Machine learning predicts which rivers, streams, and wetlands the Clean Water Act regulates

Abstract

We assess which waters the Clean Water Act protects and how Supreme Court and White House rules change this regulation. We train a deep learning model using aerial imagery and geophysical data to predict 150,000 jurisdictional determinations from the Army Corps of Engineers, each deciding regulation for one water resource. Under a 2006 Supreme Court ruling, the Clean Water Act protects two-thirds of US streams and more than half of wetlands; under a 2020 White House rule, it protects less than half of streams and a fourth of wetlands, implying deregulation of 690,000 stream miles, 35 million wetland acres, and 30% of waters around drinking-water sources. Our framework can support permitting, policy design, and use of machine learning in regulatory implementation problems.

Fig. 1:. WOTUS-ML Model Architecture Uses a ResNet-18 with 34 Input Layers.

The training data are images centered around an AJD. The model takes 34 input layers, described in Fig. S5. The ResNet-18 architecture begins with a convolutional block (Conv 1), followed by four residual blocks (Blocks 1 through 4), an average pooling operation, and finally a fully connected layer. The outputs of the fully connected layer are passed through a softmax function, producing a score in [0, 1]. We predict that a site is jurisdictional if the score exceeds 0.5.

Fig. 2:. WOTUS-ML Scores Allow Unbiased Estimates of Regulatory Probability.

(A) WOTUS-ML Estimates Regulation with High Accuracy for Many Sites. This figure finds threshold WOTUS-ML scores such that the average point beyond the threshold has at least a given accuracy in the AJD test set (0.95, or 0.90, etc.). The vertical axis plots the share of points with WOTUS-ML scores beyond this cutoff. This indicates, for example, for what share of points WOTUS-ML can predict AJD outcomes with 95% accuracy, for what share with 90% accuracy, etc. Fig. S9 provides details of calculations underlying (A). (B) WOTUS-ML Scores Reflect the Probability of Regulation. AJD test set is split into ten equal-width bins containing AJDs with WOTUS-ML scores of 0.0 to 0.1, 0.1 to 0.2, etc. The black lines show the average accuracy of AJDs in each bin, i.e., the share of AJDs that are jurisdictional. The model’s score is interpreted as the model’s confidence that a given AJD is determined to be WOTUS. The red bar shows the gap between the model’s average confidence and the accuracy of AJDs in each bin. The dashed 45-degree line is the ideal accuracy for each confidence level. If confidence and accuracy are equal, the model is calibrated and we can interpret the confidence score as a probability. If the red bar is below the diagonal, the model is too confident in its predictions, and vice-versa.

Fig. 3:. Estimated probability of CWA regulation for four million prediction points across the USA.

(A) Estimated probability of CWA regulation (WOTUS-ML score) under Rapanos. (B) Estimated regulatory probability under NWPR. (C) Estimated regulation changes from Rapanos to NWPR. (D) Estimated regulatory probability under CWR. (E) Estimated regulation changes from CWR to NWPR. A ‘regulation change’ describes when the WOTUS-ML binary classification score (>50%, <50%) changed status. Brown represents deregulation, green represents new regulation. Map creates 247×576 grid and displays mean model score in each bin (~28 prediction points per bin).

Fig. 4.. Case Studies of CWA Regulation Reveal Local Heterogeneity and Differences Across Rules.

(A) A densely sampled area around the Sackett property where jurisdiction is subject of Sackett v. EPA, near Priest Lake, Idaho. The property is marked with an orange star. Points on this property have mean model score of about 0.5, consistent with the ambiguity that produced litigation. Areas within and near wetlands have higher scores, though scores around the edge of the large centrally located wetland decrease under NWPR. Average model scores under Rapanos and NWPR are 0.34 and 0.22, respectively. (B) Prairie potholes (isolated wetlands), Benson County, North Dakota. Points near prairie potholes are more likely to be regulated under Rapanos and deregulated under NWPR. ACE categorizes 80% of AJDs in Benson County, ND, as isolated wetlands. Average model scores under Rapanos and NWPR are 0.22 and 0.17, respectively. (C) Farmland along the Mississippi River, Baton Rouge, Louisiana. WOTUS-ML predicts most points around the Mississippi are regulated under Rapanos and NWPR. Farmland, which the CWA explicitly ignores, receives lower scores under both rules. Average model scores under Rapanos and NWPR are 0.45 and 0.35, respectively. (D) Ephemeral streams north of Lake Mead, near Nevada, Utah, and Arizona borders. WOTUS-ML predicts ephemeral streams in the north-central part of the image lose regulation from Rapanos to NWPR. Lake Mead remains regulated under both rules. The southwestern part of the image includes Solar Energy Zones near Dry Lake, Nevada, where renewable energy development is occurring and requires AJDs. ACE categorizes 69% of AJDs in Utah and Nevada under NWPR as ephemeral streams. Average model scores under Rapanos and NWPR are 0.09 and 0.01, respectively. (E) Southern Florida, including the Everglades and Miami. WOTUS-ML predicts Rapanos and NWPR regulate Everglades National Park and most other protected wildlife areas. Wetlands and developed areas along the northwestern and eastern image edges have much lower model scores under NWPR. Average model scores under Rapanos and NWPR are 0.85 and 0.64, respectively. We randomly choose foreground/background ordering of points in all panels.