Wan Luwen, Kendall Anthony D, Rapp Jeremy, Hyndman David W
Department of Earth and Environmental Sciences, Michigan State University, East Lansing, MI 48824, USA; Department of Earth System Science, Stanford University, Stanford, CA 94305, USA; Institute for Human-Centered Artificial Intelligence, Stanford University, Stanford, CA 94305, USA.
Department of Earth and Environmental Sciences, Michigan State University, East Lansing, MI 48824, USA.
Sci Total Environ. 2024 Nov 10;950:175283. doi: 10.1016/j.scitotenv.2024.175283. Epub 2024 Aug 5.
There has been an increase in tile drained area across the US Midwest and other regions worldwide due to agricultural expansion, intensification, and climate variability. Despite this growth, spatially explicit tile drainage maps remain scarce, which limits the accuracy of hydrologic modeling and implementation of nutrient reduction strategies. Here, we developed a machine-learning model to provide a Spatially Explicit Estimate of Tile Drainage (SEETileDrain) across the US Midwest in 2017 at a 30-m resolution. This model used 31 satellite-derived and environmental features after removing less important and highly correlated features. It was trained with 60,938 tile and non-tile ground truth points within the Google Earth Engine cloud-computing platform. We also used multiple feature importance metrics and Accumulated Local Effects to interpret the machine learning model. The results show that our model achieved good accuracy, with 96 % of points classified correctly and an F1 score of 0.90. When tile drainage area is aggregated to the county scale, it agreed well (r = 0.69) with the reported area from the Ag Census. We found that Land Surface Temperature (LST) along with climate- and soil-related features were the most important factors for classification. The top-ranked feature is the median summer nighttime LST, followed by median summer soil moisture percent. This study demonstrates the potential of applying satellite remote sensing to map spatially explicit agricultural tile drainage across large regions. The results should be useful for land use change monitoring and hydrologic and nutrient models, including those designed to achieve cost-effective agricultural water and nutrient management strategies. The algorithms developed here should also be applicable for other remote sensing mapping applications.
由于农业扩张、集约化和气候变化，美国中西部以及全球其他地区的瓦片排水面积有所增加。尽管有这种增长，但空间明确的瓦片排水地图仍然稀缺，这限制了水文建模的准确性和养分减少策略的实施。在此，我们开发了一种机器学习模型，以30米分辨率提供2017年美国中西部瓦片排水的空间明确估计（SEETileDrain）。该模型在去除不太重要和高度相关的特征后，使用了31个卫星衍生和环境特征。它在谷歌地球引擎云计算平台内使用60938个瓦片和非瓦片地面真值点进行训练。我们还使用了多种特征重要性指标和累积局部效应来解释机器学习模型。结果表明，我们的模型具有良好的准确性，96%的点分类正确，F1分数为0.90。当瓦片排水面积汇总到县尺度时，它与农业普查报告的面积吻合良好（r = 0.69）。我们发现地表温度（LST）以及与气候和土壤相关的特征是分类的最重要因素。排名最高的特征是夏季夜间LST中位数，其次是夏季土壤湿度百分比中位数。这项研究证明了应用卫星遥感绘制大区域空间明确的农业瓦片排水图的潜力。结果应有助于土地利用变化监测以及水文和养分模型，包括那些旨在实现具有成本效益的农业水和养分管理策略的模型。这里开发的算法也应适用于其他遥感测绘应用。
