NDVI Image Classification and Polygon Detection for Agricultural Parcels

About the Project

In short, the algorithm receives a satellite image of the parcel with a NDVI filter applied, and returns a JSON/JSONLD file with information about the sub-polygons present in the parcel. The project is meant to be a component of a dashboard that helps farmers visually analyze and organize their parcels.

Algorithm pipeline:

1. Image Retrieval

• The application is built with Flask, so the coordinates of the parcel are received in the body of the request.
• The satellite image is retrieved from the open-source Terrascope service.
• The chosen image is the clearest picture taken by the satellite recently with a clarity > 50% (Less than 50% of the image is covered by clouds).

2. Image Processing

• The image is cropped and rescaled, such that only the relevant parcel is visible, not the whole bounding box.

• The parcel image is split into 3 images, according to the 3 NDVI colors that differentiate the sub-polygons.

3. Polygon Detection

• A polygon is determined by the coordinates of its corners, so it is important to find a way to detect those corners and group them in a coordinate list.
• The corners and contours of each of the 3 images is found by using the OpenCV library. Here is a representation of the corners of each image:

• The resulting corners all appear in the same list, so in order to find individual polygons we have to group the corners that appear in the same contour.
• The corners are converted from pixel coordinates to latitude and longitude in EPSG 4326 format.

4. JSON(LD) Output

• The color (classnr), area and coordinates are saved for each polygon under different names depending on the desired output.
• For simple JSON, the fields are structured in "Features" that have a class number, area and a polygon geometry which is determined by its coordinates.
• For JSONLD, the fields are organized in Management Zones which also have a class number, area and a Management Zone Geometry. The difference here however, is that the Management Zone Geometry is linked by id to the Management Zone that it belongs to.

Built With

• Python
• Flask
• JSON

Getting Started

Prerequisites

The script is written in Python 3.8.10, which can be downloaded here. The latest releases of Python also include the Python package installer (pip) which is required to install the following libraries:

• numpy
• requests
• flask
• flask-RESTful
• flask-restx
• opencv-python
• uuid
• Pillow
• pyproj

The aforementioned packages and their version are included in the file requirements.txt. They can all be installed by typing the following command in a PowerShell terminal:

pip install -r requirements.txt

Usage

The app is written with Flask and it features two endpoints that only differ in the format of the output: one for JSON and one for JSONLD. For more information about the nature of the endpoints run the application and access the Swagger documentation. Here you can find an example of a body you can use in the POST request. This body is also provided below:
{ "polygonCoordinates":"27.199243,45.910026,27.209468,45.911885,27.209607,45.906525,27.200563,45.904793",
"clientId":"36aef768-3076-43ad-8566-4818fd82c73c"
}