Optimizing Feature Selection for Solar Park Classification: Approaches with OBIA and Machine Learning

In the context of climate change mitigation, the crucial role of solar energy as an inexhaustible and renewable source is underlined by the widespread integration of photovoltaic (PV) panels. To keep track of the rapid expansion of extensive PV systems (such as solar parks), it is important to keep monitoring their quantity, distribution, and impact. Despite the many complications involved in mapping PV systems, stemming from the diversity of materials and field layouts, many researchers have focused on ways to improve this process through Remote Sensing techniques aided by Machine Learning and Deep Learning. In this study, multiple object-based image (OBIA) classification models were created for mapping large solar parks, using open-source data (Sentinel-2) and computational tools such as Google Earth Engine (GEE) and RStudio. The case study was conducted in Pavagada Solar Park, India, which is the third largest solar park in the world spanning over 153 km \(^2\) . For training, validation, and testing purposes two zones within the park were identified. Data were collected for four seasonal periods; a classification model was developed for each season while validating them. The most accurate model was then implemented in the test area. The selection of features for the Random Forest (RF) classifier was done using the Recursive Feature Elimination (RFE) method, with the main objective of eliminating non-relevant features while preserving only the significant ones. This analysis showed a substantial improvement in accuracy, thereby validating the effectiveness of the RFE algorithm in feature selection. This is particularly true in the autumn when a maximum Overall Accuracy (OA) of 88.72% was achieved.