The aim of the project is to automatically detect anthropogenic activities that have degraded agricultural soils in the past. Two main categories have been defined: "non-agricultural activity" and "land movement". The results will make it possible to identify potentially rehabilitable soils that can be used to establish a land crop rotation map.
This project was developed in collaboration with the Canton of Ticino and of the Canton of Vaud. A detailed documentation of the project and results can be found on the STDL technical website.
Table of content
The project has been run on a 32 GiB RAM machine with a 16 GiB GPU (NVIDIA Tesla T4) compatible with CUDA.
- Ubuntu 20.04
- Python version 3.8
- PyTorch version 1.10
- CUDA version 11.3
- GDAL version 3.0.4
- object-detector version 2.3.2
Install GDAL:
sudo apt-get install -y python3-gdal gdal-bin libgdal-dev gcc g++ python3.8-dev
Python dependencies can be installed with pip or conda using the requirements.txt file (compiled from requirements.in) provided. We advise using a Python virtual environment.
- Create a Python virtual environment
$ python3.8 -m venv <dir_path>/<name of the virtual environment>
$ source <dir_path>/<name of the virtual environment>/bin/activate
- Install dependencies
$ pip install -r requirements.txt
requirements.txthas been obtained by compilingrequirements.in. Recompiling the file might lead to libraries version changes:
$ pip-compile requirements.in
The folders/files of the project proj-sda (in combination with the object-detector) are organised as follows. Path names and values can be customised by the user:
. ├── config # configurations files folder │ ├── config_det.template.yaml # detection workflow template │ ├── config_det.yaml # detection workflow │ ├── config_sandbox.yaml # sandbox workflow │ ├── config_trne.yaml # training and evaluation workflow │ └── detectron2_config.yaml # detectron 2 ├── data # folder containing the input data │ ├── AoI # available on request │ ├── DEM │ ├── FP │ ├── ground_truth │ ├── layers # available on request │ └── categories_ids.json # class dictionnary ├── functions │ ├── constants.py │ ├── fct_metrics.py │ └── fct_misc.py ├── images ├── models # trained models ├── output # outputs folders │ ├── det │ └── trne ├── sandbox │ ├── clip.py # script clipping detections to the AoI │ ├── gt_analysis.py # script plotting GT characteristics │ ├── match_colour.py # script matching colour histogram to a reference image │ ├── mosaic.py # script mosaicking images │ ├── rgb_to_greyscale.py # script converting RGB images to greyscale images │ ├── rgb_to_greyscale.sh # script converting RGB images to greyscale images │ └── tiff2geotiff.py # convert tiff to geotiff dataset ├── scripts │ ├── batch_process.sh # script to execute several commands │ ├── filter_detections.py # script detections filtering │ ├── get_dem.sh # script downloading swiss DEM and converting it to EPSG:2056 │ ├── merge_detections.py # script merging adjacent detections and attributing class │ ├── merge_years.py # script merging all year detections layers │ ├── prepare_aoi.py # script preparing the aoi shapefile for inference │ ├── prepare_data.py # script preparing data to be processed by the object-detector scripts │ └── result_analysis.py # script plotting some parameters ├── .gitignore ├── LICENSE ├── README.md ├── requirements.in # list of python libraries required for the project └── requirements.txt # python dependencies compiled from requirements.in file
Below, the description of input data used for this project.
- images: SWISSIMAGE Journey is an annual dataset of aerial images of Switzerland from 1946 to today. The images are downloaded from the geo.admin.ch server using XYZ connector.
- Area of Interests (AoIs):
- swissimage footprints: image acquisition footprints by year (swissimage_footprint_*.shp) can be found here. The shapefiles of SWISSIMAGE acquisition footprint from 2015 to 2023 are provided in this repository.
- canton: shapefile of the canton's borders used to define the AoI. The limits of Canton of Ticino and Canton of Vaud.
- ground truth: labels vectorised by the domain experts.
