yasminhossam/fixMatchSeg-Muc

Semi-supervised segmentation on PV systems in Munich

PV systems semi-supervised segmentation

Semantic segmentation of PV systems based on FixMatchSeg paper and Solar panel segmentation github repository.

1. Introduction

This work uses recent overflight images from city and district of Munich and a [USA dataset] to train a semi-supervised segmentation model that identifies locations of PV systems in images.

Datasets:

1. USA dataset: The labaled dataset used in the supervised part of the model.
2. Munich dataset: The unlabeled dataset used in the unsupervised part of the model.
   It undergoes two-step augmentation using imgaug:
   1. Weak augmentation: Random rotation in the range of [−20, 20] degrees + elastic distortion.
   2. Strong augmentation on weakly augmented images: Modify sharpness, contrast and add Gaussian blur.

Model:

The segmentation model is a U-net using pixel-wise binary cross entropy as the loss function.

2. Pipeline

1. Follow the instructions in the data readme to download the data.
2. To split the unlabaled Munich data into [224,224], run:

python run.py split_images_unlabeled

3. To create masks for the labeled USA data, run:

python run.py make_masks

4. To split the images of the labeled USA data into [224,224], run:

python run.py split_images

5. Train the segmentation model

python run.py train_segmenter

For more details you can refer to the pipeline in the original Solar panel segmentation repository.

3. Results

The segmentation masks from the segmentation model were used to do image-wise classification, i.e. whether the image contains a PV system or not, since the segmentation model was originally trained on images containing PV systems and others not.

Both classification and segmentation were tested on 10% of the labeled Munich data and these were the results:
Classification:

• Accuracy: 93.26%
• Precision: 89.83%
• Recall: 99.09%
  Segmentantation dice coefficient of 0.909