手把手教程：经典STARFM时空重建模型的使用——starfm4py使用教程

starfm4py使用教程

1. 安装Anaconda3，我用的版本是Anaconda3-2024.10-1-Windows-x86_64.exe

2. 以管理员方式，打开anaconda prompt

3. 下载 starfm4py-master.zip，网址如下。不方便科学上网的，我也会在后的网盘中提供：

https://github.com/nmileva/starfm4py?tab=GPL-3.0-1-ov-file

4.解压，然后可以看到 requirements.txt 文件，以下是我的绝对路径（D:\Parper01\starfm4py-master\requirements.txt），然后在anaconda prompt中运行：

# requirements.txt 包含了需要的所有包
conda create --name starfm_v2 --file D:\Parper01\starfm4py-master\requirements.txt python=3.11.0

创建虚拟环境完成，然后激活

# 激活环境，可以安装自己需要的包。但是上述步骤后， starfm4py需要的环境就已经好了
conda activate starfm_v2
# 如果你使用 Vs Code，可能需要安装 ipykernel
conda install -n starfm_v2 ipykernel --update-deps --force-reinstall
# 退出环境
conda deactivate

5. 把D:\Parper01\starfm4py-master\src中的parameters.py和starfm4py.py复制到D:\ProgramData\anaconda3\envs\starfm_v2\Lib目录中。否则from parameters import (path, sizeSlices)

import starfm4py as stp 两行代码会出问题。我就在这里死磕了好长时间。

# 以下代码运行通后，说明环境已经没问题了
import time
import rasterio
import numpy as np
import matplotlib.pyplot as plt
from parameters import (path, sizeSlices)
import starfm4py as stp
# import starfm4py，starfm4py.py 是作者Nikolina Mileva 完成的，不需要 pip install 的方式安装。

6. 以下是运行的代码：

start = time.time()

#Set the path where the images are stored
product = rasterio.open(r'D:\Parper01\starfm4py-master\Tests\Test_3/sim_Landsat_t1.tif')
profile = product.profile
LandsatT0 = rasterio.open(r'D:\Parper01\starfm4py-master\Tests\Test_3/sim_Landsat_t1.tif').read(1)
MODISt0 = rasterio.open(r'D:\Parper01\starfm4py-master\Tests\Test_3/sim_MODIS_t1.tif').read(1)
MODISt1 = rasterio.open(r'D:\Parper01\starfm4py-master\Tests\Test_3/sim_MODIS_t4.tif').read(1)

# Set the path where to store the temporary results
path_fineRes_t0 = r'./Temporary/Tiles_fineRes_t0/'
path_coarseRes_t0 = r'./Temporary/Tiles_coarseRes_t0/'
path_coarseRes_t1 = r'./Temporary/Tiles_fcoarseRes_t1/'

# Flatten and store the moving window patches
fine_image_t0_par = stp.partition(LandsatT0, path_fineRes_t0)
coarse_image_t0_par = stp.partition(MODISt0, path_coarseRes_t0)
coarse_image_t1_par = stp.partition(MODISt1, path_coarseRes_t1)

print ("Done partitioning!")

# Stack the the moving window patches as dask arrays
S2_t0 = stp.da_stack(path_fineRes_t0, LandsatT0.shape)
S3_t0 = stp.da_stack(path_coarseRes_t0, MODISt0.shape)
S3_t1 = stp.da_stack(path_coarseRes_t1, MODISt1.shape)

shape = (sizeSlices, LandsatT0.shape[1])

print ("Done stacking!")

# Perform the prediction with STARFM
for i in range(0, LandsatT0.size-sizeSlices*shape[1]+1, sizeSlices*shape[1]):
    
    fine_image_t0 = S2_t0[i:i+sizeSlices*shape[1],]
    coarse_image_t0 = S3_t0[i:i+sizeSlices*shape[1],]
    coarse_image_t1 = S3_t1[i:i+sizeSlices*shape[1],]
    prediction = stp.starfm(fine_image_t0, coarse_image_t0, coarse_image_t1, profile, shape)
    
    if i == 0:
        predictions = prediction
        
    else:
        predictions = np.append(predictions, prediction, axis=0)
  

# Write the results to a .tif file   
print ('Writing product...')
profile = product.profile
profile.update(dtype='float64', count=1) # number of bands
file_name = path + 'prediction.tif'

result = rasterio.open(file_name, 'w', **profile)
result.write(predictions, 1)
result.close()


end = time.time()
print ("Done in", (end - start)/60.0, "minutes!")

# Display input and output
plt.imshow(LandsatT0)
plt.gray()
plt.show()
plt.imshow(MODISt0)
plt.gray()
plt.show()
plt.imshow(MODISt1)
plt.gray()
plt.show()	
plt.imshow(predictions)
plt.gray()
plt.show()

7. 看最终结果。

8. 总结：

我个人感觉starfm的效果还行，但是不够好，胜在经典与开创性，所以我复现一下。如果大家有需要，可以参考下我个人的论文。

文章名称：A Physically Constrained Downscaling Framework for Hourly, All-Sky Land Surface Temperature in Mountainous Regions
链接：https://ieeexplore.ieee.org/document/10884019
DOI: 10.1109/JSTARS.2025.3541374