Authors: Deshmukh Nishant Dhananjay, Professor Abhishek Popat Bangar, Professor S. B. Bhosale, Professor Karad V. A.
Abstract: Pests reduce both the yield and the quality of crops, and checking fields by hand is slow, tiring, and dependent on trained experts. Artificial intelligence (AI), particularly computer vision and deep learning, makes it possible to recognize pests from photographs and to respond to infestations earlier and with better precision. This review looks at AI-driven pest detection in terms of image classification, object detection, public benchmark datasets, model efficiency, and practical deployment. A focused narrative search of scholarly sources and publisher pages was carried out, covering work on the IP102 benchmark, the Pest24 dataset, machine-learning and deep-learning surveys, YOLObased pest detectors, and rice-pest detection. The literature shows that convolutional neural networks can learn the visual cues needed for pest classification, while object detectors from the YOLO family can find and label several pests within one image. The choice of dataset strongly shapes how trustworthy an evaluation is: IP102 holds more than 75,000 images in 102 categories, with bounding boxes for roughly 19,000 of them, whereas Pest24 consists of field images in which pests are small and often packed closely together. Newer model-specific studies report gains on the test sets they selected, but figures drawn from different datasets and evaluation protocols cannot be compared head to head. Challenges that keep recurring include tiny pests, look-alike species, imbalanced classes, cluttered field backgrounds, changing light, a shortage of labelled field data, and the limited computing power available on devices. The review points to the need for more representative field testing, open evaluation practices, efficient models, reporting of uncertainty, and stronger links between detection output and integrated pest management. AI-based detection is promising, yet a high benchmark score alone does not prove dependable performance on real farms. Future work should focus on external validation, ease of use for farmers, and responsible use of model predictions.
