Satsure

Step 1: Resampling Rasters

show-more>

  • All the tiff files were resampled to a resolution of 10m by 10m for all years and for both crops.
  • This was done to ensure that raster stacks of all the tiff files could be made and analysed together.
  • To accomplish this task, gdalUtils library was used which employs native C functions for resampling, which speeds up the entire process by a factor of 10.

<show-less

Step 2: Create Mandal Level Raster Bricks

show-more>

  • After resampling, the raster files were cropped as per the mandal shapefiles and stacked to form raster bricks for each vegetation index for each crop for each year.

<show-less

Step 3: Extraction of data from raster bricks

show-more>

  • Data was extracted from the raster files in csv format.

<show-less

Step 4: Preprocessing data

show-more>

  • Upon initial examination of the data, it was found that in many mandals, the NDVI values were not in the specified range of 100-200.
  • This was found to be due to the presence of cloud cover on the mandal.
  • Wherever there was cloud present on the mandal, the values in the raster were entered as 50, 250, NA, etc.
  • These odd values were imputed as per the following rules:
    • All NA’s were replaced with zeros.
    • If any column has value outside [100,200], replace by zero.
    • For each zero in column , impute by mean of repeated samples.
    • For late rice/kharif phenomenon , calculate the histogram of NDVI and check its skewness.
    • The first instance where skewness >=0 is taken as the cutoff point for both NDVI and NDWI time series.

<show-less

Step 5: Clustering

show-more>

  • Different clustering methods like k-means, DBSCAN, HDBSCAN and kShape were tried.
  • kShape is the only method which is able to classify the pixels into different groups.
  • kShape uses Dynamic Time Warping method to classify the pixels into different groups.
  • Raw NDVI and NDWI time-series data is the input to clustering algorithm.
  • Clustering is done for multiple number of clusters ranging from 2 to 6. The optimum number of clusters is obtained later using Kruskal Wallis method.
  • Mandal-wise summary of pixels along with group class for all K’s(2 to 6) for all years is stored.
  • Centroid information obtained after clustering is also stored.

<show-less

Step 6: Mandals Scoring

show-more>

  • Get the optimum number of clusters for each mandal for every year using Kruksal Wallis Method.
  • Calculate seasonal maximum of the geometric mean of NDVI and NDWI centriods for each cluster group of every mandal for all years.
  • Calculate mandal score at mandal and year level which is a weighted average (based on no of pixels in each cluster group) of above calculated season max across clustered groups.

<show-less

Step 7: Temperature Scaling

show-more>

  • No of days for which the ambient temperature is outside the optimum temperature range for a particular crop is calculated which are the bad days for that season.
  • Scaling factor is calculated which is 1 + (bad days/all days) for all mandals for every year.
  • Scaling factors are stored as csv which will be used in mandal grouping.

<show-less

Step 8: Grouping Mandals

show-more>

  • Score for a mandal for a particular year is calculated as per the following formula:
    where
    α is a hyperparameter ranging from 0.01 to 0.99
    yearn is nth year’s score
    score is the current year’s calculated score
  • For every α, the mandals are divided into two groups – High and Non-High – based on their calculated score.
  • This ranking is then compared with the ranking based on the actual score.
  • Raw NDVI and NDWI time-series data is the input to clustering algorithm.
  • Value of α with the highest percentage overlap is taken as the optimum α and the same is used to report the final scores and mandal groups.

<show-less