Soil type identification model using a hybrid computer vision and machine learning approach

Soil type identification model using a hybrid computer vision and machine learning approach

연구 분야: Artificial Intelligence

학회: Multimedia Tools and Applications

Computer vision and its technologies are being used in the area of agricultural automation to identify, locate, and track targets for further image processing. Mostly, agricultural production has been highly dependent on natural resources like soil, water, and other related natural minerals from the soil. Soil classification is a way of arranging soils that have similar characteristics into groups. Identifying and classifying soils has a great role to play in agricultural productivity as it helps to provide relevant information which aids agricultural experts to recommend the type of crop best suited for a specific type of soil. This study mainly concentrated on classifying soil types such as clay soil, loam soil, sandy soil, peat soil, silt soil, and chalk soil. The soil images were collected from Amhara region at different locations by using a sony digital camera. To reduce image noise due to handshake we used a camera stand or arm to avoid other types of noises like environmental lighting effects and shadow. Once the dataset was collected, preprocessing such as resizing and gamma correction was performed to remove noise from the images, and contrast adjustment was also performed. Experimental research was applied as a general methodology and the experiment was conducted based on two approaches. The first approach used CNN as an end-to-end classifier and the second used a hybrid approach which used CNN as a feature extractor and SVM as a classifier. When CNN was used as an end-to-end classifier, a classification accuracy of 88% was achieved, whereas when the hybrid approach which used CNN as feature extractor and SVM as classifier was employed, a classification accuracy of 95% was achieved. Finally, we conclude that the hybrid approach is better than that of the End-to-End classification using our proposed model.