PLACE

Predicting Landscapes to Advance Conservation Effectiveness

The PLACE project seeks to develop predictive models of anthropogenic land cover change based upon underlying geographic, regulatory, economic, and social data. As a proof of concept, we are conducting case studies in two focal areas: A 12-county area in central Texas. This work was funded by National Geographic and the Blumenthal Foundation.

Python code

Based on the Microsoft Landcover AI repository, this directory contains files, modules, and scripts for fine-tuning and running land cover classification models using Sentinel-2 data and the pre-trained sentinel-demo model.

R code

This directory will contain files, modules, and scripts for creating and fitting predictive models

GEE code

This directory contains javascript files used to generate and export training data and satellite imagery for land cover mapping from Google Earth Engine.

Setup

After cloning this repo, use conda env create -f conda env create -f environment_precise_landcover_prog.yml to create a conda virtual environment with all the necessary libraries to run landcover AI training.

Instructions

We fine-tune the Microsoft Landcover AI model to predict 5 land cover classes, from which changes between years can be identified: Water (1), Bare (2), Grassland/Shrub (3), Forest (4), Agriculture (5), and Impervious Surface (6). Manually curated training data points representing each desired output class are generated, and output features from the pre-trained unsuspervised sentinel U-Net model are collected at these points. This training data is passed to a Random Forest Classifier to produce a model that can predict the desired classes. We train models for each year using training data from all counties in Texas Hill Country. This method consists of Landcover-AI models consists of 6 steps.

1. Export training data points and imagery using GEE and move to Azure FileShare/Blob Storage
2. From within the PLACE repo, activate the landcover-ai virtual environment conda activate landcover-ai
3. Create config files in python/jobs for each county-year
4. Update models.json to include a model for the relevant year
5. Run 'extract.py'to collect pre-trained model output features for each county python/python3 extract.py -c {county-year config file} -m models.json
6. Run 'fine_tune.py' for the relevant year: python/python3 fine_tune.py -c {year config file} -m models.json
7. Run 'predict.py' with the relevant county config file and models.json: python/python3 predict.py -c {county-year config file} -m models.json

The result of this process will be two npy files containing pre-trained model outputs extracted at all points for a given year, and an augement.p trained model file.

Web-User Interface

Codes used to create the web-user interface are available in two other branches. Each branch has its own readme file that explains the setup and how to run the codes.