| Abstrak/Abstract |
Mangroves are vital ecosystems that require careful monitoring to prevent degradation, particularly in terms of canopy cover. Deep learning techniques are favored for developing intricate models from remote sensing data. One promising approach is deep forest, which employs tree-based learning for mapping. However, it has yet to be fully utilized in mangrove research, especially for assessing canopy cover. This study aims to examine the hyperparameter tuning of the random forest and deep forest algorithms, test the influence of input variables on canopy cover mapping, and apply and evaluate the performance of both algorithms. The random forest (RF) and deep forest (DF) algorithms were applied to PlanetScope SuperDove imagery. Several simulations were conducted to identify the optimal model, employing hyperparameter tuning through grid search optimization and a thorough analysis of input variables. The DF algorithm achieved the highest accuracy at 93.23 %, while the RF algorithm attained 88.05 %, with maximum depth being a key parameter for both. However, under different input scenarios, the RF model outperformed DF, reaching an accuracy of 69.79 % compared to DF's 68.17 %. The texture variable and the transformation index proved essential for classifying mangrove canopy cover. Overall, both algorithms effectively map mangrove canopy cover, although further research is necessary to evaluate performance across various class numbers and geographic areas. |
