Finding local maxima/minima with Numpy in a 1D numpy array

In SciPy >= 0.11

import numpy as np
from scipy.signal import argrelextrema

x = np.random.random(12)

# for local maxima
argrelextrema(x, np.greater)

# for local minima
argrelextrema(x, np.less)

Produces

>>> x
array([ 0.56660112,  0.76309473,  0.69597908,  0.38260156,  0.24346445,
    0.56021785,  0.24109326,  0.41884061,  0.35461957,  0.54398472,
    0.59572658,  0.92377974])
>>> argrelextrema(x, np.greater)
(array([1, 5, 7]),)
>>> argrelextrema(x, np.less)
(array([4, 6, 8]),)

Note, these are the indices of x that are local max/min. To get the values, try:

>>> x[argrelextrema(x, np.greater)[0]]

scipy.signal also provides argrelmax and argrelmin for finding maxima and minima respectively.

If you are looking for all entries in the 1d array a smaller than their neighbors, you can try

numpy.r_[True, a[1:] < a[:-1]] & numpy.r_[a[:-1] < a[1:], True]

You could also smooth your array before this step using numpy.convolve().

I don't think there is a dedicated function for this.

For curves with not too much noise, I recommend the following small code snippet:

from numpy import *

# example data with some peaks:
x = linspace(0,4,1e3)
data = .2*sin(10*x)+ exp(-abs(2-x)**2)

# that's the line, you need:
a = diff(sign(diff(data))).nonzero()[0] + 1 # local min+max
b = (diff(sign(diff(data))) > 0).nonzero()[0] + 1 # local min
c = (diff(sign(diff(data))) < 0).nonzero()[0] + 1 # local max


# graphical output...
from pylab import *
plot(x,data)
plot(x[b], data[b], "o", label="min")
plot(x[c], data[c], "o", label="max")
legend()
show()

The +1 is important, because diff reduces the original index number.

Another approach (more words, less code) that may help:

The locations of local maxima and minima are also the locations of the zero crossings of the first derivative. It is generally much easier to find zero crossings than it is to directly find local maxima and minima.

Unfortunately, the first derivative tends to "amplify" noise, so when significant noise is present in the original data, the first derivative is best used only after the original data has had some degree of smoothing applied.

Since smoothing is, in the simplest sense, a low pass filter, the smoothing is often best (well, most easily) done by using a convolution kernel, and "shaping" that kernel can provide a surprising amount of feature-preserving/enhancing capability. The process of finding an optimal kernel can be automated using a variety of means, but the best may be simple brute force (plenty fast for finding small kernels). A good kernel will (as intended) massively distort the original data, but it will NOT affect the location of the peaks/valleys of interest.

Fortunately, quite often a suitable kernel can be created via a simple SWAG ("educated guess"). The width of the smoothing kernel should be a little wider than the widest expected "interesting" peak in the original data, and its shape will resemble that peak (a single-scaled wavelet). For mean-preserving kernels (what any good smoothing filter should be) the sum of the kernel elements should be precisely equal to 1.00, and the kernel should be symmetric about its center (meaning it will have an odd number of elements.

Given an optimal smoothing kernel (or a small number of kernels optimized for different data content), the degree of smoothing becomes a scaling factor for (the "gain" of) the convolution kernel.

Determining the "correct" (optimal) degree of smoothing (convolution kernel gain) can even be automated: Compare the standard deviation of the first derivative data with the standard deviation of the smoothed data. How the ratio of the two standard deviations changes with changes in the degree of smoothing cam be used to predict effective smoothing values. A few manual data runs (that are truly representative) should be all that's needed.

All the prior solutions posted above compute the first derivative, but they don't treat it as a statistical measure, nor do the above solutions attempt to performing feature preserving/enhancing smoothing (to help subtle peaks "leap above" the noise).

Finally, the bad news: Finding "real" peaks becomes a royal pain when the noise also has features that look like real peaks (overlapping bandwidth). The next more-complex solution is generally to use a longer convolution kernel (a "wider kernel aperture") that takes into account the relationship between adjacent "real" peaks (such as minimum or maximum rates for peak occurrence), or to use multiple convolution passes using kernels having different widths (but only if it is faster: it is a fundamental mathematical truth that linear convolutions performed in sequence can always be convolved together into a single convolution). But it is often far easier to first find a sequence of useful kernels (of varying widths) and convolve them together than it is to directly find the final kernel in a single step.

Hopefully this provides enough info to let Google (and perhaps a good stats text) fill in the gaps. I really wish I had the time to provide a worked example, or a link to one. If anyone comes across one online, please post it here!