面向玉米隐性洪涝检测的DSA-DeepLab语义分割模型

DSA - DeepLab Semantic Segmentation Model for Hidden Waterlogging Detection in Maize

摘要:

在全球气候变化背景下，极端气象事件频发，对农业生产构成严重威胁。东北地区作为中国玉米主产区，其玉米关键生育期与暴雨高发期重叠，导致洪涝灾害频发，从而给粮食安全带来严峻挑战。高分辨率遥感技术为灾情快速检测提供了有效手段，然而现有方法多基于光学与雷达影像的水体特征提取，依赖先验地物知识，在冠层未淹没的 "隐性洪涝" 区域识别方法与效果有限。针对上述问题，本研究基于 3m Planet 与 4m 国网遥感影像，构建了首套玉米隐性洪涝灾害高分辨率语义分割数据集，弥补了高质量样本不足的问题;并提出 DSA-DeepLab 语义分割模型，通过引入 DenseASPP 模块和嵌入融合选择性核注意力的自适应特征融合模块，从而有效提升小目标与复杂边界识别能力。实验结果表明，本文模型平均交并比与平均像素精度分别达到 80.70% 与 88.60%, 较基线模型提升 1.63、1.67 个百分点，隐性洪涝正样本的 IoU 与召回率分别提升 2.72、3.38 个百分点，优于多种主流语义分割方法。此外，针对大尺寸影像输入导致的内存溢出问题，推理阶段采用重叠裁剪 - 边缘忽略策略，有效抑制了拼接伪影与边界误差。本研究实现了玉米隐性洪涝灾害的高精度自动化提取，可为灾情评估与农业保险定损提供可靠技术支撑。

Abstract:

Against the backdrop of global climate change, the increasing frequency of extreme weather events poses serious threats to agricultural production. As a major maize-producing region in China, Northeast China faces the challenge that key growth stages of maize overlap with the period of frequent regional heavy rainfall, leading to recurrent waterlogging disasters that severely endanger grain security. High-resolution remote sensing technology offers an effective means for rapid disaster monitoring. However, existing methods predominantly rely on water feature extraction from optical and radar imagery, which depends on prior knowledge of surface features and shows limited effectiveness in identifying "implicit" waterlogging areas where the canopy is not submerged. This is particularly evident in the inadequate accuracy and boundary consistency for detecting small-scale affected areas. To address these issues, the first high-resolution semantic segmentation dataset for implicit maize waterlogging disasters was constructed based on 3 m Planet and 4 m State Grid remote sensing images, filling the gap in high -quality labeled data. Furthermore, a DSA - DeepLab semantic segmentation model was proposed, which incorporated a DenseASPP module and an adaptive feature fusion module embedded with selective kernel attention, thereby effectively enhancing the detection capability for small targets and complex boundaries. Experimental results demonstrated that the proposed model achieved mean intersection over union and mean pixel accuracy of 80.70% and 88.60%, respectively, representing improvements of 1.63 percentage points and 1.67 percentage points over the baseline model. Specifically, for implicit waterlogging, the IoU and recall rates were increased by 2. 72 percentage points and 3. 38 percentage points, respectively, outperforming that of several mainstream semantic segmentation methods. Additionally, to address the memory overflow issue caused by large-size image inputs during inference, an overlapping patch extraction-edge ignoring strategy was adopted, effectively suppressing stitching artifacts and boundary errors. This research achieved high-precision automated extraction of implicit maize waterlogging disasters, providing reliable technical support for disaster assessment and agricultural insurance claims.