Development of an integrated method between deep learning and physical models in lowland drainage management with insufficient observed data

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K05838
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 41030:Rural environmental engineering and planning-related
• Research Institution: National Agriculture and Food Research Organization
• Principal Investigator: Kimura Nobuaki 国立研究開発法人農業・食品産業技術総合研究機構, 農村工学研究部門, 上級研究員 (40706842)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 少ないデータ量 / 深層学習 / 水位予測手法 / 物理モデル / 転移学習

Outline of Final Research Achievements

This study aimed to create a robust AI model that predicts water levels for0 severe flood events in rivers and agricultural water facilities (e.g., drainage pumping stations) using deep learning. For a small number of data samples, such as the number of large-flood occurrences, the AI models have poor performance in accurate prediction due to data-driven models that require large amounts of data. To improve this problem, first, we artificially generated a large amount of virtual data comparable to large floods using a physical model (e.g., a runoff analysis model), and used the data as training data for the AI model to construct a pre-learned model. Next, to incorporate with the observed data features, we established a highly accurate prediction method by retraining part of the pre-trained model (i.e., transfer learning) using few observed data samples.

Free Research Field

水工学

Academic Significance and Societal Importance of the Research Achievements

本研究の学術的意義について、一般に、深層学習モデルは、データサンプル数が少ない場合には、予測精度が劣るものの、その欠点を補うために物理モデルからの疑似データを割増し、さらに、転移学習で疑似データの特徴をサンプル数が少ない対象に転移することで、予測精度の向上が可能な手法（物理ガイド深層学習モデル）を開発した。
社会的意義について、現地の観測データのサンプルが少ない場合でも、物理モデルで疑似生成された大量のデータを併用して学習する、物理ガイド深層学習モデルは、現地への適用を通して、実用的に有用であることを明らかにし、さらに、データ保有に関して同様な条件の他の地区へ普及させる可能性が見出せた。