Parcel segmentation algorithm.

Repository structure

---

• /README.md - this file;
• /fields_recognizer.py - Python 3.8 implementation of the algorithm;
• /visualize.py - Python 3.8 script for visualizing parcel contours on the image;
• /requirements.txt - required Python packages;
• /schema.json - the json schema for validating the json-file with parcel contours.

Data description

---

MSAVI2 images

Image format: tiff.

Required data type: float32.

Required data range: [0..1].

To compute instant (obtained for a specific date) msavi2 image use follow formulas:

where , - instant satellite images of near-infrared and red channels respectively.

Aggregated msavi2 image using in the algorithm is computed with:

where (i,j) - the pixel coordinates, t - the index of the msavi2 image in the historical dataset, - the number of measurements available for (i,j)-th pixel.

Actual aggregated msavi2 lies in the range [-1..1]. For a correct result, set negative values ​​on the image to 0.

Edges map images

Image format: tiff.

Required data type: float32.

Required data range: [0..1].

The accumulated edge map is calculated as the sum of instant boundaries normalized by the number of measurements:

where - the dilated edges map extracted from the instant msavi2 image using the Canny operator. is dilated using disk structure element with radius 1.

Json-files with field contours

Information about field contours (ground truth or extracted by the algorithm) is stored in a json-file. Such file contains the following information:

• size - image size for which the field map is created;
• objects - list of field contours;
  • data - list of the contour coordinates. First item represents exterior of the contour, next items represent holes in the contour.
  • type - type of the object. It must be 'region'.

The detailed structure of the json-file with contours is described in the schema.json.

Demo materials

---

To test the program, you can download demo materials from the Zenodo repository.

The ground truth parcel contours and pre-computed images of the aggregated msavi2 and edges maps are available here.

Examples

---

Example of the algoithm usage (38ULB area).

Download the following files from repository to the folder with the fields_recognizer.py script:

• ground truth parcel contours 38ULB_markup.json (from markup.zip);
• aggregated msavi2 image 38ULB_msavi2.tiff (from demo.zip);
• edges map 38ULB_edges.tiff (from demo.zip).

Next run the aglorithm with:

python3 fields_recognizer.py 38ULB_msavi2.tiff 38ULB_edges.tiff output.json

To visualize recognized parcel contours output.json on the msavi2 image run:

python3 visualize.py output.json vis_output.tiff -src 38ULB_msavi2.tiff

The result image with contours is saved in vis_output.tiff.

You can also visualize ground truth parcel contours:

python3 visualize.py 38ULB_markup.json vis_markup.tiff -src 38ULB_msavi2.tiff

Notes

---

The algorithm does not include the step of aggregating vegetation indices and edges. Compute the aggregated msavi2 and aggregated edges map before running the script or download demo from repository.