Abstract Land cover mapping in complex environments can be challenging due to their landscape heterogeneity. With the increasing availability of various open-access remotely sensed datasets, more images acquired by different sensors and on different dates tend to be used to improve land cover classification accuracy. Selecting an appropriate feature domain with the best landscape separability is therefore crucial in meeting the requirement of computational efficiency and model interpretability. Variable selection is widely used in pattern recognition to enhance model parsimony. This study focused on the variable selection process and proposed a series of methods to select the optimal feature domain to improve land cover classification in a complex urbanized coastal area. Two decision tree models (CART-Classification and Regression Tree and CIT-Conditional Inference Tree) and five variable importance measures (GINI, PVIM-Permutated Variable Importance Measure, MD- Minimum Depth, IPM-Intervention of Prediction Measure, and CPVIM-Conditional Permutation Variable Importance Measure) based on random forests were considered. Variable importance measures were applied to a set of spectral, spatial and temporal features derived from medium-resolution satellite images. Backward elimination methods were used to select the optimal feature subset. It is found that compared to the traditional band-only model, the variable selection process can significantly improve the model parsimony and computational efficiency. The CPVIM based on CIT decision tree model was more reliable in selecting relevant features regardless their correlations, but CART tended to generate higher classification accuracy. Therefore, the combination of the CART model and the ranking from the CPVIM variable measure is recommended to achieve higher classification accuracy and better data interpretability. The novelty of our work is with the insight into the merits of integrating variable selection in the land cover classification process over complex environments.

中文翻译：

---

通过随机森林改善城市化沿海地区的土地覆盖分类：变量选择的作用

摘要 由于景观异质性，复杂环境中的土地覆盖制图可能具有挑战性。随着各种开放式遥感数据集的可用性不断提高，不同传感器在不同日期获取的更多图像倾向于用于提高土地覆盖分类的准确性。因此，选择具有最佳景观可分离性的适当特征域对于满足计算效率和模型可解释性的要求至关重要。变量选择广泛用于模式识别以增强模型简约性。本研究着眼于变量选择过程，提出了一系列选择最优特征域的方法，以改善复杂城市化沿海地区的土地覆盖分类。两个决策树模型（CART-分类和回归树和 CIT-条件推理树）和五个变量重要性度量（GINI、PVIM-置换变量重要性度量、MD-最小深度、预测度量的 IPM-干预和 CPVIM-条件排列）考虑了基于随机森林的变量重要性度量。变量重要性度量被应用于一组源自中等分辨率卫星图像的光谱、空间和时间特征。后向消除方法用于选择最优特征子集。发现与传统的band-only模型相比，变量选择过程可以显着提高模型的简约性和计算效率。基于 CIT 决策树模型的 CPVIM 在选择相关特征时更可靠，而不管它们的相关性，但 CART 倾向于产生更高的分类精度。因此，推荐结合 CART 模型和来自 CPVIM 变量度量的排序，以实现更高的分类准确率和更好的数据可解释性。我们工作的新颖之处在于深入了解在复杂环境下的土地覆盖分类过程中整合变量选择的优点。