應用深度學習於時空資料預測

Applications of Deep Learning in GIS – Spatiotemporal Data Mining and Forecasting

郭有富(Supasin Wuthikulphakdi)

指導教授 ： 蔡憶佳

淡江大學/工學院/資訊工程學系全英語碩士班/碩士(2022年)

https://doi.org/10.6846/TKU.2022.00692

摘要

在本論文中，對時空數據挖掘網絡進行了廣泛的探討並使用火災事件數據集對這些網絡模型進行了比較。本文解決兩個問題： 1. 在最近提出的 STDM-DL（時空數據挖掘，深度學習）模型中，比較這些模型的預測能力？ 2. 當應用於火災數據集時，這些模型的性能如何？本論文進行了兩個實驗。第一個是使用他們的數據運行最先進的 STDM-DL 模型並比較它們的性能。本研究下的模型由 METR-LA 或 PEMS-BAY 數據集訓練，預測空間和時間域中的交通。在第二個實驗中，我們使用了新北市的火警數據集 (NTPC-Fire 2015-17) 並實現了一些熟悉但簡單的模型，例如自動編碼器和 GAN，以重建（預測）光柵化熱圖。然後，我們使用 LSTM-RNN、FBProphet 和 ARIMA 處理時間表示，以比較每日和每週事件頻率的時間序列預測的性能。我們的第一個實驗發現一些最先進的模型，例如 ST-MetaNet、STGCN 和 Spacetimeformer，都具有相似的性能。“Deepforecast Multi-LSTM”是迄今為止最好的交通預測模型。令人驚訝的是，在我們的第二個實驗中，對於我們的數據集，FBProphet 模型是我們最好的時間模型，具有 6.97231 RMSE 和 5.045342 MAE。同樣，我們重建（預測）柵格熱圖的最佳空間模型是具有 1.04198155 RMSE 和 0.3522904 MAE 的 9 批變分自動編碼器 (VAE)。鑑於這些發現，我們進一步使用數據可視化並為 STDM 任務中的每個域實施組合模型和架構。這項研究表明，這些現有模型可用於解決時空領域的預測問題。

關鍵字

調查 ； 數據挖掘 ； 深度學習 ； 地理信息系統 ； 時空數據挖掘

並列摘要

In this thesis, an extensive survey on spatiotemporal data mining networks was done, and a comparison of those network models using a fire event dataset was made. This thesis addresses two issues: 1. Among those recently proposed STDM-DL (spatial-temporal data mining, deep learning) models, how to compare those models in forecasting capability? And 2. What are the performances of those models when applied to the fire dataset from our previous search? Two experiments are performed in this thesis. The first one was that we ran the state-of-the-art STDM-DL models using their data and compared their performances. The models under this study are trained by either METR-LA or PEMS-BAY dataset, predicting the traffic in spatial and temporal domains. In the second experiment, we used the fire-call dataset of the New Taipei City (NTPC-Fire 2015-17) and implemented some familiar yet straightforward models, such as Autoencoders and GANs, to reconstruct (predict) rasterised heatmaps. Then, we processed the temporal representations using an LSTM-RNN, FBProphet, and ARIMA, to compare performance in time series forecasting of daily and weekly incident frequencies. Our first experiment discovered that some state-of-the-art models, such as ST-MetaNet, STGCN, and Spacetimeformer, all have similar performance. The “Deepforecast Multi-LSTM” is the best traffic prediction model to date. Surprisingly, in our second experiment, for our small dataset, the FBProphet model is our best temporal model with 6.97231 RMSE and 5.045342 MAE. Likewise, our best spatial model to reconstruct (predict) a raster heatmap was the 9-Batch Variational Autoencoder (VAE) with 1.04198155 RMSE and 0.3522904 MAE. Given these findings, we further use data visualisation and implement combined models and architectures for each domain in the STDM task. This study suggested that those existing models can be used to solve issues in the spatial-temporal domain, and the use of deep learning networks is a fast-growing research field that depends intensely on big data.

並列關鍵字

Survey ； Data Mining ； Deep Learning ； Geo-Information System ； Spatiotemporal Data Mining

參考文獻

[1] J. Doshi, S. Basu, and G. Pang, “From satellite imagery to disaster insights,” arXiv preprint arXiv:1812.07033, 2018.

[2] J. Doshi, “Residual inception skip network for binary segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 216–219, 2018.

[3] S. N. K. B. Amit and Y. Aoki, “Disaster detection from aerial imagery with convolutional neural network,” in 2017 International Electronics Symposium on Knowledge Creation and Intelligent Computing (IES-KCIC), pp. 239–245, 2017.

[4] V. Iglovikov, S. Mushinskiy, and V. Osin, “Satellite imagery feature detection using deep convolutional neural network: A kaggle competition,” arXiv preprint arXiv:1706.06169, 2017.

[5] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “Yolov4: Optimal speed and accuracy of object detection,” arXiv preprint arXiv:2004.10934, 2020.