TL;DR
This study investigates the optimal way to feed environmental data into deep learning models for water quality evaluation across Japan’s 109 major watersheds. The researchers discovered that training models on raw numerical data outperforms human-preprocessed categorical data, as preprocessing often introduces bias and leads to significant information loss.
What Problem Does This Paper Address?
Evaluating and planning for river health has traditionally relied on complex mathematical models that simulate physical and chemical processes. However, these models often struggle to incorporate "big data" or the nuanced expertise of planners across different regions.
While Artificial Intelligence (AI) is a powerful alternative, a critical question remains: What is the best way to present water environment data to an AI? For example, a metric like Biochemical Oxygen Demand (BOD) can be recorded as a raw number (mg/l), a class (Class AA, A, etc.), or a simple "yes/no" regarding whether a management goal was met. This paper tests which of these formats leads to the most reliable and accurate AI predictions.
Methods
The researchers developed a layered neural network model using Deep Learning technologies. The core of their experiment involved comparing three distinct data presentation methods for the "teacher data" (training set):
- Raw Numerical Data: Continuous values (e.g., exact pH or BOD levels) without any preprocessing.
- Environmental Quality Standard (EQS) Scores: Data converted into scores (0.1 to 0.9) based on established water quality classes.
- Management Goal Achievement: Categorical "yes/no" data indicating whether administrative goals were met.
The model utilized the Error Back Propagation Algorithm and the Method of Gradient Descent to minimize error, employing a Sigmoid function for unit output.
Datasets
The study utilized a massive dataset from the Ministry of Land, Infrastructure, Transport and Tourism of Japan, covering:
- Scope: 104 of the 109 "Class-A" watersheds in Japan (5 were excluded due to missing data).
- Timeframe: 21 years of records from 1998 to 2018 for training, with 2019 data used for evaluation.
- Features: 58 items per river, categorized into Time of Sampling, River/Flow Conditions, Watershed Conditions, and various Water Quality Indexes (Living Environment, Human Health, etc.).
Key Insights and Ideas
- Preprocessing is a Double-Edged Sword: While humans use categories (like "Class A") to make data easier to understand, this study suggests that such "filtering" damages the reliability of AI models. It strips away the subtle gradients that the AI needs to understand natural evolution processes.
- The Bias of Human Logic: Preprocessing methods like EQS scoring are often subjective or arbitrary, designed for human convenience rather than mathematical precision. Feeding this "biased" information to an AI limits its ability to find the true underlying patterns.
- AI's Native Language is Raw Data: AI models are inherently capable of processing various data types simultaneously. The study confirms that they are most "reliable" when allowed to interpret the rawest form of information available.
Key Results
The models were evaluated using Accuracy and "Threat Scores" (measuring how well the model hits the target for achieved vs. unachieved goals).
| Presentation Way | Accuracy | Threat Score (Achieved) | Threat Score (Unachieved) |
|---|---|---|---|
| Raw Data | 96.88% | 96.22% | 84.71% |
| EQS Class Scores | 92.55% | 90.96% | 70.19% |
| Management Goals | 97.12% | 96.57% | 84.62% |
While the "Management Goal" approach showed slightly higher accuracy, the authors argue this is a "circular" advantage because the output format perfectly matches the evaluation metric. When looking at the ability to capture the "un-achievement" rate (Threat Score U), the Raw Data model proved exceptionally robust, confirming it as the superior method for general water environment evaluation.
Limitations and Future Work
The authors acknowledge that while raw data is best for model reliability, preprocessing might still be necessary if the only goal is a simple binary answer (e.g., "Did we pass the test?").
Future research directions include: * Identifying which specific factors (out of the 58 items) are the most influential on water quality. * Determining the most effective specific improvement measures for each of the 104 watersheds. * Refining the methodology to help establish more reliable river environment planning and management standards globally.
Summary updated