-learning methods.Additional disadvantages include difficulties with small or substantially different datasets, optimization challenges from model complexity, and hardware restrictions.

4. Deep Learning Applications in Agriculture

The survey identifies 40 deep-learning studies across 16 agricultural areas, with weeds, land-cover classification, plant recognition, fruit counting, and crop-type classification among the most common applications. Most studies use image data and report strong performance, although limited train/test variation constrains confidence in many results.

  • Data and preprocessing: Most studies use large image datasets, sometimes containing thousands of real or synthetic images, while 36 papers (90%) perform image pre-processing.Datasets may also originate from well-known, publicly available sources.
  • Performance: CA is the most common metric (24 papers, 60%), followed by F1 (10 papers, 25%); 19 of 24 CA studies report accuracy above 90%.Five F1 studies exceed 0.90, with the highest values above 0.99.
  • Generalization: Only 8 papers (20%) test on datasets different from training data, while 27 make no training/testing changes, lowering confidence in reported results.Generalization tests include simulated-to-real data, later plant growth stages, different fields, fruits, trees, rooms, and imaging conditions.

5. Discussion

The survey finds that deep learning generally outperforms traditional approaches in within-study comparisons, while highlighting advantages in feature extraction, generalization, and agricultural applicability. Its adoption remains constrained by data requirements, annotation and preprocessing burdens, limited dataset coverage, and weak generalization beyond training distributions.

  • Comparative performance: 28 out of 40 papers (70%) performed direct, valid comparisons, and deep learning outperformed traditional approaches in the vast majority of related work.Comparisons were limited to techniques evaluated under the same paper-specific experimental conditions because datasets, preprocessing, metrics, models, and parameters differed across papers.
  • Feature extraction: Deep learning’s automatic feature extraction was more effective than manual approaches, and a combined CNN+LSTM model achieved 25% higher CA than an LSTM using hand-crafted descriptors.The survey specifically contrasts deep learning with SVM, RF, ANN, LMC, SIFT, GLCM, histogram, area-, statistical-, texture-, color-, and shape-based methods.
  • Agricultural applications: Performance was especially strong for leaf and plant disease detection, plant recognition, and fruit counting, with reported values of CA > 95%, F1 > 0.92, or RFC > 0.9.Land-cover classification, crop-type classification, plant-phenology recognition, and weed detection showed more average performance, reported as CA < 87% or F1 < 0.8.
  • Generalization and efficiency: Deep learning generalized well in several studies, with only small performance reductions on different datasets, and enabled simulated datasets and fast implementations.However, models could not generalize beyond the dataset’s expressiveness, with failures involving unseen fruit colors, smaller objects, and heavily occluded or distant objects.
  • Limitations: A major barrier is the need for at least some hundreds of images, while agricultural datasets are scarce, costly to create, and often require expert annotation.Additional problems include low class variation, sensor noise, occlusion, clustering, high-dimensional hyperspectral data, limited training samples, and datasets that incompletely represent the target problem.

6. Conclusion

The survey examined 40 deep-learning studies across agricultural problems, models, data, preprocessing, augmentation, and performance. It found that deep learning outperformed popular image-processing techniques and supports further use in smarter, more sustainable farming and secure food production.

  • 6. Conclusion: 40 relevant papers were surveyed across agricultural problems, models, data sources, preprocessing, augmentation, and performance metrics.The survey also examined comparisons with other existing techniques in terms of performance.
  • 6. Conclusion: Deep learning offered better performance and outperformed other popular image-processing techniques.
  • 6. Conclusion: Future work will apply the survey’s deep-learning concepts and best practices to underused agricultural areas identified in the discussion.
  • 6. Conclusion: The survey aims to motivate researchers to apply deep learning to agricultural classification, prediction, computer vision, image analysis, and data analysis.
  • 6. Conclusion: Deep learning’s encouraging benefits support further use toward smarter, more sustainable farming and more secure food production.

