Machine learning system and method for predicting clean water act jurisdiction

Abstract

An approach for determining regulatory status of water resources under the Clean Water Act includes receiving input data layers for a geographic location, preprocessing the layers into a multi-channel image, feeding the image into a trained convolutional neural network, obtaining an output score indicating a probability of regulated waters, and determining regulatory status based on the score. The input layers may include aerial imagery, wetland / stream data, soil / elevation / climate data, land cover classification, and regulatory boundary data. The neural network may be trained on approved jurisdictional determinations. The method enables efficient, consistent predictions of Clean Water Act jurisdiction across diverse water resources and geographic regions. By incorporating regulatory information alongside environmental data, the approach accounts for variations in Clean Water Act implementation across jurisdictions and regulatory regimes. The scalable method supports rapid evaluation of large areas for policy analysis and environmental planning.

Description

MACHINE LEARNING SYSTEM AND METHOD FOR PREDICTING CLEAN WATERACT JURISDICTIONCROSS REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims priority to U. S. Provisional Application No. 63 / 513,464, entitled “WATERS OF THE UNITED STATES - MACHINE LEARNING”, and fded on July 13, 2023. The entire contents of the above-listed application are hereby incorporated by reference for all purposes.GOVERNMENT SUPPORT

[0002] This invention was made with government support under Grant Number AG079914 awarded by the National Institutes of Health. The government has certain rights in the invention.TECHNICAL FIELD

[0003] The present description relates to machine learning, and more particularly to automatically predicting whether a given water resource is subject to regulation under the Clean Water Act (CWA) based on geospatial data and regulatory determinations.BACKGROUND

[0004] The Clean Water Act (CWA) of 1972 represents a cornerstone of federal water quality regulation in the United States. The CWA aims to protect waterways for flood protection, clean drinking water, ecosystem health, and economic activity. However, the CWA does not explicitly define which waters are protected under the phrase "Waters of the United States" (WOTUS), leading to ongoing debates and legal challenges regarding the scope of CWA jurisdiction.

[0005] In recent years, several judicial and executive rules have substantially rewritten CWA coverage, including the Rapanos decision, the Clean Water Rule (CWR), and the Navigable Waters Protection Rule (NWPR). These rules have provided differing interpretations of which water resources fall under CWA jurisdiction. For example, the 2006 Rapanos decision found that the CWA protects water resources with a "significant nexus" to navigable waters, while the NWPR excluded isolated wetlands and ephemeral streams from protection.

[0006] The lack of a clear definition for WOTUS has made it challenging to preciselyunderstand which waters gain or lose protection under different rules. Existing tools and datasets, such as the National Wetlands Inventory (NWI) and National Hydrography Dataset (NHD), provide useful but incomplete information about water resource types. However, these tools alone are insufficient to determine whether a particular water resource is jurisdictional under the CWA.

[0007] From a technical standpoint, automatically determining CWA jurisdiction for water resources presents several challenges. The complex nature of CWA rules makes it difficult to translate legal criteria into quantifiable features that can be used in convolutional neural networks. For instance, the "significant nexus" standard established in Rapanos requires assessing biological, chemical, and physical connections between water resources, which may not be readily apparent from remote sensing data or geospatial information alone. Furthermore, the relevant features needed to automatically determine CWA jurisdiction can vary widely depending on the specific rule being applied and the characteristics of the water resource in question. For example, factors such as flow regime, proximity to navigable waters, and the presence of surface water connections may all influence regulatory status, but their relative importance can change based on the applicable rule and local conditions. This variability makes it challenging to develop a single, comprehensive set of features that can accurately predict jurisdiction across diverse water resources and geographic regions.

[0008] The process of determining CWA jurisdiction for specific water resources typically involves Approved Jurisdictional Determinations (AJDs) made by Army Corps of Engineers (ACE) engineers. These case-by-case decisions require evaluation of geophysical conditions and can be time-consuming and resource-intensive. Scaling this process to cover large geographic areas or numerous water resources is particularly challenging, as it requires significant human expertise and on-site assessments. Attempts to automate or streamline jurisdictional determinations using machine learning techniques face several obstacles. First, the limited availability of high-quality, labeled training data that accurately represents the full range of water resource types and jurisdictional decisions across different rules and geographic regions can hinder model development and generalization. Second, the complex spatial relationships and contextual information that inform jurisdictional decisions may be difficult to capture using traditional machine learning approaches. Additionally, the dynamic nature of water resources and the potential for changes in legal interpretations over time introduce temporal challenges. Models trained on historical data may not accurately reflect current jurisdictional standards or account forseasonal variations in water resource characteristics.

[0009] Given these challenges, there is a need for more accurate and efficient methods to determine CWA jurisdiction across diverse water resources and geographic regions. Improved approaches could benefit various stakeholders, including regulatory agencies, developers, environmental organizations, and policymakers, by providing more reliable information on the scope of CWA protection and streamlining the jurisdictional determination process.SUMMARY

[0010] The present disclosure at least partially addresses the issues identified above, by providing a method for automatically determining regulatory status of water resources under the CWA. The method includes receiving a plurality of input data layers for a geographic location, where the input data layers comprise aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data for the geographic location. The plurality of input data layers are preprocessed into a multichannel image, wherein each pixel of the multi-channel image corresponds to the geographic location and comprises values from the plurality of input data layers. The multi-channel image is then fed into a trained convolutional neural network comprising a plurality of convolutional layers. An output score is obtained from the trained convolutional neural network, indicating a probability that the geographic location contains waters regulated under the Clean Water Act. Based on the output score, a regulatory status for the geographic location is determined, wherein the geographic location is determined to contain regulated waters when the output score exceeds a threshold value.

