Also known as not supported.

Methods discussed and provided here have no guarantee to work or provide any meaningful results. These are somehow abandoned projects and unfortunately mid-way through. They aren’t completely discarded simply because of hope that maybe someday someone will come and help fix them. We all know that the best motivation to do something is to be annoyed by the current state. Seriously though, mode decomposition in 2D and multi-dim is an interesting topic. Please?



To use JitEMD you need to install PyEMD wiht jit option. If you’re using pip then use this command

pip install EMD-signal[jit]

Uses Numba (https://numba.pydata.org/) as a Just-in-Time (JIT) compiler for Python, mostly Numpy. Just-in-time compilation means that the code is compiled (machine code) during execution, and thus shows benefit when there’s plenty of repeated code or same code used a lot.

This EMD implementation is experimental as it only provides value when there’s significant amount of computation required, e.g. when analyzing HUGE time series with a lot of internal complexity, or reuses the instance/method many times, e.g. in a script, iPython REPL or jupyter notebook.

Additional reason for this being experimental is that the author (me) isn’t well veristile in Numba optimization. There’s definitely a lot that can be improved. It’s being added as maybe it’ll be helpful or an inspiration for others to learn something and contribute to the PyEMD.


There’s an example which compares the pefromance between JIT and classic EMD.

❯ python example/emd_comparison.py

Comparing EEMD execution on a larger signal with classic and JIT EMDs. Signal is random (uniform) noise of length: 2000. The test is done by executing EEMD with either classic or JIT EMD 20 times and taking the average. Such setup favouries JitEMD which is compiled once and then reused 19 times. Compiltion is quite costly.

Classic EEMD on 2000 length random signal: 5.7 s per EEMD run

JitEMD EEMD on 2000 length random signal: 4.2 per EEMD run


There are two ways of interacting with JIT EMD; either as a function, or as a class compatible with the rest of PyEMD ecosystem. Please take a look at code examples for a quick start.

Class JitEMD

When using class experimental.JitEMD it’ll be compatible with other PyEMD classes, for example with EEMD. That’s why it’ll accept the same inputs and will provide the same outputs. The only difference is in configuring the class. It now has to be a copy of default_emd_config with updated values.

from PyEMD.experimental.jitemd import default_emd_config, JitEMD, get_timeline

rng = np.random.RandomState(4132)
s = rng.random(500)
t = get_timeline(len(s), s.dtype)

config = default_emd_config
config["FIXE"] = 4
emd = JitEMD(config=config, spline_kind="akima")
imfs = emd(s, t)

Function JIT emd

When using emd function directly, you only get the imfs as the result and only once. There’s also a differnce where the config is passed; it’s directly to the method. Other than that, it should be the same. The class JitEMD uses this function and provides some abstraction on to of it for, hopefully, easier use, but there might be benefits to access the function directly.

from PyEMD.experimental.jitemd import default_emd_config, emd, get_timeline

s = rng.random(500)
t = get_timeline(len(s), s.dtype)

config = default_emd_config
config["FIXE"] = 4
imfs = emd(s, t, spline_kind="cubic", config=config)



Important This is an experimental module. Please use it with care as no guarantee can be given for obtaining reasonable results, or that they will be computed index the most computation optimal way.


BEMD performed on bidimensional data such as images. This procedure uses morphological operators to detect regional maxima which are then used to span surface envelope with a radial basis function.




Important This is an experimental module. Please use it with care as no guarantee can be given for obtaining reasonable results, or that they will be computed index the most computation optimal way.


EMD performed on images. This version uses for envelopes 2D splines, which are span on extrema defined through maximum filter.


class PyEMD.EMD2d.EMD2D(**config)[source]

Empirical Mode Decomposition on images.

Important This is an experimental module. Experiments performed using this module didn’t provide acceptable results, either in actual output nor in computation performance. The author is not an expert in image processing so it’s very likely that the code could have been improved. Take your best shot.

Method decomposes images into 2D representations of loose Intrinsic Mode Functions (IMFs).

The current version of the algorithm detects local extrema, separately minima and maxima, and then connects them to create envelopes. These are then used to create a mean trend and subtracted from the input.

Threshold values that control goodness of the decomposition:
  • mse_thr — proto-IMF check whether small mean square error.

  • mean_thr — proto-IMF chekc whether small mean value.

__call__(image, max_imf=-1)[source]

Call self as a function.


list of weak references to the object (if defined)

check_proto_imf(proto_imf, proto_imf_prev, mean_env)[source]

Check whether passed (proto) IMF is actual IMF. Current condition is solely based on checking whether the mean is below threshold.

proto_imfnumpy 2D array

Current iteration of proto IMF.

proto_imf_prevnumpy 2D array

Previous iteration of proto IMF.

mean_envnumpy 2D array

Local mean computed from top and bottom envelopes.


Whether current proto IMF is actual IMF.

emd(image, max_imf=-1)[source]

Performs EMD on input image with specified parameters.

imagenumpy 2D array,

Image which will be decomposed.

max_imfint, (default: -1)

IMF number to which decomposition should be performed. Negative value means all.

IMFsnumpy 3D array

Set of IMFs in form of numpy array where the first dimension relates to IMF’s ordinary number.

classmethod end_condition(image, IMFs)[source]

Determins whether decomposition should be stopped.

imagenumpy 2D array

Input image which is decomposed.

IMFsnumpy 3D array

Array for which first dimensions relates to respective IMF, i.e. (numIMFs, imageX, imageY).


Calculates top and bottom envelopes for image.

imagenumpy 2D array
min_envnumpy 2D array

Bottom envelope in form of an image.

max_envnumpy 2D array

Top envelope in form of an image.

classmethod find_extrema(image)[source]

Finds extrema, both mininma and maxima, based on local maximum filter. Returns extrema in form of two rows, where the first and second are positions of x and y, respectively.

imagenumpy 2D array

Monochromatic image or any 2D array.

min_peaksnumpy array

Minima positions.

max_peaksnumpy array

Maxima positions.

classmethod prepare_image(image)[source]

Prepares image for edge extrapolation. Method bloats image by mirroring it along all axes. This turns extrapolation on edges into interpolation within bigger image.

imagenumpy 2D array

Image for which interpolation is required,

imagenumpy 2D array

Big image based on the input. Grid 3x3 where the center block is input and neighbouring panels are respective mirror images.

classmethod spline_points(X, Y, Z, xi, yi)[source]

Interpolates for given set of points