Satellite-based field delineation has entered a quasi-operational stage due to recent advances in machine learning for computer vision. Transfer learning allows for the resource-efficient transfer of pre-trained field delineation models across heterogeneous geographies. However, the scarcity of labeled data for complex and dynamic smallholder landscapes remains a major bottleneck. The key innovation of this study is to overcome this challenge by using pre-trained models to generate sparse (i.e., not fully annotated) field delineation pseudo-labels for fine-tuning models across geographies and sensor characteristics. We build on a FracTAL ResUNet trained for crop field delineation in India (median field size of 0.24 ha) based on multi-spectral imagery at 1.5 m spatial resolution. We use this model to generate pseudo-labels for the use in Northern Mozambique (median field size of 0.06 ha) based on sub-meter resolution true-color satellite imagery. We designed multiple pseudo-label selection strategies based on field-level probability scores and compared the quantities, area properties, seasonal distribution, and spatial agreement of the pseudo-labels against human-annotated training labels (n = 1,512). We then used the human-annotated labels and the pseudo-labels for model fine-tuning and compared predictions against human field annotations (n = 2,199). We evaluated performance with regards to object-level spatial agreement and site-level field size estimation. Our results indicate i) a good baseline performance of the pre-trained model in both field delineation (mean intersection over union (mIoU) of 0.634) and field size estimation (mean root mean squared error (mRMSE) of 0.071 ha), and ii) the added value of regional fine-tuning with performance improvements in nearly all experiments (mIoU increases of up to 0.060, mRMSE decreases of up to 0.034 ha). Moreover, we found iii) substantial performance increases when using only pseudo-labels (up to 77 % of the mIoU increases and 68 % of the mRMSE decreases obtained by human-annotated labels), and iv) additional performance increases (mIoU+0.008, mRMSE: −0.003 ha) when complementing human annotations with pseudo-labels. Pseudo-labels are architecture-agnostic, can be efficiently generated at scale, and thus facilitate domain adaptation in label-scarce settings. The workflow presented here is a stepping stone for overcoming the persisting challenges in mapping heterogeneous smallholder agriculture.

中文翻译：

---

更进一步：利用伪标签在标签稀缺的小农地区进行田地划分


由于计算机视觉机器学习的最新进展，基于卫星的场地描绘已进入准操作阶段。迁移学习允许跨异构地理区域以资源高效的方式迁移预训练的野外描绘模型。然而，复杂且动态的小农景观的标记数据的缺乏仍然是一个主要瓶颈。这项研究的关键创新是通过使用预训练模型生成稀疏（即未完全注释）的场描绘伪标签来克服这一挑战，以便跨地理位置和传感器特性微调模型。我们建立在 FracTAL ResUNet 的基础上，该网络经过训练，用于基于 1.5 m 空间分辨率的多光谱图像来描绘印度作物田地（中位田地大小为 0.24 公顷）。我们使用该模型根据亚米分辨率真彩色卫星图像生成伪标签，供莫桑比克北部使用（中位面积为 0.06 公顷）。我们根据字段级概率得分设计了多种伪标签选择策略，并将伪标签的数量、区域属性、季节分布和空间一致性与人工注释的训练标签进行了比较（n = 1,512）。然后，我们使用人工注释标签和伪标签进行模型微调，并将预测与人工现场注释进行比较 (n = 2,199)。我们评估了对象级空间一致性和站点级场大小估计方面的性能。我们的结果表明 i) 预训练模型在场描绘（平均交并集 (mIoU) 为 0.634）和场大小估计（平均均方根误差 (mRMSE) 为 0）方面均具有良好的基线性能。071 ha)，ii) 几乎所有实验中区域微调的附加值和性能改进（mIoU 增加最多 0.060，mRMSE 减少最多 0.034 ha）。此外，我们发现 iii) 仅使用伪标签时性能显着提高（通过人工注释标签获得的 mIoU 增加高达 77%，mRMSE 降低 68%），以及 iv) 额外的性能增加（mIoU+0.008， mRMSE：-0.003 ha）当用伪标签补充人类注释时。伪标签与架构无关，可以大规模有效地生成，从而促进标签稀缺环境中的域适应。这里介绍的工作流程是克服异构小农农业绘图中持续存在的挑战的基石。