Simulated Hyperspectral Dataset of Forest Canopies for Unmixing Problems

Dataset specification

This is a simulated hyperspectral dataset of five forest and orchard scenes as imaged on top of the canopy. The simulation was done using HyperBlend spectral canopy simulator. HyperBlend is hosted at github.com/silmae/HyperBlend/. The frozen version of the source code is provided here in file HyperBlend.zip. The included software version was used to generate the dataset and future versions of the code will likely not work with the files provided in here.

There are five geometrical variations of the scenes with two soil spectra in five spatial resolution. The scenes are F1, F2, I1, O1, and O2 (provided in zip files named accordingly). F1 and F2 are general fairly flat forest scenes, while the I1 is a forest scene on a 15* inclined ground. All of them have a heterogenous distribution of trees. O1 and O2 are orchards with homogenous lines of trees. If you want to use HyperBlend to manipulate the scenes, simply unpack the scene zips to /HyperBlend/System simulation/. 

The two soil types are wet peat and dry sand indicated by "WP" and "DS", respectively, in the file and directory names. The soil spectra are simulated with GSV (https://doi.org/10.1016/j.jag.2019.101932).

The five spatial resolutions are 1024, 256, 64, 16, and 4 pixel squares. Rendering samples (how many light rays are cast through each pixel) is 32 for the highest resolution, 128 for the next, and so on, so each spectral band is rendered with 33.5 million samples.

The camera is at 100 m altitude at the center of the world. It has a 28* field of view, resulting in roughly 50*50 m image area. The scene itself is 75*75 m, so that even though no all of the trees are in the imaged area, they can still affect the result.

The sun and sky spectra are simulated with SSloar GOA (https://doi.org/10.5194/gmd-15-1689-2022) at coordinates lat=45.1845 lon=5.7151 (Grenoble, France) at 30.6.2025 13:00 local time. The sun angle at that time is elevation=66.45*, azimuth=155.92*.

Ground truth abundance maps are available for all image cubes

Ground truth enmember spectra not yet available. These will be added at the latest in early 2026. In the meantime, you can use the visibility maps (not abundance maps) to calculate approximate endmembers manually.

Data structure

As an example of how the data is structured, we use the scene FDS1_1024 (unzip F1.zip and you find FDS1_1024 inside). In this name, the DS refers to dry sand soil and F1 to the geometry. (Yes, it would have made more sense to name it F1DS.) The number after refers to spatiel resolution, so in this case _1024 means that the spectral cube is 1024 pixels wide and high. The scene I1 with wet peat soil and 256 px spatial resolution is therefore called IWP1_256. The directories with just the scene and soil shortnames without the resolution definition (like just FDS1) has been used as a common ancestor to generate the other scenes and does not have a simulated spectral cube; you can safely ignore these directories.

Inside the scene directory, there are files needed by HyperBlend to simulate the spectral cubes. We do not go through all of those but focus on the ones that are usable for unmixing. The leaf data from LOTUS dataset (10.1016/j.rse.2024.114424) is named by 5 character codes like CCAN1 and ELMDK. These names come directly from the LOTUS dataset. For example, CCAN1.JPG is the image of a leaf sample, CCAN1.toml contains biochemical and biophysical measurements of the sample (chlorophyll content, mass per area, and so on), CCAN1_0_slab_sim_result.png is a visualization of HyperBlend leaf simulation result for that sample, and corresponding .toml file contains the numerical values. The leaf spectra are called Slab materials internally in HyperBlend. To make it easier for humans to read, the file leaf_material_map.toml contains mapping from slab material index to LOTUS leaf codes.

Spectra of the sun, sky, and soil are visualized in files grenoble_sun.png, grenoble_sky.png, and soil_reflectance_dry_sand_reflectance.png, respectively.

The generated spectral cube can be found from subdirectory Spectral cube/ in ENVI format (i.e., header data is in .hdr file and the pixel values in .img file).

Abundance maps can be found found from subdirectory Abundance maps/. The .pngs visualize the abundance of each material in the scene and abundances.npy is a single numpy array containing all the abundances in shape (width, height, material index).  The file map_name_indices.toml relates the material index in the abundance maps array into the LOTUS leaf indices.

Visual representation of the directory structure in ASCII art:

└── scene_FDS1_1024/
├── CCAN1.JPG
├── CCAN1.toml
├── CCAN1_0_slab_sim_result.png
├── CCAN1_0_slab_sim_result.toml
├── ELMDK.JPG
├── ELMDK.toml
├── ELMDK_0_slab_sim_result.png
├── ELMDK_0_slab_sim_result.toml
├── ...
├── grenoble_sky.png
├── grenoble_sun.png
├── leaf_material_map.toml
├── leaf_spectrum_plotSlab material 1.png
├── leaf_spectrum_plotSlab material 2.png
├── ...
├── soil_reflectance_dry_sand_reflectance.png
├── Abundance maps/
│ ├── Abundance CCAN1.png
│ ├── Abundance ELMDK.png
│ ├──...
│ ├── Abundance Soil.png
│ ├── Abundance Trunk.png
│ ├── abundances.npy
│ └── map_name_indices.toml
├── rend/
│ ├── drone_preview.png
│ ├── sleeper_preview.png
│ ├── trees_preview.png
│ ├── walker_preview.png
│ └── Visibility maps/
└── Spectral cube/
├── spectral_cube_FDS1_1024.hdr
└── spectral_cube_FDS1_1024.img

