Sampling procedure of land use and soil fertility map under uncertainty

The dataset is composed of input data, the code to process the data and the output data from a sampling procedure for spatial data. The sampling procedure is used to derive samples, that result in extreme lower and upper bound single-objective values. The extreme lower and upper bound samples are used to approximate the extreme Pareto fronts when performing the multi-objective optimization with the samples as input data. To the input data belong a land use map, a patch ID map, a soil fertility map and an uncertainty map, which all have 10*10 cells. In addition, two matrices are included. One matrix defines transition constraints for the land uses of the land use map, the second matrix defines related classification accuracies per land use. The land use map, patch ID map, transition matrix and soil fertility map were reused from the CoMOLA project, produced for research from Strauch et al. 2019 (https://doi.org/10.1016/j.envsoft.2019.05.003). The outputs are divided into three categories. One category is about creating samples from the uncertain land use map with the original soil fertility map. The second category is about creating samples from the uncertain soil fertility map with the original soil fertility map. The third category is about creating samples from both uncertain maps combined.
For the first category, the land use maps and the corresponding patch ID maps were stored that resulted in extremely low or high objective values of habitat heterogeneity, forest species richness, crop yield and water yield from 1000 samples (computation time: <1 minute). For category 2, the soil fertility maps are stored that resulted in extremely low and high crop yield objective values with the original land use map were stored from 1000 samples, because the soil fertility map only influences the objective crop yield (computation time: <10 seconds). For category 3, two different soil fertility maps and two different land use map with the corresponding patch ID maps that lead to the extreme lower and upper bound crop yield values are stored. Here, the 1000 samples of category 1 and 2 are reproduced and each sampled land use map is evaluated against every sampled soil fertility map (computation time: ~2 hours). For convenience, additional pickle files were saved containing extreme solutions, extreme objective values and extreme objective values per iteration of the sampling procedure to plot the progression. The execution of the Python script SpatialSampling.py (Python 3.8) needs to be executed to reproduce the sampling procedure. After the simulations, plots are generated in case that the boolean parameter "show_visualizations" (beginning of script) is set to True. Pseudo-random states are used to assure reproducibility despite the stochastic processes.

本数据集由空间数据采样流程对应的输入数据、数据处理代码，以及采样得到的输出数据三部分组成。该采样流程用于生成可获取极端下界与上界单目标值的样本集；当以该类样本作为输入数据开展多目标优化时，极端下界与上界样本可用于近似极端帕累托前沿（Pareto front）。
输入数据包含土地利用图（land use map）、斑块ID图（patch ID map）、土壤肥力图（soil fertility map）与不确定性图（uncertainty map），所有图件均为10×10的栅格单元。此外，数据集还包含两类矩阵：其一为土地利用类型转换约束矩阵（transition constraint matrix），其二为各土地利用类型对应的分类精度矩阵（classification accuracy matrix）。本数据集所复用的土地利用图、斑块ID图、转换矩阵与土壤肥力图均来自CoMOLA项目，该项目由Strauch等人于2019年开展相关研究（DOI: 10.1016/j.envsoft.2019.05.003）。
输出数据分为三类：第一类基于原始土壤肥力图，从不确定土地利用图中生成采样样本；第二类基于原始土地利用图，从不确定土壤肥力图中生成采样样本；第三类则同时结合两类不确定图件生成采样样本。
第一类场景中，我们从1000个采样样本中筛选得到了生境异质性、森林物种丰富度、作物产量与产水量的极端下界或上界目标值对应的土地利用图及其配套斑块ID图，并将其存储（单次采样计算时长<1分钟）。第二类场景中，我们基于原始土地利用图，从1000个采样样本中筛选得到了作物产量目标值处于极端下界与上界的土壤肥力图并予以存储；由于土壤肥力图仅对作物产量目标值产生影响，该类采样的计算时长<10秒。第三类场景中，我们存储了两组可产生极端下界与上界作物产量值的土壤肥力图，以及两组配套的土地利用图与斑块ID图。该场景下，我们复现了第一类与第二类的1000个采样样本，并将每一张采样得到的土地利用图与每一张采样得到的土壤肥力图进行配对评估，整体计算时长约为2小时。
为便于使用，本数据集还额外存储了pickle格式文件（pickle），其中包含采样流程各迭代轮次的极端解、极端目标值与迭代过程目标值，用于绘制采样进展曲线。若需复现该采样流程，需运行Python脚本SpatialSampling.py（适配Python 3.8版本）。模拟完成后，若将脚本起始处的布尔型参数"show_visualizations"设置为True，则可生成可视化图表。尽管采样流程包含随机过程，但通过使用伪随机状态（pseudo-random states）可确保实验结果可复现。