This repository presents a method of using vegetation cycle characterizations to train an irrigation detection model. A full description of this methodology can be found in Chapter IV of the following PhD dissertation.
The method of vegetation cycle characterization implemented here was first introduced in Small (2012), and involves first extracting temporal endmembers (tEMs) from the exterior of the point cloud representation of a region's vegetation phenologies. In this repository, Climate Hazards Group InfraRed Precipitation with Station Data (CHIRPS) precipitation estimates are used in the endmember selection process: Two tEMs are selected as "in-phase" and "out-of-phase" with rainfall based on the CHIRPS timeseries. These tEMs are then used to invert a linear mixture model, the pixelwise results of which constitute an "abundance map" and represent the contribution of each tEM to a vegetation signature. All spatiotemporal characterization is performed using MODIS 250m imagery, downloaded from here.
The resulting endmember abundances at 250m resolution are used to train a classifier, with labels collected via visual inspection on Google Earth Pro. All endmembers, abundance maps, imagery, labels, and predictions are contained in the linked Google Cloud bucket.
ethiopia_irrigation_detection
├── chirps_processing.py
├── dataloader.py
├── environment.yml
├── main.py
├── model.py
├── params.yaml
├── plotting.py
├── spectral_unmixing.py
Note: Each file described below also contains substantial line-by-line documentation.
-
chirps_processing.py: This script contains the functions necessary for processing CHIRPS rainfall estimates. These functions include those for loading data, clustering the rainfall timeseries, and extracting tEMs in-phase and out-of-phase with rainfall. -
dataloader.py: This script contains functions necessary for reading in abundance maps, associating certain pixels in the maps with irrigated/non-irrigated labels, and preparing these labeled samples for model training. -
environment.yml: This file specifies the Python packages required to run the code contained in this repository. Users can create the necessary Pyhton environment viaconda env create -f environment.yml. -
main.py: This is the main script for training the neural network (NN) based irrigation detection model. Run this script to execute the repository's functionality. -
model.py: This file contains code that instantiates a simple deep NN to serve as the irrigation classification model. -
params.yaml: This file contains user-specified parameters for model training. -
plotting.py: This script contains functions for plotting vegetation timeseries. -
spectral_unmixing.py: This script contains code for applying a spectral unmixing model to determine the contribution of certain temproal endmembers to vegetation phenologies. See Small (2012) for a full description of the process of using endmember-based unmixing approaches for spatiotemporal vegetation characterization.