Automation in Agriculture by Machine and Deep Learning Techniques: A Review of Recent Developments

• Published in: Precision agriculture Vol. 22; no. 6; pp. 2053 - 2091
• Main Authors: Saleem, Muhammad Hammad, Potgieter, Johan, Arif, Khalid Mahmood
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2021; Springer Nature B.V
• Subjects: Agricultural equipment; Agricultural land; Agriculture; Algorithms; Artificial intelligence; Artificial neural networks; Atmospheric Sciences; Automation; Biomedical and Life Sciences; Chemistry and Earth Sciences; Classification; Computer Science; Crop diseases; Crops; Deep learning; Discrimination; Disease detection; Feature extraction; fruits; Land cover; Learning algorithms; Life Sciences; Machine learning; Multilayers; Neural networks; Physics; Plant diseases; precision agriculture; Remote Sensing/Photogrammetry; Review; robots; Soil Science & Conservation; Statistics for Engineering; Support vector machines; weeds; Deep learning; Fruit harvesting; Plant disease detection; Agricultural robotics; Convolutional neural network; Machine learning
• DOI: 10.1007/s11119-021-09806-x
• ISSN: 1385-2256

Recently, agriculture has gained much attention regarding automation by artificial intelligence techniques and robotic systems. Particularly, with the advancements in machine learning (ML) concepts, significant improvements have been observed in agricultural tasks. The ability of automatic feature extraction creates an adaptive nature in deep learning (DL), specifically convolutional neural networks to achieve human-level accuracy in various agricultural applications, prominent among which are plant disease detection and classification, weed/crop discrimination, fruit counting, land cover classification, and crop/plant recognition. This review presents the performance of recent uses in agricultural robots by the implementation of ML and DL algorithms/architectures during the last decade. Performance plots are drawn to study the effectiveness of deep learning over traditional machine learning models for certain agricultural operations. The analysis of prominent studies highlighted that the DL-based models, like RCNN (Region-based Convolutional Neural Network), achieve a higher plant disease/pest detection rate (82.51%) than the well-known ML algorithms, including Multi-Layer Perceptron (64.9%) and K-nearest Neighbour (63.76%). The famous DL architecture named ResNet-18 attained more accurate Area Under the Curve (94.84%), and outperformed ML-based techniques, including Random Forest (RF) (70.16%) and Support Vector Machine (SVM) (60.6%), for crop/weed discrimination. Another DL model called FCN (Fully Convolutional Networks) recorded higher accuracy (83.9%) than SVM (67.6%) and RF (65.6%) algorithms for the classification of agricultural land covers. Finally, some important research gaps from the previous studies and innovative future directions are also noted to help propel automation in agriculture up to the next level.