Hybrid Learning Model for Satellite Forest Image Segmentation

Abstract

Image segmentation is an essential image processing technique as the quality of individual object detection significantly affects subsequent global image classification accuracy. The segmentation process can be performed by a varying number of different algorithms, but to date, these different algorithms are not yet able to guarantee a level of performance similar to or superior to human capability. This study adopts a supervised approach toward satellite forest image segmentation. The proposed model used a feature vector obtained through transfer learning from ResNet50; these features were then passed to a Random Forest for segmentation. The satellite images used for training and testing were obtained from the Land Cover Classification Truck in DeepGlobe Challenge. Metrics such as precision, recall, F1-Score, accuracy, Root Mean Square Error (RMSE), and Mean Average Error (MAE) were used to assess the performance of the model. The model achieved a testing accuracy of 94%, RMSE value of 0.2499, and MAE value of 5.92. For detecting forest areas the proposed model obtained a precision of 0.94, recall of 0.96, and F1-Score of 0.95. For non-forest areas, the proposed model achieved a precision of 0.93, recall of 0.89, and F1-Score of 0.91.