Appendix I: Applications of computer vision in agriculture and popular techniques used.

The appendix organizes computer-vision applications in agriculture around remote sensing, including soil and vegetation/crop mapping and leaf area index and crop canopy analysis. It lists imaging modalities and popular analytical techniques used for these applications.

  • Remote sensing applications use hyperspectral imaging from satellite and airborne platforms, alongside satellite multi-spectral imaging and synthetic aperture radar.
  • Listed techniques include image fusion, SVM, decision trees, logistic regression, ANN, NDVI, and several polarization- and classification-based methods.The methods also include end-member extraction, co-polarized phase differences, linear polarizations, distance-based classification, linear mixing models, and related approaches.
  • Soil and vegetation/crop mapping is identified as a primary agricultural application.
  • Mapping datasets include airborne hyperspectral and multi-spectral imaging.
  • Leaf area index and crop canopy analysis uses linear regression analysis and NDVI.

3. Crop phenology Satellite remote sensing

Satellite remote sensing for crop phenology uses diverse sensing modalities, including LIDAR, hyperspectral and multispectral imaging, SAR, red-edge cameras, and thermal infrared. Applications include crop-height and yield estimation, fertilizer-effect and biomass assessment, using vegetation indices, regression, ANN, and other analytical methods.

  • Analytical methods: Analytical approaches include wavelet-based filtering, Fourier transforms, NDVI, linear and exponential regression, linear polarizations, and ANN.
  • Sensing modalities: Remote-sensing modalities include LIDAR, hyperspectral and multispectral imaging, SAR, red-edge cameras, and thermal infrared.
  • Applications: Crop-phenology applications include crop-height, yield, fertilizer-effect, and biomass estimation.
  • Analytical methods: Additional methods include DISCRIM feature extraction, PPD, multiple linear polarizations, and classification and regression tree analysis.

5. Crop monitoring

Crop monitoring applies remote sensing and imaging to seed identification, soil and leaf assessment, water-stress detection, erosion assessment, and pest management. The surveyed approaches use hyperspectral, multispectral, thermal, optical, radar, and near-infrared data alongside classical image-processing, regression, classification, and machine-learning methods.

  • 5. Crop monitoring: Crop-monitoring targets include seed identification, species reorganization, soil and leaf nitrogen treatment, salinity detection, and plant water stress or drought conditions.These applications draw on remote sensing, cameras, photo-detectors, hyperspectral or multispectral imaging, and thermal imaging.
  • 5. Crop monitoring: Classical analytical techniques across crop-monitoring tasks include principal component analysis, feature extraction, linear and exponential regression, decision trees, unsupervised clustering, density slicing, and NDVI.One listed workflow also uses the Fraunhofer Line Depth principle, contour tracing, and linear polarizations HH and VV.
  • 5. Crop monitoring: Satellite crop monitoring uses hyperspectral, multispectral, radar, optical, thermal-infrared, near-infrared, and red-edge imagery for crop and plant-condition assessment.The passages also identify SAR and satellite optical and radar images among the sensing modalities.
  • 10. Water erosion assessment: Water-erosion assessment uses hyperspectral and multispectral imaging, microwave remote sensing, thermal cameras, sample-image processing, regression, statistical analysis, CEM nonlinear signal processing, and NDVI.The listed processing methods include linear and exponential regression analysis and NDVI.
  • 11. Pest detection and management: Pest detection and management combines optical, hyperspectral, and multispectral remote sensing with clustering, Bayes classification, feature extraction, wavelets, Gabor filtering, genetic algorithms, fuzzy techniques, and neural networks.Additional methods include erosion and dilation segmentation, logistic regression, edge and color detection, and principal component analysis.

12. Weed detection

The weed-detection section references remote sensing, optical cameras, and photo-detectors, alongside fuzzy techniques and discriminant analysis.

