FLR-Priority-Potential

This repository contains code pipeline to assess Forest Landscape Restoration (FLR) priority scores for different types of sites, and assess the maximum restoration potential score if necessary interventions are undertaken.

Introduction

This pipeline is divided into three parts:

1. Forest Restoration Priority Score : This generates a ‘priority score’ for suitability of restoration computed from underlying variables like risk of soil erosion, distance to forest edge etc. Separate maps are generated for deforested areas and intact/degraded forests, and filtering can also be done by the history of the area (e.g. how long it has been forested).

2. Canopy Cover Density Estimation : This generates a Canopy Density Classification map over India with 3 classes - High, Medium and Low.

3. Ecological Clustering: This generates ecologically similar clusters in an area of interest based on underlying variables like mean rainfall, temperature, elevation etc. The clustering is in three levels - on each ACZ, first, a coarse 5 km resolution clustering is done using SNIC Segmentation; second, these SNIC superpixels are further divided into polygons based on FAO soil type features; third, these soil polygons are clustered using K-Means. You can find the links to all the scripts and assets here.

Forest Restoration Priority Score

1. Generate underlying variables

   a. Risk of soil erosion: This map is available publicly as the IWED dataset and is an existing GEE asset. No computation necessary.

   b. LTP_masks_and_derived_rasters.ipynb : This Python notebook is used to generate the following layers at 100m resolution for 2023:

        i. LTP mask
   
        ii. LTP area and perimeter
   
        iii. Using LTP mask, Distance from Forest Edge
   
        iv. Using LTP mask, LTP Interpatch Distance
   

   c. settlement_rasters.ipynb: This Python notebook is used to generate Distance from Settlements raster at 100m resolution for 2023. This raster is based on two datasets - WRIS’ settlement features (based on 2011 Census data) and in-lab IndiaSAT LULC’s Built-up class, moded over the previous 3 years (2021, 2022, 2023). The union of these two settlement datasets gives the mask from which distance is computed.

   d. road_water_rasters.ipynb: This Python notebook is used to generate Distance from Roads raster at 100m resolution. The Road Network is taken from the PMGSY DRRP II project and is available as GEE assets.

2. LTP_masks_and_derived_rasters.ipynb: This Python notebook is used to generate the following historical tree cover masks, available as GEE assets:

    i. forested_since_1985: Pixels that have been under tree cover continuously since 1985 to 2023.
   
    ii. forested_since_2000 : Pixels that have been under tree cover continuously since 2000 to 2023.
   
    iii. forested_1985_2000_union: Pixels that were under tree cover at any point in 1985-2000.
   
    iv. forested_2000_onwards_union: Pixels that were under tree cover at any point in 2000-2023.
   

3. restoration_priority.js: This script is used to generate and save the final Restoration Priority Score layers on a given ROI, for deforested and forested areas, masked on the type of forest (historically forested, recently forested, intact, degraded etc).

Canopy Cover Density Estimation

1. Generate binary Canopy Density Classification map (classes: High, Low) over India thresholded at the median value

2. Generate binary Canopy Density Classification map (classes: High, Low) over India thresholded at a value close to either the 25th or the 75th percentile

   a. Train Canopy Density Model.ipynb : This Python notebook was used to train the binary classifier given the classification threshold for each ACZ. It generates XGBoost models for each ACZ.

   b. Predict CCD Results.ipynb : Given the trained models from Step 2(a), this notebook was used to predict Canopy Density in 2022. It generates multiple CSVs for each district for each ACZ, containing geometry, features and prediction.

   c. uploadAssets.ipynb: This notebook is used to compile the multiple CSVs generated in Step 2(b) into larger chunks for manual uploading to GEE as Feature Collections.

   d. Next, manually upload the CSVs to GEE as Feature Collections (collections of points).

   e. fc_to_image.js: This GEE script converts the Feature Collections into rasters containing the CCD predictions. This is the final binary Canopy Density Classification map over India.

3. cd_analysis.ipynb : This notebook is used to merge the maps generated in Steps 1 and 2, resolve conflicts based on the model with the higher F1-score, and export the resulting 3-class Canopy Density Classification map.

Ecological Clustering

1. Generate underlying variables

   a. rainfall_temperature_elevation_rasters.ipynb: This notebook is used to generate all the underlying variables for the coarse SNIC-based clustering step (Step 2) at 5 km resolution.

    i. Annual rainfall and number of wet months (months with rainfall > 100 mm)
    averaged from 2003-2023.
   
    ii. Seasonal mean temperature and mean diurnal temperature range
    averaged from 2003-2023. Seasons are Rabi, Kharif, Zaid.
   
    iii. Mean elevation based on SRTM-DEM.
   

   It is also used to generate the following terrain-based variables from SRTM-DEM for the fine K-Means clustering (Step 4) at 30 m resolution.

    iv. Elevation
   
    v. Slope
   
    vi. Aspect
   

   b. road_water_rasters.ipynb: This notebook is used to generate the Distance to Water Source raster for the year 2022 at 100 m resolution. It is based on in-lab IndiaSAT LULC’s Water classes.

2. Do coarse clustering of an ACZ based on rainfall, temperature, coarse elevation (5 km)

   a. Decide how many SNIC superpixels (K) should be in a particular ACZ based on the K-Means elbow curve. Use compute_params_kmeans.js to search the K-Means elbow curve for a given feature set and ROI, compute the optimal K based on the point of maximum curvature, and save this parameter K in a GEE asset.

   b. Find size of SNIC spacing grid that will yield K superpixels. Use compute_params_snic.js after running Step 2(a) to compute the size of the spacing corresponding to the optimal K and save this parameter S in a GEE asset.

   c. Do SNIC segmentation. Use run_experiments.js after running Steps 2(a) and 2(b) to do the SNIC segmentation and visualize/export as needed.

3. Get FAO soil types features from Harmonized World Soil Database v2 and intersect these features with the SNIC superpixels generated in Step 2.

4. In each of these intersected features, do K-Means clustering based on fine elevation, slope and distance to water source

   a. Use compute_params_kmeans.js to get the optimal K based on maximum curvature of the elbow curve and save this parameter in a GEE asset.

   b. Use run_experiments.js after running Step 4(a) to do the K-Means clustering and visualize/export as needed.

5. After generating SNIC and K-Means clusters, use join_images.js to mosaic together all the ACZ-wise outputs and stack them into a single pan-India multilevel ecological clustering asset.

6. After obtaining the final clusters, use assign_potential.js to generate and save the CCD Potential table containing one row for each K-Means cluster.

7. To add new features (like rainfall, temperature, pH) to the pipeline, modify FEATURE_BUILDERS in feature_prep.js. FEATURE_BUILDERS is a dictionary mapping a feature name to a function that returns an ee.Image.

8. To visualize new features, add a visparams dictionary to VIS_PARAMS in visualization.js.

9. Helper modules: kmeans.js and snic.js.

Important Links:

1. FLR Scripts and Assets
2. FLR Report
3. India Multilevel Ecological Clusters Asset Description
4. FLR Project Presentation