Imbalanced learning of remotely sensed data for bioenergy source identification in a forest in the Wallacea region of Indonesia

Remote sensing technologies have been increasingly crucial to support policy-makers in achieving their ecological strategies. The data provided by such technology can estimate the bioenergy source production rate and monitor deforestation. This work participates in the cause by contributing an aerial dataset and developing an intelligent tree-detection system usable for counting trees with the bioenergy potential. Low-altitude flying units have been vastly used for such a purpose due to their ability to capture high-quality data from distant locations. Despite these potentials, collected images that compose a dataset are often characterized by imbalanced distribution among classes. The class disproportion can affect the overall model performance, as it severely deprives key features of under-represented classes. This study proposes data-level approaches that adopt and extend prior sampling algorithms for object detection problems. The devised techniques try to reduce the number of redundant outputs obtained from sampling methods and reduce the iteration required to achieve the target imbalance ratio by employing a systematic flow. In such a process, the class distribution of an original dataset is used as a guideline for selecting candidates for subsequent processes. Our results show that the modified dataset can reduce the length of a training process shown by fewer iterations required to achieve the final metrics of its original dataset version and lower training losses in each iteration. Additionally, the modified dataset can improve the F-score (F₁) and precision metric of object detection algorithm by up to 6%.