運用深度學習建立多時空地表覆蓋辨識模型

Use deep learning to build a multi-temporal and spatial landcover segmentation and classification model

吳兆鴻(Chao-Hung Wu)

指導教授 ： 朱子豪

國立臺灣大學/理學院/地理環境資源學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102970

摘要

土地覆蓋的變遷中地表覆蓋物是非常重要的一項內容，在許多研究中地表覆蓋的分布屬性或特徵對於其研究成果或是核心都影響甚鉅，地表覆蓋的變化、分佈與人類的土地使用行為有著密不可分的關係，其通常造成快速的地物變遷情形，因此如何達到與建立時空快速變化下地物的精準判釋是極具必要性與需求性的。
而在過去直接使用多時序地物變化特徵資料建立模型的方式，通常伴隨著龐大的資料庫、運算資源與模型判釋的時間，因此如何精準的萃取與快速建立地物變化的特徵模型將是改善目前困境的重點，本研究中嘗試使用可用於時序性問題並兼備考慮空間特徵的深度學習模型，萃取地物的時空變化關係特徵並藉此特徵進行地物判釋，在本研究分析方法中提出利用ConvLSTM Autoencoder的方法進行特徵萃取，檢視過去未加入這樣萃取特徵方法與不同卷積架構大小的ConvLSTM Autoencoder的方法來進行萃取特徵並進行非監督地物分類的判釋結果比較，進而了解本研究中提出的ConvLSTM Autoencoder的深度學習特徵萃取方法對於地物分類判釋的幫助與貢獻。

關鍵字

圖像分割 ； 深度學習 ； 特徵萃取 ； ConvLSTM ； Autoencoder

並列摘要

The change of land cover is a very important topic in the geographical research. In many studies, the attributes of the land cover is a core of the research results. The change of land cover is closely related to human behavior which usually cause rapid changes in land cover. Therefore, how to build accurate model of land cover features with temporal and spatial data is extremely necessary and demanding.
In the past, the method of using multiple temporal feature data to build a model was usually accompanied by a huge database, computing resources, and running time. Therefore, how to accurately extract and build feature models of feature changes in short time, that is the core in this research. In study, we try to use a deep learning model to extract the temporal and spatial feature of sentinel 2 image. That deep learning model is autoencoder model, and it is built with ConvLSTM layer.
In this research result, we will compare the segmentation which use ConvLSTM Autoencoder and not use.

並列關鍵字

Image segmentation ； Deep learning ； Feature extraction ； ConvLSTM ； Autoencoder

參考文獻

Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S. (2010). Slic superpixels.

Byeon, W., Liwicki, M., Breuel, T. M. (2014). Texture classification using 2d lstm networks. Paper presented at the 2014 22nd international conference on pattern recognition.

Haralick, R. M., Shanmugam, K., Dinstein, I. H. J. I. T. o. s., man,, cybernetics. (1973). Textural features for image classification. (6), 610-621.

Herold, M., Scepan, J., Clarke, K. C. J. E., A, P. (2002). The use of remote sensing and landscape metrics to describe structures and changes in urban land uses. 34(8), 1443-1458.

Hochreiter, S., Schmidhuber, J. J. N. c. (1997). Long short-term memory. 9(8), 1735-1780.