Growth Monitoring of Rapeseed Seedlings in Multiple Growth Stages Based on Low-Altitude Remote Sensing and Semantic Segmentation

Abstract

Growth monitoring of rapeseed seedlings can enable timely identification of problems, such as leakage of rapeseed sowing, uneven seedbeds, pests, and diseases, understanding their growth status, and taking emergency measures, which play an important role in improving the sowing strategy of rapeseed, decision-making of rapeseed fertilizer prescription maps, and improving economic efficiency. To address the current problems of low efficiency and high cost in monitoring rapeseed seedling growth, a method of rapeseed seedling multi-growth stage growth monitoring based on low-altitude remote sensing and semantic segmentation was proposed by considering the visible low-altitude remote sensing image of unmanned aerial vehicle (UAV) in early rapeseed seedlings (mainly including two-, four-, and six-leaf stages) as the research object. Rapeseed seedling growth was classified into three categories: excellent, average, and poor. First, the problem of difficult segmentation of rapeseed seedlings and furrows attributed to complex field scenes and dense rapeseed planting was addressed to determine the optimal rapeseed seedling and furrow segmentation model. The lightweight network MobileNetV2 was selected as the backbone feature extraction network based on the original DeeplabV3+ model to reduce the number of model parameters and the computational complexity. Furthermore, the coordinate attention (CA) mechanism module was integrated to improve the attention of the model to important features. The interference of small border areas was removed using the regional pixel area thresholding method. The maximum outer rectangle of each border area was filled according to the image width. The maximum outer rectangles within the same border were combined, and the maximum border centerline extraction algorithm was used to obtain the minimum outer rectangle of the largest border and identify the optimal centerline of each border. Finally, eight relevant eigenvalues were extracted by combining the characteristics of rapeseed seedling growth and inputted as eigenvectors into the random forest (RF) model to construct a multifactor rapeseed seedling growth monitoring model for multiple growth stages. The test results indicated that the improved DeeplabV3+ network outperformed the original DeeplabV3+, with the mean pixel accuracy (mPA) increasing from 78.43% to 87.47% (an improvement of 9.04%) and the average intersection over union (mIoU) increasing from 67.45% to 76.89% (an improvement of 9.44%). The mean positional deviation of the centerline was –5.29 pixels with a standard deviation of 9.51 and a mean angular deviation of –0.01848 rad with a standard deviation of 0.00791, which allowed for the effective detection of field drain centerlines. Meanwhile, the precision, sensitivity, specificity, and accuracy of the proposed algorithm for rapeseed seedling multi-growth stage growth monitoring were 96.35%, 96.34%, 97.20%, and 96.34%, respectively. The algorithm of this study can effectively detect rapeseed seedlings and field drain areas, obtain the best centerline of the field drain area, and be used for rapeseed seedling multi-growth stage growth monitoring, which provides a theoretical basis and technical reference for rapeseed seedling multi-growth stage growth monitoring.

Keywords: Unmanned Aerial Vehicle, Rapeseed seedlings, Growth monitoring, Image Processing, Semantic segmentation