  • 12. Weed detection: Remote sensing, optical cameras, and photo-detectors are referenced for weed detection.
  • 12. Weed detection: Fuzzy techniques and discriminant analysis are also referenced as methods.

13. Herbicide

Herbicide-related agricultural analysis used optical cameras and photo-detectors to capture monochrome images under different illumination conditions. Processing and analysis incorporated clustering, image fusion, color histograms, machine learning, discriminant analysis, and Bayesian filtering.

  • Optical cameras and photo-detectors captured monochrome images under different illumination conditions.
  • Image processing included K-means clustering, image fusion, and color histogram techniques.
  • Analysis methods included machine learning, especially SVM, alongside Bayesian and linear discriminant approaches.

14. Fruit grading

Fruit-grading research addresses packaged food and food-product inspection, including contaminant, disease, defect, and bruise detection. The surveyed approaches use diverse imaging modalities and conventional image-analysis methods.

  • Fruit grading: Methods span 3D vision, pattern recognition, clustering, SVM, discriminant analysis, classification trees, nearest neighbors, feature extraction, and thermography.
  • Fruit grading: Packaged food and food products are inspected for contaminants, diseases, defects, and bruises.
  • Fruit grading: Imaging inputs include X-ray or transmitted light, CCD cameras, thermal cameras, and multispectral or hyperspectral NIR imaging.
  • Fruit grading: Additional sensing approaches include multispectral imaging and polarimetric radar imagery.

Appendix II: Applications of deep learning in agriculture. · Appendix III: Publicly-available datasets related to agriculture.

Appendix II catalogs agricultural deep-learning applications spanning plant, crop, soil, fruit, weed, obstacle, and yield tasks, using diverse image and remote-sensing datasets. Reported results include strong classification, segmentation, detection, and regression performance, alongside recurring confusion and occlusion limitations.

  • Appendix II: Applications of deep learning in agriculture.: Deep-learning applications cover plant-species and disease recognition, crop and land-cover classification, phenology, root identification, yield estimation, fruit detection, and agricultural obstacle detection.Tasks use leaf, aerial, UAV, hyperspectral, multispectral, X-ray tomography, and time-series imagery.
  • Appendix II: Applications of deep learning in agriculture.: Reported limitations include poor generalization from hand-crafted features, confusion between visually similar classes, and reduced robustness under occlusion or uncontrolled conditions.Examples include confusion between cultivated and forest land, maize classes, fruit and foliage, and overlapping weeds and wheat.
  • Appendix II: Applications of deep learning in agriculture.: 75.34% (CA), 0.7463 (F1) on the first dataset and 84.61% (CA), 0.8441 (F1) on the second dataset exceeded RF and SVM results.The RF and SVM baseline achieved 74.20% (CA), 0.7158 (F1) and 83.82% (CA), 0.8274 (F1), respectively.
  • Appendix II: Applications of deep learning in agriculture.: 0.968 (RFC), 13.8 (L2) for oranges and 0.913 (RFC), 10.5 (L2) for apples were reported for fruit-count estimation.The ZF network achieved 0.892 (apples), 0.876 (mango), and 0.726 (almonds).
  • Appendix II: Applications of deep learning in agriculture.: 99.9% in row crops and 90.8% in grass were reported for detecting distant, heavily occluded, and unknown obstacles.The task used agricultural field imagery containing row crops, grass, and mowing contexts.
  • Appendix II: Applications of deep learning in agriculture.: 94% CA, 0.71 IoU (crops), 0.70 IoU (weeds) and 0.93 IoU (soil) were achieved in crop, weed, and soil segmentation.The application targeted sugar beet plants and weeds in field image data.
  • Appendix II: Applications of deep learning in agriculture.: 18% reduction in RMSE was reported for the multi-layer perceptron MCA model in crop-yield estimation.The application estimated county-level corn yield using historical yields and MODIS Enhanced Vegetation Index.
Loading 1807.11809v1…