Agoro Field Boundary Detector

Detect field boundaries using satellite imagery.

Installation

To install this package in your environment, run:

pip install git+https://github.com/radix-ai/agoro-field-boundary-detector.git

Usage

To use the field boundary detector model together with the Google Earth Engine (model's input data), use the FieldBoundaryDetectorInterface class. This class provides an interface on top of the field boundary detector model by adding these next steps:

Translate a (latitude,longitude) coordinate into an image extracted from GEE
Predict the field-boundary of the field overlapping with the provided coordinate (image's center pixel)
Translate the extracted polygon from pixel-coordinates to geographical (lat,lng) coordinates

This interface also allows you to easily make predictions with a higher certainty, by changing the default 0.5 certainty in the class' call.

from agoro_field_boundary_detector import FieldBoundaryDetectorInterface

# Load in the model, will start GEE session
model = FieldBoundaryDetectorInterface(model_path=...)

# Make the prediction
pred = model(lat=39.6679328199836, lng=-95.4287818841267)
# Result
# [
#     (39.683761135289785, -95.4369042849299),
#     (39.6837431689841, -95.43695096539429), 
#     ...
#     (39.683761135289785, -95.4368809446977), 
#     (39.683761135289785, -95.4369042849299)
# ]

# Make the prediction with a higher certainty threshold
pred_certain = model(lat=39.6679328199836, lng=-95.4287818841267, thr=0.9)
# Result
# [
#     (39.68350960701023, -95.43681092400112), 
#     (39.683491640704545, -95.43685760446552), 
#     ...
#     (39.683491640704545, -95.43678758376893), 
#     (39.68350960701023, -95.43681092400112)
# ]

If you want to use the field boundary detector model in separation (i.e. without GEE), use the FieldBoundaryDetector class. This class will predict pixel-level polygon boundaries when fed an image. It also enables you to fetch all other polygons found in this image as well.

import numpy as np
from agoro_field_boundary_detector import FieldBoundaryDetector

# Load in the model, will start GEE session
model = FieldBoundaryDetector(model_path=...)

# Make the prediction
im = np.asarray(...)
single_polygon = model(im)
# Result
# [
#     (34, 67),
#     (41, 67), 
#     ...
#     (26, 87), 
#     (34, 67),
# ]

# Get all polygons found in the image
im = np.asarray(...)
all_polygons = model.get_all_polygons(im)
# Result
# [[
#     (34, 67),
#     ...
#     (34, 67),
# ],
# ...
# [
#     (41, 67), 
#     ...
#     (41, 67), 
# ]]

Development

Setup Environment

In order to initialise your environment properly, run the following command in the root of this project:

tasks/init.sh

This creates a conda environment with all the necessary dependencies (run and development dependencies).

If you only need the run dependencies, use:

pip install .

Exporting images

A script on how to export Google Earth Engine images - more specifically those from the National Agriculture Imagery Program dataset - can be found under i_export_field_data.py. This script utilises the functions and classes found in the google_earth_engine subfolder. Note that you need to sign-up for Google Earth Engine first (here), before you can export images from it.

Data Annotation

A script on how to augment the extracted images can be found under ii_augment_data.py. This script utilises the functions and classes found in the augmentation subfolder.

Model Training

A script on how to train, evaluate, and infer the field boundary detector model can be found under iii_train_mask_rcnn.py. This script utilises the function and classes found in the field_detection subfolder.

Model inference

The main.py file combines both the field boundary detector model, and the Google Earth Engine code in a single interface. This script allows you to easily query the model. The only things you need are a pre-trained model and the (latitude,longitude) coordinate of the field for which you want to extract the boundaries for.

dojowahi / agoro-field-boundary-detector