| 董一波,刘立群.YOLO-LT的无人机航拍苹果果园小目标检测及产量预测方法[J].干旱地区农业研究,2025,(5):252~262 |
| YOLO-LT的无人机航拍苹果果园小目标检测及产量预测方法 |
| A small target detection and yield prediction method with YOLO-LT under UVA aerial photography in apple orchards |
| |
| DOI：10.7606/j.issn.1000-7601.2025.05.25 |
| 中文关键词: 苹果果园 小目标检测 产量预测 特征提取 YOLO-LT 无人机航拍 |
| 英文关键词:apple orchard small target detection yield forecasting feature extraction YOLO-LT UVA aerial imagery |
| 基金项目:国家自然科学基金（ 32460440）；甘肃省高校教师创新基金（2023A-051） |
|
| 摘要点击次数: 223 |
| 全文下载次数: 140 |
| 中文摘要: |
| 针对无人机拍摄条件下小目标果实识别及产量预测精度不高的问题，提出一种YOLO-LT的无人机航拍苹果果园小目标检测及产量预测方法。首先，提出YOLO-LT模型，在YOLOv8n模型主干网络中添加图像增强去雾模块（feature fusion attention network, FFA-Net），增强了雾天等复杂环境下图像的清晰度；用焦点调制网络（focal modulation networks, Focal Nets）替代快速空间金字塔池化模块（spatial pyramid pooling fast, SPPF），提高了小目标特征的提取能力；在neck部分引入扩张式残差分割网络（dilation-wise residual segmentation , DWR seg），优化C2f和bottleneck模块的性能；同时，在目标检测层中集成混合注意力变换器（hybrid attention transformer, HAT），进一步提升模型对小目标的关注度和检测敏感性。其次，利用YOLO-LT模型检测果树图像中的果实数量，并结合无人机图像获取的果树树冠面积，将上述检测结果作为输入特征，构建CNN-LSTM空间时序产量预测模型。试验表明，提出的YOLO-LT模型不仅在检测精度上有显著提升，其精确率、召回率和平均精确率均值（mAP）分别提高了5.3、3.9和4.3个百分点，推理速度达到49帧·s-1，且模型大小仅为12.6 MB。CNN-LSTM产量预测模型的决定系数（ R2 ）达到了0.8060，均方根误差（RMSE）为1.8167 kg。此方法能够满足自然环境下果树测产的实际需求，为现代果园智能化管理提供了有效的技术支持。 |
| 英文摘要: |
| In order to solve the problem of insufficient detection accuracy of small target fruit under unmanned aerial vehicle (UAV) shooting conditions and yield prediction of fruit trees in natural environments, a YOLO-LT UAV aerial photography method for detecting small targets and predicting yield in apple orchards was proposed. Firstly, the YOLO-LT model was proposed, and the Feature Fusion Attention Network (FFA-Net) was added to the backbone network of the YOLOv8n model to enhance the clarity of the image in complex environments such as foggy days. Focal Modulation Networks (Focal Nets) were used to replace the Spatial Pyramid Pooling Fast (SPPF) to improve the extraction ability of small target features. In the neck part, a dilation\|wise residual segmentation (DWR Seg) network was introduced to optimize the performance of the C2f and bottleneck modules. At the same time, a Hybrid Attention Transformer (HAT) was integrated into the object detection layer to further improve the attention and detection sensitivity of the model to small targets. Secondly, the YOLO-LT model was used to detect the number of fruits in the fruit tree image, combined with the canopy area of the fruit tree obtained by the UAV image, and the above detection results were used as input features to construct the CNN-LSTM spatial time series yield prediction model. Experiments showed that the proposed YOLO-LT model not only had a significant improvement in detection accuracy but also increased its precision, recall, and mean average precision (mAP) by 5.3, 3.9, and 4.3 percentage points, respectively; the inference speed reaches 49 frames per second, and the model size was only 12.6 MB. The coefficient of determination (R2) of the CNN-LSTM yield prediction model reached 0.8060, and the root mean square error (RMSE) was 1.8167 kg. This method can meet the actual needs of fruit tree yield measurement in the natural environment and provides effective technical support for the intelligent management of modern orchards. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| | |