3-Dimensional Clustering

5 次查看（过去 30 天）

Tyler Smith 2018-8-14

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/414804-3-dimensional-clustering

评论： Pauline Combarnous 2019-6-21

I have a 3D matrix (73 X 144 X 25738) where the first 2 dimensions are data in a lat / long grid and the 3rd dimension is time. I have transformed the data so that it is Boolean, where all 1's are the data points that meet a criteria and every other cell has a 0. I need suggestions on how to group/cluster the data by closeness in the first 2 dimensions (aka: the cells with 1's should be contiguous or within a specified distance - 2 cells would be grouped if they were both 1's and were touching/adjacent within the 2D (73 x 144) array). Furthermore, I need to maintain the 2D groups/clusters through time (aka: if there is a cluster of 5 cells on Day 1 at a given lat/long and a cluster of 8 cells around the same lat/long values on Day 2, I would like to be able to identify that the cluster of 8 cells on day 2 belongs to the same cluster as day 1). Thanks for any suggestions.

3 个评论
显示 1更早的评论隐藏 1更早的评论

Tyler Smith 2019-6-20

在 MATLAB Online 中打开

One thing I found to be close but not quite what I was looking for:

dynamic time warping: https://www.mathworks.com/help/signal/ref/dtw.html

Ultimately I ended up identifying local maximums by summing the 3D matrix I wanted to cluster. This showed me where potential regions may be. I then turned the local maximums into single grid points by calculating the weighted centroid (max_centroids) based on the area of each lat/long grid cell. The I used correlations between all the vectors in the lat/long matrix and the vectors identified by the max_centroids to produce a rough regionalization/clustering of my timeseries data. This was entirely case specific, so a similar approach may or may not work for you.

regmax = imextendedmax(Data,50);
max_centroids = regionprops(regmax,gridpoint_area,'WeightedCentroid');

Pauline Combarnous 2019-6-21

Thank you but indeed that's not quite what I was looking for, even thougth your method is interesting. I think I can find a solution using the linkage and cluster functions, I'm just trying to figure out how I could put different cutoffs on the cluster function along the different directions (lat/long and time). Anyway, thanks for your quick answer.

请先登录，再进行评论。

请先登录，再回答此问题。