Algorithm for detection of dead trees in multispectral images.

The work flow is developed in the following steps.

The first step is the initialization (initialize.py). It calls the clipshape functions. It clips the raster with the shape of the training set. The clipshape creates a dictionary, in which the keys are the categories provided by the training set (shapefile, see detailed description of the input below), and the values are numpy arrays with the clips of the raster. Once this object is created, it is saved into a pickle object in pickle/clip/. It overwrites, so make sure to chose a different name for the clip if you want a new one.

The training is performed by train.py. It takes in input the pickle clip (created by initialize.py) and the name you want to use for the model. It creates the model in the directory pickle/model/. If the folder doesn't exist, it creates it. It overwrites, so make sure to chose a different name for the model if you want a new file.

The classification is performed by predict.py. It takes in input the path where the orthophotos are and loops over them. It also takes as input the folder where you want to save the results, and the pickle model created by train.py. The classification is performed with KNN algorithm.

Path to training ortophoto.

The shapefile is a training set, defined by polygons containing different classes. The classes or categories are indicated in the field "zona" of the attribute table.

Example:

It is important to make sure that all the classes are well represented in the training set. The training set is created manually.

Path to orthophotos to be classified.

(Automatically created) where the result of the training is written as a pickle object.

(Automatically created) Where the clip created by the initialization will be written.

• Libspatialindex (confusionmatrix.py)
• GDAL
• Numpy
• Matplotlib
• Sklearn
• Scikit-Image (eq.py)