Home

Awesome

LULC mapping using time series data & spectral bands

LULC mapping is difficult because of the complexity of the land cover types and the variability of the spectral signatures. This project aims to use time series data and spectral bands to map land cover types.

In our recent work with a partner, we faced a few challenges while trying to map different forest types, such as:

  1. Data
  1. Model

In this project, we aim to solve these problems by using time series data and spectral bands. This approach has 2 advantages:

  1. We only label areas where we are most confident about the class. We don't have to aasign a classes to every pixel, this reduces the time spent on labeling and increases the accuracy of the model.
  2. We will use 1D convolutions that learn from time-series data. This will help the model generalize better and reduce the effect of clouds and shadows.

Because this is a simpler model, it takes less time to train and we can experiment faster.

As a part of labs POC:

Data

Getting the training dataset

We have data for training these models hosted on Development Seed's S3 bucket.

aws s3 ls s3://ds-data-projects/data/
                           PRE geojson/
                           PRE images_dt_12B/
                           PRE images_dt_4B/
                           PRE img_2020-03_2020-06/
                           PRE img_2020-06_2020-09/
                           PRE img_2020-09_2020-12/
                           PRE img_2020-12_2021-03/
                           PRE img_2021-03_2021-06/
                           PRE img_2021-06_2021-09/
                           PRE img_2021-09_2021-12/
                           PRE img_2022-06_2022-09/
                           PRE img_2022-12_2023-03/
                           PRE labels/
                           PRE no_rescale_color_formula/
2023-04-03 19:20:01      14340 .DS_Store
2023-04-03 19:20:01   57821721 clean_features.geojson
2023-03-04 16:10:56   42339856 clean_features.geojson.0bB292D3
2023-04-03 19:20:01      41808 clean_features.json

geojson/ - has vector masks labels/ - has raster masks images_dt_4B/ - has 4 band images images_dt_12B/ - has 12 band images no_rescale_color_formula/ - has 04 band images with no rescaling (by default, we rescale the images to 0,10000 while pulling the data from planetary computer) Ignore all the other folders

Code

I have added my nbs/ to the repository - this is where I was trying to pull data for HABs. Feel free to ignore this & start from scratch.