甘蔗干旱等级识别的低空遥感分割模型研究OA
Research on a Low-Altitude Remote Sensing Segmentation Model for Sugarcane Drought Level Identification
针对单一气象指标为主要依据的干旱评估与预警方法难以精细化反映甘蔗生长状况的问题,研究构建了一种融合低空遥感影像与深度学习模型的甘蔗干旱等级智能识别方法.以健康/干旱甘蔗为研究对象,在三个典型种植区域采集地块低空遥感影像,计算归一化植被指数,并结合专家判读建立干旱等级标签.在 U-Net 语义分割框架基础上,引入残差网络 ResNet-34 作为主干,并采用加权跳跃连接提升边缘识别能力,构建了 CaneUNet-W 模型.通过区域交叉验证评估模型性能,结果显示模型在甘蔗地块识别任务中的交并比(IoU)、F1 值和 Kappa 系数平均分别为 0.924、0.960和 0.912,表现出较高的精度与泛化能力,模型推演结果的空间分布特征与专家基于高分影像的视觉判读趋势基本吻合.本方法能客观反映甘蔗冠层在干旱胁迫下的遥感表型变异,验证了其在甘蔗旱情智能识别与精细化评估中的可行性与有效性.但受限于地面同步实测数据的缺乏,未来研究将引入土壤墒情等生理指标进行定量标定,从而提升甘蔗生产的防灾减灾能力.
To address the limitation of drought assessment and early warning methods relying primarily on single meteorological indicators in accurately reflecting sugarcane growth,this study constructs an intelligent method for identifying sugarcane drought levels by integrating low-altitude remote sensing imagery and a deep learning model.Low-altitude remote sensing imagery of healthy and drought-stressed sugarcane fields was collected in three typical planting areas.The Normalized Difference Vegetation Index(NDVI)was calculated and drought level labels were established based on expert interpretation.Based on the U-Net semantic segmentation framework,a residual network(ResNet-34)was introduced as the backbone,and weighted skip connections were used to enhance edge recognition capabilities,resulting in the CaneUNet-W model.Model performance was evaluated through regional cross-validation.Results showed that the model achieved average Intersection over Union(IoU),F1-score,and Kappa coefficient of 0.924,0.960,and 0.912,respectively,for sugarcane fields identification,demonstrating high accuracy and generalization.The spatial distribution characteristics of the model's inference results were basically consistent with the visual interpretation trends of experts based on high-resolution imagery.The method presented in this paper can objectively reflect the remote sensing phenotypic variation of sugarcane canopy under drought stress,verifying its feasibility and effectiveness in intelligent identification and refined assessment of sugarcane drought conditions.However,due to the lack of synchronous ground-based measured data,future research will introduce physiological indicators such as soil moisture for quantitative calibration,thereby improving the disaster prevention and mitigation capabilities of sugarcane production.
王政锋;李莉;马瑞升;莫妘捷;匡昭敏
桂林理工大学计算机科学与工程学院(人工智能学院)/广西嵌入式技术与智能系统重点实验室,广西 桂林 541006广西壮族自治区气候中心,南宁 530022广西壮族自治区气象科学研究所,南宁 530022广西壮族自治区气象技术装备中心,南宁 530022广西壮族自治区气象科学研究所,南宁 530022
甘蔗低空遥感深度学习语义分割U-Net
sugarcanelow-altitude remote sensingdeep learningsemantic segmentationU-Net
《农业大数据学报》 2026 (3)
281-295,15
中国气象局科技项目"揭榜挂帅"(No.CMAJBGS202322)广西自然科学基金项目(No.2026GXNSFAA00641192)桂林理工大学科研启动基金(No.GUTQDJJ6613034)广西嵌入式技术与智能系统重点实验室基金项目(No.2020-2-3).
评论