In-field cotton detection via region-based semantic image segmentation

Abstract

Crop detection from the images taken in the natural field is not only a complex task, but also stands an important data for precision agriculture in obtaining crop growth information of field crops. This paper reports on a novel in-field cotton detection via region-based semantic image segmentation with two perspectives of observation, including unsupervised region generation and supervised semantic labeling prediction. First, simple linear iterative clustering (SLIC) and density-based spatial clustering of applications with noise (DBSCAN) on Wasserstein distance are employed to generate regions, with superiority in edge-preserving and density contrast distribution. Then histogram-based color and texture features extracted from each region proposal are passed to random forest, achieving semantic labeling prediction on in-field cotton images. Finally, to evaluate the robustness and accuracy of the proposed method on cotton detection, 46 test images taken from the year 2012 to 2015 are utilized to compare the proposed method with other well-established methodologies based on four metrics. Experiments and comparisons demonstrate that our method outperforms other mentioned methods with highest mean values and lowest standard deviations. Furthermore, the method can also detect the boll opening stage automatically, which provides support for precision agriculture. The dataset and source code will be made available online.