Time Series Analysis (Python)

This code is a python reimplementation integration of typical time series analysis packages in R. Detailed API/tutorial please see Github . It contains a range of R packages and methods reimplement like GeenBrown bfast, Strucchange.

For now, It could be used to

  • Detect Annual, Seasonal, and Short Structure change and Breakpoints based on OLS-SUM/differential method (or Differential or Recursive)
  • Execute trend test like MK. Sctest
  • Seasonal Decomposition: 1.Yt=Tt+An+Tm+St+rem 2. Yt=Tt+St+Res
  • Preprocess and Spatial Statistics according to Annual Data.

For example: Seasonal Decomposition:

Summary

Spatial Statistics

Detection Label Tool for DL

There are two Modes Can be selected Object Label Tool(only line support)

Change detection(only polygon support)

Wheat Spike Counting

For some reasons, the complete code cannot be released yet.

Here is a demo on Google Colab

Dataset

  • Create labeled Wheat_multi_scale dataset(WMSD) including 33 images for current.
  • Diversity includes spatial resolution(D), Plant growth stage, Camera View

Model: MCNN model on Google Colab WDCN model on Google Colab Article on Google Doc

  • Propose a adaptive line density map algorithm based on adaptive Gaussian kernel (Point Gaussian kernel sequence according to the length)

  • Proposed a End to End Dilated network suitable for different scale scenes

  • Evaluated two network MCNN and WDCN and compare thier accuracy on WMSD and spike

  • Three images were used as pre-training due to average low value of WMSD line density map( easy to converge to 0)

Network ovall-MAE overall-MSE SPIKE MAE overall
MCNN 8.73 12.38 7.25
Tasselnet 9.7 13.8 5.96
WDCN 4.41 6.27 2.83

Change Detection

For some reasons, the complete code cannot be released yet.

Here is a demo on Google Colab

HD map for Auto-Driving

The photos are collected by mobile phone through crowdsourcing . So accuracy is higher than 10m. In order to achieve cm level, SFM needs to be applied to reconstructure relative position.

Study Area

Nerual Network

Use SFM method to enhance Position

Result Sample