Application of machine vision for classification of soil aggregate size

• Texture-based method shows good performance in determining soil tillage quality.
• This approach is more useful in secondary tillage than primary tillage.
• The height of 60 cm was the most appropriate camera height.

Tillage operations demand more than half of the total energy consumed in mechanized agriculture. Simultaneous measurement of tillage quality during the operation, would present the possibility of real time adjustments of the tillage tool parameters. The development of such a method would result in a desirable plough with the least possible running cost. On that basis, the purpose of this study was to develop an algorithm that supplies the potential of real-time measurement of tillage quality using image processing. Photography was performed at three camera heights and covering nine different sizes of soil aggregates. Textural information from tilled soil images was extracted by four methods, including first order statistics of image histogram, gray level co-occurrence matrix; gray level run length matrix and local binary pattern. A data mining procedure by CfsSubsetEval was used for feature selection. Networks with topology of 19-19-1, 14-22-1, and 17-20-1 neurons represented the best classification performance for photography heights of 60, 80, and 100 cm, respectively. The best overall accuracy of the ANN classifier was obtained from images taken at the height of 60 cm (72.04%). Results indicated that the present approach for estimating mean weight diameter up to about 35 mm had the best performance with an accuracy of over 80%. The technique suggested in this study is feasible for implementation in variable rate secondary tillage machines.