Understanding the crop cycle shift across years using Image Processing and Remote Sensing…

Original Source Here



Given an average NDVI signature for a target crop cycle for a specific year with a fixed time window (for eg: May-September, 20 weeks approximate) of a region of interest, calculate the distance between the target signature (mentioned above) and the average NDVI signature for all fixed time windows starting from January-May, February-June all the way till August-December for the year we want to calculate the crop cycle shift in a sliding window fashion. The fixed window in the comparison year with the least distance in juxtaposition with the target year will be the time window that crop cycle is shifted to. The offset is the shift in weeks is the time the crop cycle is shifted by.


While we get only a few images per month from L2A because of our cloud threshold, the idea is to compare target year time window with comparison year time windows in a weekly manner instead of monthly with some error rate to understand the shift in weeks instead of months. The algorithm below defined the steps taken to develop the above-

  1. Monthly prepared data was used to do an initial probe of missing weeks present in the datasets.
  2. Data was translated from data points available at certain days to weeks. Three cases presented itself upon doing it
  • Two dates present in the same week- average of those dates were taken to represent that week
  • Two dates present in the subsequent week- the dates were converted to week numbers and the data was just copied
  • Two dates present months/weeks apart- The missing weeks were imputed with missing values which would represent the missing week/weeks number.

3. After weekly data preparation, average signatures were extracted for the pixels in which the crop was present for target signature and for comparison windows to calculate the distance between them.

4. Number of data points considered for an entire signature were 20 and the same window was maintained throughout the sliding windows to understand the shift in crop cycle

5. Sliding was done by shifting 1 week (1 column) from the start point and 1 week from the endpoint till the last week is reached (reference dataset snapshot is found below)

6. For each comparison window , DTW distance (discussed further in the blog) is computed with the target signature and stored in order to understand which comparison window has the smallest distance with the target signature window.

7. Error interval — since multiple weeks data is missing in the analysis, there is a possibility that 2 subsequent windows can have the same data points implying that there can be an error rate of potentially multiple weeks depending on the least distance window found and its neighbouring windows distance.



Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot

%d bloggers like this: