Swiss houseshold-level solar panel identification from ortho-rectified satellite images with deep learning. Using the Inception-v3 framework, we performed transfer learning with the pre-trained weights from US images as a part of the DeepSolar project. The model was developed using Keras, on Python 3.6.5.
The necessary librairies are explicited in requirements.txt. To install them all at once, run
pip install -r requirements.txt
Note: if your machine is unable to find Tensorflow 1.9.0, the code should run on any version of Tensorflow between 1.15.0 and 1.9.0.
The data is obtained from SwissTopo. Unfortunately, we cannot share it publicly.
(for the CS-433 TA team) However, you can contact Dr. Roberto Castello of the LESO-PB lab if the data is absolutely necessary for grading the project.
To start off, one needs to create the folders that will contain the models. To do this, run
mkdir ckpt
cd ckpt
mkdir deepsolar_classification
mkdir inception_tl_load
mkdir inception_tl_save
The final models can be downloaded here,
and placed into inception_tl_load.
Alternatively, one can start off from scratch by downloading DeepSolar's model and creating the folders that will
contain the models. To do this, while in ckpt, run
curl -O https://s3-us-west-1.amazonaws.com/roofsolar/inception_classification.tar.gz
tar xzf inception_classification.tar.gz
We then have to transform DeepSolar's model from TensorFlow to Keras. To perform this task, return to the base directory and run
python deepsolar_to_tf.py
python tf_to_keras.py --with_aux=False
where with_aux=False creates a model with a single output, while with_aux=True uses an intermediary part of the model for predictions too and therefore produces two outputs. These two scripts take a total of approximately 20 minutes to run on a regular computer.
The script train_classification_tl.py handles all aspects of the classification task (loads data, trains and evaluates model, saves checkpoints). Have a look at the possible flags for specific use. The saved Keras model is called keras_swisspv_untrained.h5 and is found in /BASE_DIR/ckpt/inception_tl_load/.
An example is provided in run.sh
Alternatively, one can use the file train_classification_tl_aux.py to run the classification using the augmented model.
An example is provided in run_aux.sh
--ckpt_load Load the given model from a saved .h5 file. [str]
--ckpt_load_weights Load the weights by giving the path with respect to the base directory. [str]
--epochs Number of training iterations. [int]
--verbose Print out information gradually in terminal output. [bool]
--epochs_ckpt Number of training iterations after which the model is saved. [int]
--batch size Batch size. [int]
--train_set Pickle file where distribution of train images is stored. If these files exist already, the program will load from them. [str]
--test_set Pickle file where distribution of test images is stored. If these files exist already, the program will load from them. [str]
--validation_split Fraction of images used for validation. [float]
--skip_train Skip training. [bool]
--skip_train Skip testing. [bool]
--data_dir Path to the dataset. Folder must contain two sub-folders "PV" and "noPV" containing the images with PV and withouth PV, respectively. [str]
The script train_segmentation_tl.py handles all aspects of the segmentation task (loads data, trains and evaluates model, saves checkpoints). Have a look at the possible flags for specific use.
An example is provided in run_seg.sh
--ckpt_load Load the given model from a saved .h5 file. [str]
--ckpt_load_weights Load the weights by giving the path with respect to the base directory. [str]
--epochs Number of training iterations. [int]
--verbose Print out information gradually in terminal output. [bool]
--epochs_ckpt Number of training iterations after which the model is saved. [int]
--batch size Batch size. [int]
--train_set Pickle file where distribution of train images is stored. If these files exist already, the program will load from them. [str]
--test_set Pickle file where distribution of test images is stored. If these files exist already, the program will load from them. [str]
--validation_split Fraction of images used for validation. [float]
--skip_train Skip training. [bool]
--skip_train Skip testing. [bool]
--second_layer_from_ckpt Load second layer from checkpoints. [bool]
--two_layers Use two layers. [bool]
--seg_1_weights Load layer 1 weights. [str]
--seg_2_weights Load layer 2 weights. [str]
--data_dir Path to the dataset. Folder must contain two sub-folders "PV" and "noPV" containing the images with PV and withouth PV, respectively. [str]
JAKOB Anthony, DHAENE Arnaud, and ROMERO-GRASS Maëlle
This is a project in the framework of the Machine Learning CS-433 course given at EPFL (École Polytechnique Fédérale de Lausanne).
Project idea, data, and mentoring was provided by Roberto Castello from the Solar Energy and Building Physics Laboratory LESO-PB.
A detailed description of the project, as well as a concise list of references, can be found in our report in the repository.