Solar park detection from publicly available satellite imagery

The rapid increase in large-scale photovoltaic installations, or solar parks, causes a need to monitor their amount and allocation, and assess their impacts. While their spectral signature suggests that solar parks can be identified among other land covers, this detection is challenged by their low occurrence. Here, we develop an object-based random forest (RF) classification approach, using publicly available satellite imagery, which has the advantage of requiring relatively little training data and being easily extendable to large spatial extents and new areas. First, we segmented Sentinel-2 imagery into homogenous objects using a Simple Non-Iterative Clustering algorithm in Google Earth Engine. Thereafter, we calculated for each object the mean, standard deviation, and median for all 10- and 20-meter resolution bands of Sentinel-1 and Sentinel-2. These features are subsequently used to train and validate a range of RF models to select the most promising model setup. The training datasets consisted of subsampled presence/absence data, oversampled presence/absence data, and multiple land-cover categories. The best-performing model used an oversampled dataset trained on all 10- and 20- meter resolution spectral bands and the radar backscatter properties of one period. Independent test results show an overall classification accuracy of 99.97% (Kappa: 0.90). For this result, the producer accuracy was 85.86% for solar park objects and 99.999% for non-solar park objects. The user accuracy was 92.39% for solar park objects and 99.999% for non-solar park objects. These high classification accuracies indicate that our approach is suitable for transfer learning and is able to detect solar parks in new study areas.