堆疊集成深度學習方法應用於建物災損影像辨識

Abstract

本研究探討了各種旨在使用無人機（Unmanned Aerial Vehicle）影像和機器學習來提升建築物損害分類的方法。以 YOLO 和 MobileNet作為基礎模型，我們提出並測試了一種創新的k1 k2重新加權方法，該方法旨在調整模型預測的類別分數並糾正分類錯誤。此研究還探究了使用元學習模型進行堆疊學習，以神經網絡和梯度提升機器模型為例，作為整合基礎模型優勢的一種方法。我們使用兩個不同的數據集，ISBDA 102和DoriaNet 216，來評估這些技術的有效性。 我們的調查強調了訓練數據多樣性對損害分類模型性能的重大影響。K1K2權重方法展示出了一些改進，但其局限性也變得明顯，並在某些錯誤類型條件和'Minor'和'Major'損害實例之間的比率分佈下顯得更為顯著。這一發現指出了需要更細緻的方法來解決預測錯誤。 我們探索了兩種類型的元學習者 - 僅具有三個隱藏層的神經網絡（Neural Network）和基礎的梯度提升機器模型（GBM）。這些元學習模型的性能在不同數據集之間產生變化。當他們在ISBDA 205上進行訓練並在ISBDA 102上進行測試時，我們觀察到了些微的性能上升和mAP分數的下降，這突出了多樣化和靈活的訓練數據的重要性。當在DoriaNet 216上進行測試時，我們還注意到了不同損害類別之間的精確度和召回率的權衡，這表明需要針對性的模型優化。 通過添加DoriaNet 55數據集，我們取得了顯著的改進，特別是元學習者在'Destroyed'建築物的分類上。對於K1K2權重方法，在大多數損害類別中，除了'Destroyed'類別，都有顯著的改善。這表明，適度增加訓練數據可以顯著提升預測準確性，強調了數據多樣性大於數量的重要性。 然而，我們還識別出K1K2權重方法的另一個局限性，主要關於被忽視的'Major'和'Destroyed'損害之間的關係。這導致在DoriaNet 216數據集上測試'Destroyed'預測的性能下降。 總的來說，這項研究強調了數據多樣性和先進的機器學習方法在提高損害分類準確性中的關鍵作用。它也強調了通過添加數據提高模型性能與保持計算效率之間的平衡。這些發現為進一步使用機器學習進行建築物損害分類的進步開闢了有希望的道路。

This research investigates various methods aimed at enhancing the classification of building damage using UAV (Unmanned Aerial Vehicle) imagery and machine learning methods. With YOLO and MobileNet as our base models, the study focuses on the application of the K1K2 reweighting methods and meta learners as tools to correct class prediction errors and improve model accuracy. Our investigation highlighted the significant impact of training data diversity on the performance of damage classification models. The K1K2 reweighting method demonstrated some improvements, but its limitations became evident, proving more effective under certain error type conditions and ratio distributions between 'Minor' and 'Major' damage instances. This finding pointed to the need for a more nuanced approach to address prediction errors. We explored two types of meta learners—neural networks with only three hidden layers and basic Gradient Boosting Machines (GBM). The performance of these meta learners varied across different datasets. Moderate performance with a decline in mAP scores was observed when they were trained on ISBDA 205 and tested on ISBDA 102, underscoring the importance of diverse and versatile training data. A trade-off between precision and recall across different damage classes was also noted when testing on DoriaNet 216, indicating the need for targeted model optimization. Significant improvements were achieved with the addition of the DoriaNet 55 dataset, especially in the classification of 'Destroyed' buildings by meta learners. For K1K2 reweighting methods, a substantial improvement was evident across most damage classes, except for 'Destroyed'. This suggests that a moderate increase in training data can notably enhance prediction accuracy, emphasizing data diversity over quantity.

However, another limitation of the K1K2 reweighting method was identified, mainly concerning the overlooked relationship between 'Major' and 'Destroyed' damages. This resulted in a decline in the performance of 'Destroyed' predictions when tested on the DoriaNet 216 dataset. In conclusion, this research underscores the critical role of data diversity and advanced machine learning methods in improving damage classification accuracy. It also stresses the balance between enhancing model performance through additional data and maintaining computational efficiency. The findings open up promising avenues for further advancements in the field of building damage classification using machine learning.