Farmland boundary extraction based on the AttMobile-DeeplabV3+ network and least squares fitting of straight lines

Abstract

The rapid extraction of farmland boundaries is key to implementing autonomous operation of agricultural machinery. This study addresses the issue of incomplete farmland boundary segmentation in existing methods, proposing a method for obtaining farmland boundaries based on unmanned aerial vehicle (UAV) remote sensing images. The method is divided into two steps: boundary image acquisition and boundary line fitting. To acquire the boundary image, an improved semantic segmentation network, AttMobile-DeeplabV3+, is designed. Subsequently, a boundary tracing function is used to track the boundaries of the binary image. Lastly, the least squares method is used to obtain the fitted boundary line. The paper validates the method through experiments on both crop-covered and non-crop-covered farmland. Experimental results show that on crop-covered and non-crop-covered farmland, the network’s intersection over union (IoU) is 93.25% and 93.14%, respectively; the pixel accuracy (PA) for crop-covered farmland is 96.62%. The average vertical error and average angular error of the extracted boundary line are 0.039 and 1.473°, respectively. This research provides substantial and accurate data support, offering technical assistance for the positioning and path planning of autonomous agricultural machinery.