Abstract Terricolous macrolichens are an important food source for caribou (Rangifer tarandus) and can greatly influence their movement, distribution and demography over time. Mapping the spatial distribution and cover of macrolichens with remote sensing can serve as an important approach for assessing the impact of disturbances (e.g. fire, grazing, trampling) on lichen cover at the landscape scale and for monitoring post-disturbance rates of recovery. Previous remote sensing-based efforts to retrieve the distribution and abundance of lichen have been restricted to particular regions and thus are not indicative of the potential for large extent mapping and monitoring. In this study, we assessed the effectiveness of machine learning methods for retrieving lichen cover and change across different regions in northern Canada and Alaska using Landsat-5 images, topographic and climate data. Global and regional-scale models were evaluated to assess whether regionally specific analyses would improve performance. Of the models tested, the deep neural network was the most accurate for predicting lichen cover (model efficiency (ME) = 0.58, mean absolute error (MAE)

中文翻译：

---

加拿大北部和阿拉斯加驯鹿范围内基于 Landsat 的大地衣覆盖检索和变化分析的评估

摘要 Terricolous macrolichens 是北美驯鹿 (Rangifer tarandus) 的重要食物来源，随着时间的推移可以极大地影响它们的运动、分布和人口统计。用遥感绘制大型地衣的空间分布和覆盖可以作为评估干扰（例如火灾、放牧、践踏）对景观尺度上地衣覆盖的影响和监测干扰后恢复率的重要方法。以前基于遥感的努力来检索地衣的分布和丰度仅限于特定区域，因此并不能表明大范围绘图和监测的潜力。在这项研究中，我们使用 Landsat-5 图像、地形和气候数据评估了机器学习方法在检索加拿大北部和阿拉斯加不同地区的地衣覆盖和变化方面的有效性。对全球和区域尺度模型进行了评估，以评估区域特定分析是否会提高性能。在测试的模型中，深度神经网络在预测地衣覆盖率方面最准确（模型效率 (ME) = 0.58，平均绝对误差 (MAE)）