A fusion approach to forest disturbance mapping using time series ensemble techniques
A fusion approach to forest disturbance mapping using time series ensemble techniques
复制标题
DOI：
10.1016/j.rse.2018.11.025
复制
发表时间：
2019-02
影响因子：
13.5
通讯作者：
S. Hislop;S. Jones;M. Soto‐Berelov;A. Skidmore;A. Haywood;T. Nguyen
中科院分区：
文献类型：
--
作者：
S. Hislop;S. Jones;M. Soto‐Berelov;A. Skidmore;A. Haywood;T. Nguyen
Landsat数据的时间序列分析被广泛用于评估大面积尺度的森林变化。已经提出了各种变化检测算法，每种算法都采用不同的技术来表征突发干扰事件和长期趋势。然而，根据所选择的算法、参数和光谱指数，结果可能会有很大差异。这种结果的不匹配导致研究人员假设基于集合的方法可能会提高准确性。在这项研究中，我们评估了两种变化检测算法（LandTrendr和R package strucchange），每种算法都有三个指数（归一化植被指数或NDVI，归一化燃烧比或NBR和流穗帽湿度或TCW）。我们测试了它们在29 年的时间内检测硬叶林突变干扰的能力，并随后使用简单的融合规则和随机森林模型评估了一些集合。总共4087个人工解译的参考像元，从900 百万 公顷的森林中采样，用于训练和验证。此外，我们评估了随机森林分类器与混淆情况（时间序列算法的委托误差）的启动效果。我们的结果清楚地表明，组合多种变化检测技术的集成优于任何一种方法。我们最准确的随机森林模型，使用所有6种算法输出的集合，以及3个双时间变化光栅（NBR， NDVI和TCW的变化），与最准确的单一算法/指数方法（LandTrendr with NBR）相比，其总体错误率为7%，其总体错误率为21%。我们的研究结果还表明，通过使用鲁棒参考数据和随机森林分类，可以在不使用传统变化检测算法的情况下获得可接受的结果（误差14%）。然而，通过用变化检测算法通知的混淆情况启动分类器，委托错误大大减少，代价是遗漏错误略有增加。事实上，使用初始训练数据和仅3个双时态变化光栅的随机森林集合比任何单独的算法都更准确，总体误差为11%。通过包括一些来自Landsat时间序列的附加指标（例如2 年变化和总体平均值），该误差进一步降低到8%。考虑到大多数变化检测算法都有很大的处理要求，这表明算法只能应用于像素样本，唯一的目的是训练机器学习分类器。我们通过在大面积森林（900 百万 公顷）和长时间（29 年）上创建年度扰动图，证明了这种以前未探索过的方法的可行性。
Time series analysis of Landsat data is widely used for assessing forest change at the large-area scale. Various change detection algorithms have been proposed, each employing different techniques to characterise abrupt disturbance events and longer term trends. However, results can vary significantly, depending on the algorithm, parameters and the spectral index (or indices) chosen. This mismatch in results has led to researchers hypothesizing that an ensemble based approach may increase accuracy. In this study we assess two change detection algorithms (LandTrendr and the R package strucchange), each with three indices (the Normalized Difference Vegetation Index or NDVI, the Normalized Burn Ratio or NBR, and Tasseled Cap Wetness or TCW). We test their ability to detect abrupt disturbances in sclerophyll forests over a 29 year time period, and subsequently evaluate a number of ensembles, using simple fusion rules and Random Forests models. A total of 4087 manually interpreted reference pixels, sampled from 9 million ha of forest, were used for training and validation. In addition, we assess the effects of priming the Random Forests classifier with confusing cases (commission errors from the time series algorithms). Our results clearly show that ensembles combining multiple change detection techniques out-perform any one method. Our most accurate Random Forests model, using an ensemble of all 6 algorithm outputs, along with 3 bi-temporal change rasters (change in NBR, NDVI and TCW), had an overall error rate of 7%, compared with the most accurate single algorithm/index approach (LandTrendr with NBR), which had an overall error of 21%. Our findings also indicate that acceptable results (14% error) can be achieved without the use of traditional change detection algorithms, by using robust reference data and Random Forests classification. However, by priming the classifier with confusing cases informed by the change detection algorithms, commission errors decreased substantially, at the expense of slight increases in omission errors. In fact, a Random Forests ensemble, using the primed training data and only 3 bi-temporal change rasters, was more accurate than any one individual algorithm, with an overall error of 11%. By including some additional metrics derived from Landsat time series (e.g. 2 year changes and overall means) this error was further reduced to 8%. Given that most change detection algorithms have large processing requirements, this suggests that algorithms can be applied to a sample of pixels only, for the sole purpose of training a machine learning classifier. We demonstrate the feasibility of this previously unexplored approach, by creating annual disturbance maps over a large area of forest (9 million ha) and long time period (29 years).