基于无人机遥感与作物分布感知的撂荒地种植推荐策略

Cropping Recommendation Strategy for Abandoned Farmland Based on UAV Remote Sensing and Crop Distribution Perception

摘要:

近年来，受农田监测与管理手段滞后影响，部分农村地区撂荒现象频发，引发耕地资源利用率下降和粮食产能受限等问题。针对上述问题，提出一种融合作物监测与空间分布感知的撂荒地种植推荐策略。该策略以无人机航拍获取的复杂背景农田遥感影像为研究对象；在DeepLabv3+模型基础上，使用轻量级MobileNetv4作为主干特征提取网络，以压缩模型参数量并降低计算复杂度；同时，在解码阶段引入自适应细粒度通道注意力机制，以增强作物轮廓边缘与纹理细节的特征感知；采用风车卷积替换传统3×3卷积，提升模型在无人机俯视视角下农田地块的小尺寸特征提取能力；此外，构建FocalDiceLoss复合损失函数，缓解类别不平衡及相似类别间难区分的负面影响。最后，结合遥感图像解析结果、地理位置及作物空间分布统计，在撂荒地邻域范围内整合周边作物信息，利用农时与邻域优势作物，对撂荒地进行种植推荐。实验结果表明，改进后的DeepLabv3+模型的准确率、平均像素准确率和平均交并比分别为96.64%、96.37%、92.82%，相比原模型分别提高1.85、3.71、6.10个百分点。该模型可为农田作物精准监测与撂荒地资源再利用提供关键技术支撑，促进农业智能化管理与可持续发展。

Abstract:

In recent years, due to the lag in farmland monitoring and management methods, abandoned land has become increasingly prevalent in some rural areas, resulting in decreased utilization efficiency of cultivated land and constraints on grain production capacity. To address this issue, a cropping recommendation strategy for abandoned farmland was proposed, which integrated crop monitoring with spatial distribution perception. The approach utilized UAV-acquired remote sensing imagery with complex farmland backgrounds as the primary research object. On the basis of the DeepLabv3+ semantic segmentation model, a lightweight MobileNetv4 network was introduced as the backbone feature extractor to reduce parameter complexity and computational cost. Additionally, an adaptive fine-grained channel attention mechanism was incorporated in the decoder to enhance the model’s sensitivity to crop boundary contours and texture details. To improve the extraction of small-scale farmland features under UAV nadir perspectives, the conventional 3×3 convolution was replaced with windmill convolution. Furthermore, a hybrid focal-dice loss function was constructed to mitigate the effects of class imbalance and the difficulty in distinguishing between visually similar crop categories. Finally, by combining the remote sensing analysis results with geolocation data and crop spatial distribution statistics, the model aggregated surrounding crop information over a broad spatial domain and recommended suitable crops for abandoned plots based on seasonal farming schedules and regional crop dominance. Experimental results demonstrated that the improved DeepLabv3+ model achieved an ACC of 96.64%, mPA of 96.37%, and MIoU of 92.82%, representing increases of 1.85, 3.71, and 6.10 percentage points, respectively, over the baseline model. This approach can provide a critical technical foundation for precision crop monitoring and abandoned land reutilization, promoting intelligent agricultural management and sustainable farmland development.