Reindeer lichens (Cladonia spp.) are an essential food source for caribou especially during winter. They can also be a valuable indicator for ecosystem health and climate change. Inventory of lichen abundance at regional scales is required to assess availability within caribou ranges, and assess potential declines from natural and anthropogenic disturbances. Previous studies have mapped lichen cover and volume using remote sensing, but these efforts were often constrained by the limited availability of ground truth information needed for model calibration and validation. In this study, we leveraged unoccupied aerial vehicle (UAV) surveys and WorldView (WV) satellite scenes in a nested upscaling approach in order to expand the number of training samples at the 30 m Landsat resolution. These were used to develop machine learning models to map fractional reindeer lichen cover in Eastern Canada. We found that the best correlation between UAV and WV derived lichen coverages exists at an optimal scale that is slightly larger than 30 m and varies with landscape type and observation geometry. Based on training data from UAV-calibrated lichen coverage from WV data, a neural network model with simple structure achieved a root mean square error (RMSE) = 0.09, a mean absolute error (MAE) = 0.07 and R² = 0.79 for mapping fractional lichen cover from Landsat without the use of ancillary data. We then applied our model and Landsat data to produce a lichen fractional cover map for the Red Wine Mountain caribou herd range in Labrador, NL and the Manicouagan caribou herd range in Québec. Validation against domain-averaged lichen cover in eight UAV survey sites suggests an accuracy with RMSE = 0.04, MAE = 0.03 and R² = 0.62 for low lichen cover. Compared to aggregated lichen cover at 30 m from UAV surveys, map accuracy decreases to RMSE = 0.09, MAE = 0.06, and R² = 0.49, partially due to registration error between UAV and Landsat images. Our study demonstrates that upscaling of lichen cover from UAV data to Landsat via an intermediate image scale is an effective regional-scale mapping approach.

驯鹿地衣（Cladoniaspp.) 是驯鹿的重要食物来源，尤其是在冬季。它们也可以成为生态系统健康和气候变化的重要指标。需要在区域尺度上清查地衣丰度，以评估驯鹿范围内的可用性，并评估自然和人为干扰造成的潜在下降。以前的研究已经使用遥感绘制了地衣覆盖和体积的地图，但这些努力通常受到模型校准和验证所需的地面实况信息的有限可用性的限制。在本研究中，我们以嵌套升级方法利用无人飞行器 (UAV) 调查和 WorldView (WV) 卫星场景，以扩大 30 m Landsat 分辨率下的训练样本数量。这些被用来开发机器学习模型来绘制加拿大东部的部分驯鹿地衣覆盖物。我们发现 UAV 和 WV 派生的地衣覆盖率之间的最佳相关性存在于略大于 30 m 的最佳尺度上，并且随景观类型和观测几何形状而变化。基于来自 WV 数据的 UAV 校准地衣覆盖的训练数据，结构简单的神经网络模型实现了均方根误差 (RMSE) = 0.09、平均绝对误差 (MAE) = 0.07 和 R²  = 0.79，用于在不使用辅助数据的情况下绘制来自 Landsat 的部分地衣覆盖。然后，我们应用我们的模型和 Landsat 数据为荷兰拉布拉多的 Red Wine Mountain 驯鹿群和魁北克的 Manicouagan 驯鹿群生成地衣部分覆盖图。针对八个无人机调查站点的域平均地衣覆盖率的验证表明， 对于低地衣覆盖率，RMSE = 0.04、MAE = 0.03 和 R ² = 0.62的准确性。与无人机调查的 30 m 处聚合地衣覆盖相比，地图精度降低到 RMSE = 0.09、MAE = 0.06 和 R ² = 0.49，部分是由于 UAV 和 Landsat 图像之间的配准错误。我们的研究表明，通过中间图像尺度将地衣覆盖从 UAV 数据放大到 Landsat 是一种有效的区域尺度制图方法。