Introduction

Why RSP

The goals of RSP are to:

• Provide a full-cycle workflow for geospatial machine learning - from data processing to training the model and making predictions with it

• Make geospatial machine learning simple and accessible while keeping it functional and customizable

Usually processing remote sensing data and training geospatial ML models in Python is complicated and need lots of code, because standard GIS libraries like GDAL or Rasterio and machine learning libraries like Scikit-learn or Pytorch provide only low-level functions. To preprocess Landsat of Sentinel image with Rasterio you need to define all the stages of preprocessing: reading data, atmospheric correction, pansharpening, cloud masking, reprojecting and writing result to a file manually with lots of code.

RSP provides high-level functions that automate routine processing operations like remote sensing data preprocessing, merging, calculating vegetation indices, training and testing models. For example, you can preprocess Sentinel-2 image from archive with operations of atmospheric correction, 20- and 60-m bands superresolution, cloud masking and reprojecting to needed projection with one line of code.

Another key idea of RSP is easy pipeline construction, where outputs from one function can be used as inputs to other functions. For example, you can preprocess several Sentinel-2 images with sentinel2 function, then merge preprocessed images with mosaic function, and then cut merged band rasters into tiles with generate_tiles function.

output_sentinels = rsp.sentinel2(sentinel2_imgs)
x = rsp.mosaic(output_sentinels, "/home/rsp_test/mosaics/sentinel/")
x_tiles, y_tiles = rsp.semantic.generate_tiles(x, y)

Also, RSP writes the outputs of the most of the functions to files, which makes possible to resume the interrupted pipeline from the last successful stage, easily return to previous stage or explore the intermediate data in traditional GIS systems. By default, RSP also saves dataset metadata in a STAC format in JSON files alongside the data itself.

FAQ

What exactly does RSP do?

General data preprocessing

With process you can clip, reproject, match with another raster, fill the gaps in a raster

With replace_value you can replace specific value in a raster.

With replace_nodata you can replace nodata value in a raster.

With rasterize you can rasterize a vector file (shapefile, geopackage, geojson etc.)

With match_hist you can match histogram of a raster to a histogram of another raster.

With clip_values you can clip the values in a raster to a specific range to remove outliers.

Data normalization

normalize module is for data normalization.

With normalize.min_max you can apply min-max normalization.

With normalize.z_score you can apply z-score normalization.

With normalize.dynamic_world you can apply log-transform + sigmoid (dynamic world) normalization.

With denormalize.min_max you can restore the original values from min-max normalized data.

With denormalize.z_score you can restore the original values from z-score normalized data.

With denormalize.dynamic_world you can restore the original values from data normalized with dynamic world normalization.

With get_normalization_params.min_max you can get min and max values of a raster.

With get_normalization_params.z_score you can get mean and standard deviation of a raster.

With get_normalization_params.percentile you can get specific percentiles of a raster.

With get_normalization_params.dynamic_world you can get parameters for dynamic world normalization.

Satellite imagery

With sentinel2 you can preprocess Sentinel-2 imagery. Preprocessing include upscaling 20- and 60-m bands to 10-m resolution, cloud masking, reprojection, clipping and normalization.

With landsat you can preprocess Landsat imagery. Preprocessing include DOS-1 atmospheric correction, cloud masking, pansharpening for Landsat 7 and 8, calculating temperature from thermal band, reprojection and clipping.

Vegetation indices

With calculate_index you can calculate normalized difference indexes like NDVI.

DEM

With dem.aspect you can calculate aspect from a DEM.

With dem.slope you can calculate slope from a DEM.

With dem.curvature you can calculate curvature from a DEM.

With dem.hillshade you can calculate hillshade from a DEM.

Mosaics

With mosaic you can merge several rasters (or multi-band products) into mosaic, fill the gaps in it, match their histograms, match it to a reference raster and clip it to ROI.

Semantic segmentation

With semantic module you can train semantic segmentation model and use it for predictions.

With semantic.generate_tiles you can generate an ML-ready dataset ,which includes cutting rasters into tiles, splitting data to subdatasets and shuffling the samples.

With semantic.train you can train a machine learning model for semantic segmentation using generated tiles.

With semantic.test you can test a semantic segmentation model.

With semantic.generate_map you can create map from predictions of pre-trained segmentation model.

With semantic.band_importance you can estimate importance of different bands for the modeling.

With semantic.confusion_matrix you can calculate a confusion matrix for a modeling result.

Regression

With regression module you can train regression model and use it for predictions.

With regression.generate_tiles you can generate an ML-ready dataset ,which includes cutting rasters into tiles, splitting data to subdatasets and shuffling the samples.

With regression.train you can train a machine learning model for regression using generated tiles.

With regression.test you can test a regression model.

With regression.generate_map you can create map from predictions of pre-trained regression model.

With semantic.band_importance you can estimate importance of different bands for the modeling.

Are you planning to add preprocessing of other imagery types (Sentinel-1, MODIS, GEOS etc.)?

Yes, but it is a long-term goal. First, I will focus on improving current functionality and adding other ML tasks (object detection, panoptic segmentation etc.) Also, you can contribute by adding your code!

I keep running into memory errors.

RSP is mostly optimized for performance rather than for memory efficiency. Despite using Dask arrays at the backend, it is still likely to fail when processing data that does not fit into memory. So, I highly recommend using a swap file on Linux and Mac or a pagefile on Windows. Also, it is planned to add Dask cluster support, but, if you read this, I still have no success with it.

I got a pytorch-related error.

You need to have Nvidia GPU in your PC to run pytorch models. Actually, it should also work on CPU, but the processing can be very slow.

If you have GPU and received this error, try to re-install pytorch using official guide. Or try another pytorch version.

If you are get such error while Sentinel-2 processing, try to run sentinel2 with upscale = "resample" or upscale = None.

I want to report an error / suggest adding a new feature

Feel free to open new tickets at https://github.com/simonreise/remote-sensing-processor/issues anytime.

How can I cite RSP?

If you use RSP in a scientific publication, we would appreciate citations: https://doi.org/10.5281/zenodo.19238835

I got error ‘Sen2Cor not working. Is it installed correctly?’.

Looks like you did not install Sen2Cor. RSP uses Sen2Cor which is installed via SNAP plugin installer. Here is the instruction how you can do it. If you don’t want to install SNAP, you can manually install Sen2Cor 2.11 or Sen2Cor 2.9 to %HOME%/.snap/auxdata/. If you installed Sen2Cor correctly, but it still does not work, you can set flag sen2cor = False.

Also, as Level-2 imagery is now widely available, we will soon discontinue support of Sen2Cor.

License

RSP is an open source software distributed under GNU General Public License v3.0