0.3.0

ML module changes

• Now semantic and regression are separate modules
• Added models from segmentation_models_pytorch and torchgeo libraries
• Added more sklearn models
• Completely changed the syntax
• Code base is almost completely rewritten
• ML datasets are now stored in .rspds format - a custom ML dataset format based on Huggingface Datasets library
• Now any custom Pytorch or Sklearn-based model can be trained
• More metrics available, including custom metrics
• Loss selection available, including custom losses
• Now user can define the augmentations that will be applied to train dataset, including custom augmentations
• Now basic multiscale features can be generated to improve the modeling quality of Sklearn-based models
• Added band_importance functions that use SHAP to estimate band importance for the modeling
• Added confusion_matrix function that calculate confusion matrix for semantic segmentation models

Other major changes

• Added process function that can clip, reproject, reproject match and change dtype of a single raster
• Added dem group of functions: slope, curvature, aspect, hillshade
• Added match_hist function that matches histograms of two images/datasets
• Added clip_values function that clips raster values to a certain range
• Added denormalize functions that restore original values from normalized data
• Added get_normalization_params functions that retrieve optimal normalization parameters
• Added zscore and dynamicworld normalization
• Now data is saved with a metadata file in STAC format (can be controlled with write_stac argument)
• Almost every function now supports not only file paths, but also STAC Items as inputs
• Most of the functions now can process multi-band datasets and STAC datasets
• Reworked calculate_index function, now it supports all the indices supported by spyndex library

Minor changes

• replace now supports multiple values replacement via values arg
• Added clip_values and nodata args to normalize function
• input_file and output_file args are renamed to input_path and output_path
• process now accepts dtype arg, which will convert input dataset to the requested dtype
• Landsat imagery is now processed by satpy
• Sentinel-2 superresolution models are now stored on Huggingface Hub
• sen2cor parameter of sentinel2 function is now False. Sen2Cor support is going to be deprecated in the future.
• The required Sen2Cor version is now 02.12.03
• Multiple performance optimisations

Deprecations

• landsat no longer supports Collection-1 products because they are no longer available to download

Remote Sensing Processor 0.2.2

This update reworks semantic segmentation functions and improves processing speed and stability

• Now uses xarray, dask and rioxarray instead of numpy and rasterio
• Now stores tiles in zarr containers instead of hdf5
• Syntax, inputs and outputs of all segmentation functions are reworked
• Custom kwargs can be used when initialising models
• Augmentations can be applied while training with augment arg
• Dataset size can be increased by repeating it n times while training with repeat arg
• Raster histograms now can be matched while creating mosaic using match_hist arg
• Specific value in a raster can be replaced using replace_value function
• Nodata value in a raster can be replaced using replace_nodata function
• Vector file can be rasterized using rasterize function
• Sentinel2 now can be upscaled using resampling algorithm. superres arg is renamed to upscale, resample arg added
• Sentinel2 now can be normalized using normalize arg
• Landsat thermal bands now can be normalized using normalize_t arg
• clipper argument is renamed to clip
• projection argument is renamed to crs

Remote Sensing Processor 0.2.1

• Added normalize function that applies min/max normalization to data
• Segmentation train and test now support multiple datasets input
• Segmentation train, test and generate_map now support multiprocessing
• Added support for more Landsat products
• Various bug fixes

Remote Sensing Processor 0.2

Remote Sensing Processor 0.2 adds image segmentation module

• Added train and test functions that train and test pytorch and sklearn segmentation models
• generate_tiles and generate map functions reworked and moved to rsp.segmentation module
• Sentinel-2 superresolution algorithm rewritten in pytorch
• normalized_difference function renamed to calculate_index

Remote Sensing Processor 0.1

Remote Sensing Processor 0.1

This is the first release of Remote Sensing Processor.

It includes Sentinel-2 and Landsat preprocessing, creating raster mosaics, calculating normalized difference indices (for now NDVI only), cutting rasters to tiles and creating maps using pre-trained models.