Image segmentation merging on adjacent regions

Sammanfattning

Segmentation of remote sensing images plays a crucial role in forestry management, enabling the identification and delineation of distinct land segments for planning and analysis. This thesis explores novel segmentation methods aimed at improving the accuracy and efficiency of image segmentation for forested areas. The primary focus is on two segmentation methods developed in this work: the Distance Transform method and the Segment Heterogeneity method. These methods are evaluated against manually segmented ground truth data and the original AI segmentation model currently utilized by the forestry company. The Distance Transform method generates segmentation boundaries by calculating the distance of each pixel to the nearest segment boundary, while the Segment Heterogeneity method relies on analyzing the variation within each segment to improve the delineation process. Both methods are assessed using two metrics: the Distance Measurement Error, which measures the average distance between segmentation outputs and ground truth, and the Original AI Results Error, which evaluates the similarity between the new segmentation methods and the original AI model’s output. The evaluation demonstrates that both methods show promise, with the Distance Transform method achieving a mean distance measurement error of 11.4/20.2, while the Segment Heterogeneity method shows a slightly higher error of 12.8/22.2. However, when comparing similarity to the original AI output, the Segment Heterogeneity method achieves a higher mean similarity percentage of 80.94%, compared to 78.93% for the Distance Transform method, indicating its ability to retain the original segmentation characteristics better. The thesis also discusses the potential integration of alternative segmentation approaches, such as region-growing methods, and reflects on the challenges and limitations encountered throughout the study, such as computational efficiency and result consistency. Overall, this work contributes to the understanding and development of segmentation techniques in remote sensing imagery, providing valuable insights into optimizing segmentation for practical forestry applications. Future research may further refine these methods, explore alternative approaches, and evaluate their applicability to different types of terrain and imagery.