Low-altitude UAV Visible Light Remote Sensing Tobacco Identification Dataset for Complex Scenarios

To address the challenges of low resolution in traditional satellite remote sensing, inefficient ground surveys, and insufficient existing datasets for precision monitoring of tobacco in complex mountainous regions, this paper employs manual visual interpretation to delineate plant outlines for sample annotation. Following geometric correction, radiometric calibration, and image mosaicking, orthorectified imagery with a spatial resolution of 6.4 cm is generated, thereby constructing a precisely annotated semantic segmentation dataset tailored for tobacco. The dataset comprises five 5000×5000-pixel UAV remote sensing images, randomly segmented into an initial dataset (2300 samples) and an optimised dataset (9500 samples), each consisting of 224×224-pixel segments, alongside corresponding manually annotated labels in PNG format. Results indicate: Among eight scenario types, the highest accuracy (0.85 precision) was achieved in fragmented terrain without weeds, while the lowest accuracy (0.49 precision) occurred in flat plots with weeds. Whole-image recognition accuracy without scenario differentiation was 0.68, with significantly higher accuracy achieved after deconstructing complex scenarios. This dataset provides crucial data and methodological support for deep learning models to accurately identify surface crops in complex mountainous terrain, thereby enhancing precision agricultural decision-making.

针对传统卫星遥感分辨率偏低、地面调查效率低下，以及复杂山区烟草精准监测领域现有数据集不足的痛点，本文采用人工目视解译勾勒植株轮廓以完成样本标注。经几何校正、辐射定标与图像拼接后，本研究生成空间分辨率为6.4 cm的正射校正影像，进而构建出适配烟草种植场景的精准标注语义分割数据集。本数据集包含5张5000×5000像素的无人机（Unmanned Aerial Vehicle, UAV）遥感影像，经随机切分为初始数据集（2300个样本）与优化数据集（9500个样本），二者均由224×224像素的图像块组成，并附带PNG格式的人工标注标签。实验结果表明：在8种场景类型中，无杂草的破碎地形场景的精确率最高（0.85），有杂草的平坦地块场景精确率最低（0.49）；未区分场景的全图识别精确率为0.68，在对复杂场景进行解构后，识别精度得到显著提升。本数据集可为深度学习模型精准识别复杂山区地表作物提供关键的数据与方法支撑，助力提升农业精准决策水平。