Disclaimer: the ground truth dataset is unofficial and has been produced specifically for the purposes of this project. - layers: list of vector layers provided by the domain experts to be spatially intersect with the results to either excluded or to add intersection information in the final attribute table. The data is available on the cantonal geoportals (Ticino and Vaud) or are available on request.
- category_ids.json: categories attributed to the detections.
- models: the trained models used to produce the results presented in the documentation are available on request.
The proj-sda repository contains scripts to prepare and post-process the data and results. Hereafter a short description of each script and a workflow graph:
prepare_aoi.py: produce an aoi shapefile compatible with a SWISSIMAGE year and desired geographical boundaries.prepare_data.py: format labels and produce tiles to be processed in the OD.results_analysis.py: plot some parameters of the detections to help understand the results (optional).merge_detections.py: merge adjacent detections cut by tiles into a single detection and attribute the class (the class of the maximum area).filter_detections.py: filter detections by overlap with other vector layers. The overlapping portion of the detection can be removed or a new attribute column is created to indicate the overlapping ratio with the layer of interest. Other information such as score, elevation, slope are also displayed.merge_years.py: merge all the detection layers obtained during inference by year.get_dem.sh: download the DEM of Switzerland.batch_process.sh: batch script to perform the inference workflow over several years.
Object detection is performed with tools present in the object-detector git repository.
The workflow can be executed by running the following list of actions and commands. Adjust the paths and input values of the configuration files accordingly. The contents of the configuration files in square brackets must be assigned.
Training and evaluation:
Prepare the data:
$ python scripts/prepare_data.py config/config_trne.yaml
$ stdl-objdet generate_tilesets config/config_trne.yaml
Train the model:
$ stdl-objdet train_model config/config_trne.yaml
$ tensorboard --logdir output/trne/logs
Open the following link with a web browser: http://localhost:6006 and identify the iteration minimising the validation loss and select the model accordingly (model_*.pth) in config_trne. For the provided parameters, model_0002999.pth is the default one.
Perform and assess detections:
$ stdl-objdet make_detections config/config_trne.yaml
$ stdl-objdet assess_detections config/config_trne.yaml
Some characteristics of the detections can be analysed with the help of plots:
$ python scripts/result_analysis.py config/config_trne.yaml
Finally, the detection obtained by tiles can be merged when adjacent and a new assessment is performed:
$ python scripts/merge_detections.py config/config_trne.yaml
Inference:
Colour processing on images can be performed if needed prior to inference with scripts available in the sandbox folder.
Copy the selected trained model to the folder models:
$ mkdir models
$ cp output/trne/logs/<selected_model_pth> models
Process images:
$ python scripts/prepare_aoi.py config/config_det.yaml
$ python scripts/prepare_data.py config/config_det.yaml
$ stdl-objdet generate_tilesets config/config_det.yaml
$ stdl-objdet make_detections config/config_det.yaml
$ python scripts/merge_detections.py config/config_det.yaml
$ scripts/get_dem.sh
$ python scripts/filter_detections.py config/config_det.yaml
The inference workflow has been automated and can be run for a batch of years (to be specified in the script) by executing these commands:
$ scripts/get_dem.sh
$ scripts/batch_process.sh
Finally, all the detection layers obtained for each year are merged into a single geopackage.
$ python scripts/merge_years.py config/config_det.yaml
Additional scripts can be used to process images. Their use is optional.
clip.py: clip a vecor layer with another one.gt_analysis.py: plot GT characteristics.match_colour.py: normalise the colour histogram to the one of a reference image.mosaic.py: mosaic images.rgb_to_greyscale.py: convert RGB images to greyscale images.rgb_to_greyscale.sh: convert RGB images to greyscale images.tiff2geotiff.py: convert tiff to geotiff files. Provide geotiff files with the same size and extent as the tiff files to georeference.
This project uses a multi-year dataset comprising greyscale and RGB images. The historical greyscale images are colourised using the method developed by Farella et al. 2022, for which the code is available here.
Depending on the end purpose, we strongly recommend users not take for granted the detections obtained through this code. Indeed, results can exhibit false positives and false negatives, as is the case in all approaches based on deep learning.