Forest Segmentation: Spatio-Temporal Ground Truth Labelling via Assisted Annotation

Accurate ground truth annotation is essential for training and evaluating deep learning models for remote sensing applications, particularly for tasks such as forest and non-forest segmentation. This paper aims to validate spatio-temporal datasets of ground truth labels generated through assisted annotation. This paper has three primary objectives: proposing an enhanced annotation algorithm, validating individual spatial and temporal datasets, and validating the combined spatio-temporal dataset. The methodology starts with the systematic acquisition of nine satellite images from the Endau Rompin, Royal Belum, and Tasik Chini areas, covering the years 1990, 2000, and 2010. Followed by automated ground truth annotation for forest segmentation using existing approaches. Subsequently, noise reduction is performed using SimpleBlobDetector, before being validated using the U-Net model. The results demonstrate high average precision and recall performance across multiple datasets, either evaluated individually or in combined form. Specifically, the combined training dataset achieved an average precision of 97.949
^%
and an average recall of 98.411%, confirming the reliability and effectiveness of the proposed methods for improving the accuracy of ground truth annotations. This research underscores the critical role of meticulous data collection and validation in enhancing the predictive capabilities of deep learning models for the continual detection of environmental changes.