数据集说明

本数据集为针对5种森林与果园冠层成像场景构建的模拟高光谱数据集（hyperspectral dataset），模拟过程使用HyperBlend（HyperBlend）光谱冠层模拟器完成。HyperBlend的托管地址为github.com/silmae/HyperBlend/，本次附带的源代码冻结版本已打包在HyperBlend.zip文件中。本次数据集生成即使用该附带软件版本，后续代码版本大概率无法兼容本数据集附带的文件。

本数据集包含5种场景几何变体，搭配2种土壤光谱，并提供5种空间分辨率（spatial resolution）。场景分别为F1、F2、I1、O1与O2（对应命名的zip压缩包提供）。其中F1与F2为整体相对平缓的森林场景，I1为位于15°倾斜地面上的森林场景，所有场景均具备非均匀分布的林木。O1与O2为具备均匀林木行列布局的果园场景。若需使用HyperBlend对场景进行编辑，只需将场景压缩包解压至/HyperBlend/System simulation/目录即可。

两种土壤类型分别为湿泥炭与干沙，在文件与目录命名中分别以"WP"与"DS"标识。土壤光谱通过GSV（https://doi.org/10.1016/j.jag.2019.101932）模拟生成。

5种空间分辨率分别为1024、256、64、16与4像素见方。渲染采样数（即每像素投射的光线数量）在最高分辨率下为32，次一级分辨率为128，依此类推，每个光谱波段均搭载3350万次采样。

相机位于世界中心上空100米处，视场角（field of view）为28°，对应成像区域约为50×50米。场景本身尺寸为75×75米，因此即便部分林木未处于成像区域内，仍可对成像结果产生影响。

太阳与天空光谱通过SSloar GOA（https://doi.org/10.5194/gmd-15-1689-2022）模拟，模拟坐标为纬度45.1845、经度5.7151（法国格勒诺布尔），时间为2025年6月30日13:00当地时间。此时太阳高度角为66.45°，方位角为155.92°。

所有光谱数据立方体均配有地面实况丰度图，但端元光谱暂未公开，预计最晚将于2026年初发布。在此期间，用户可借助可见性图（而非丰度图）手动计算近似端元。

数据结构

以场景FDS1_1024为例（解压F1.zip后即可在内部找到FDS1_1024目录）。该命名中，DS代表干沙土壤，F1代表场景几何类型（注：原命名逻辑本应设为F1DS更为合理）。后缀数字代表空间分辨率，因此_1024表示该光谱立方体的宽与高均为1024像素。例如，搭载湿泥炭土壤、空间分辨率为256像素的I1场景，其命名应为IWP1_256。仅包含场景与土壤短名称、未标注分辨率的目录（如仅FDS1）为生成其余场景的公共父目录，未包含模拟光谱立方体，可安全忽略。

在场景目录中，包含HyperBlend模拟光谱立方体所需的全部文件。下文不再逐一赘述，仅聚焦于可用于光谱解混的文件。来自LOTUS数据集（10.1016/j.rse.2024.114424）的叶片数据以5位字符代码命名，例如CCAN1与ELMDK，此类命名直接沿用自LOTUS数据集。以CCAN1为例，CCAN1.JPG为叶片样本图像，CCAN1.toml包含该样本的生化与生物物理参数（如叶绿素含量、单位面积质量等），CCAN1_0_slab_sim_result.png为HyperBlend针对该样本的叶片模拟结果可视化图，对应的.toml文件则存储相关数值。在HyperBlend内部，叶片光谱以Slab材料命名。为便于人工识别，文件leaf_material_map.toml存储了Slab材料索引与LOTUS叶片代码的映射关系。

太阳、天空与土壤的光谱分别存储于grenoble_sun.png、grenoble_sky.png与soil_reflectance_dry_sand_reflectance.png文件中，用于可视化展示。

生成的光谱立方体存储于Spectral cube/子目录，采用ENVI格式（ENVI format），即头文件为.hdr，像素值存储于.img文件。

丰度图存储于Abundance maps/子目录：.png文件可视化了场景中每种材料的丰度分布，abundances.npy为单个NumPy数组，以(宽, 高, 材料索引)的形状存储所有丰度数据。文件map_name_indices.toml建立了丰度图数组中的材料索引与LOTUS叶片索引的对应关系。

以下为目录结构的ASCII艺术图：
└── scene_FDS1_1024/
├── CCAN1.JPG
├── CCAN1.toml
├── CCAN1_0_slab_sim_result.png
├── CCAN1_0_slab_sim_result.toml
├── ELMDK.JPG
├── ELMDK.toml
├── ELMDK_0_slab_sim_result.png
├── ELMDK_0_slab_sim_result.toml
├── ...
├── grenoble_sky.png
├── grenoble_sun.png
├── leaf_material_map.toml
├── leaf_spectrum_plotSlab material 1.png
├── leaf_spectrum_plotSlab material 2.png
├── ...
├── soil_reflectance_dry_sand_reflectance.png
├── Abundance maps/
│ ├── Abundance CCAN1.png
│ ├── Abundance ELMDK.png
│ ├──...
│ ├── Abundance Soil.png
│ ├── Abundance Trunk.png
│ ├── abundances.npy
│ └── map_name_indices.toml
├── rend/
│ ├── drone_preview.png
│ ├── sleeper_preview.png
│ ├── trees_preview.png
│ ├── walker_preview.png
│ └── Visibility maps/
└── Spectral cube/
├── spectral_cube_FDS1_1024.hdr
└── spectral_cube_FDS1_1024.img