sig-gis/gambia_pv

Training a neural network model on human labelled photovoltaic panels (pv) across the greater Banjul, The Gambia area. # Gambia PV Training a neural network model on human labelled photovoltaic panels (pv) across the greater Banjul, The Gambia area. ## Introduction A very simple workflow for preparing a dataset and training a standard unet is provided. The dataset is generated by grouping all human labelled polygons within 4 meters from each other and drawing images chips around the centroids of each grouping by windowing into the provided high resolution (30cm) imagery. This is to prevent drawing redundant image chips around photovoltatic panels where there are multiple units on a roof top or area and isolate the test samples such that those in that split are completely unseen locations. Additionally chips are generated around background polygons of confirmed locations with no photovoltaic cells. These background polygons correspond to locations where previous model iterations produced false positives. The background and pv image chips/masks are distinguished by either _pv or _bg suffix in the train and test sample directories (see below). ~100 512px pv chips and ~253 512px background chips are generated. Note: while there is little to no errors in the human labeleld polygons, there is potentially unlabelled pv panels in and around both pv and background polygons. This is however difficult to prevent due to the vast area cover and size of some panels. We consider this part of the problem to solve. Also provided is a geojson with major and minor roads across the greater Bajul area taken from open street map. From previous model iterations, we find road networks depending on locations are easily confused with pv panels due to spectrally similar properties and dust obscuring some panels. This is especially useful for filtering the final product but is by no means a comprehensive filter. Due to the small size of the training data, simple logic is provided to augment the images as they are passed through the network. This is done after post data generation to save disk space and p …