[0011] The above approach leverages machine learning techniques to analyze complex geospatial data and regulatory decisions, enabling more efficient and consistent jurisdictional determinations compared to conventional case-by-case assessments. By incorporating regulatory boundary data, such as state borders, Army Corps of Engineers (ACE) district boundaries, and applicable Clean Water Act rules, as input data layers alongside geophysical data, the convolutional neural network can capture both environmental characteristics and regulatory context that inform jurisdictional decisions. This integration of regulatory information allows the model to account for variations in Clean Water Act implementation across different jurisdictions and regulatory regimes. The multi-channel input image, combining diverse geospatial layers with encoded regulatory data, enables the trained model to learn relationships between physicallandscape features and jurisdictional outcomes specific to particular regulatory contexts. As a result, the method can provide more nuanced and locally-relevant predictions of regulatory status across diverse water resource types and geographic regions. Additionally, the model's ability to output a regulatory probability, rather than a binary classification, provides valuable uncertainty information that can be used to prioritize manual review of ambiguous cases or inform risk assessments in development projects, taking into account the specific regulatory landscape of a given location.

[0012] Furthermore, the scalability of the approach allows for rapid evaluation of CWA jurisdiction over large areas, potentially supporting more comprehensive policy analysis and environmental planning. The method's flexibility in incorporating different regulatory rules through appropriate training data also enables adaptation to evolving legal interpretations of CWA jurisdiction. Overall, this approach significantly streamlines the process of determining CWA jurisdiction while providing a quantitative basis for decision-making in water resource management and protection.

[0013] It should be understood that the summary above is provided to introduce in simplified form a selection of concepts that are further described in the detailed description. It is not meant to identify key or essential features of the claimed subject matter, the scope of which is defined uniquely by the claims that follow the detailed description. Furthermore, the claimed subject matter is not limited to implementations that solve any disadvantages noted above or in any part of this disclosure.BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present disclosure will be better understood from reading the following description of non-limiting embodiments, with reference to the attached drawings, wherein below:

[0015] FIG. 1 is a schematic diagram illustrating an example data processing system for determining regulatory status of water resources, in accordance with an embodiment;

[0016] FIG. 2 is a flowchart illustrating a method for determining regulatory status of water resources using a convolutional neural network, in accordance with an embodiment;

[0017] FIG. 3 is a block diagram illustrating components of a system for determining regulatory status of water resources using a convolutional neural network, in accordance with an embodiment;

[0018] FIG. 4 is a flowchart illustrating a method for generating a training dataset for training a convolutional neural network to determine regulatory status of water resources, in accordance with an embodiment;

[0019] FIG. 5 is a flowchart illustrating a method for generating multi-channel input images for training or using a convolutional neural network to determine regulatory status of water resources, in accordance with an embodiment;

[0020] FIG. 6 is a flowchart illustrating a method for training a convolutional neural network to determine regulatory status of water resources, in accordance with an embodiment;

[0021] FIG. 7 is a flowchart illustrating a method for incorporating updated regulatory rules into a convolutional neural network for determining regulatory status of water resources, in accordance with an embodiment;

[0022] FIG. 8 is a flowchart illustrating a method for determining the importance of input data layers in a convolutional neural network for determining regulatory status of water resources, in accordance with an embodiment;

[0023] FIG. 9 shows two charts comparing feature importance at national and state levels for various environmental and geographical factors, in accordance with an embodiment; and

[0024] FIG. 10 is a graphical depiction of estimated probabilities of Clean Water Act regulation across the United States under different regulatory frameworks, in accordance with an embodiment.DETAILED DESCRIPTION

[0025] The present disclosure provides systems and methods for automatically determining the regulatory status of water resources under the Clean Water Act (CWA) using machine learning techniques applied to geospatial data. The systems and methods described herein address several challenges in the field of water resource management and environmental regulation.

[0026] One issue in the field is the lack of a clear, consistent definition of "Waters of the United States" (WOTUS) under the CWA. This ambiguity has led to ongoing debates and legal challenges regarding the scope of CWA jurisdiction. Recent judicial and executive rules, such as the Rapanos decision, the Clean Water Rule (CWR), and the Navigable Waters Protection Rule (NWPR), have provided differing interpretations of which water resources fall under CWA jurisdiction. For example, the 2006 Rapanos decision established that the CWA protects water resources with a"significant nexus" to navigable waters, while the NWPR excluded isolated wetlands and ephemeral streams from protection. This lack of consistency makes it challenging to determine precisely which waters are protected under different regulatory regimes. Another challenge is the resource-intensive nature of traditional jurisdictional determinations. Typically, Approved Jurisdictional Determinations (AJDs) are made by Army Corps of Engineers (ACE) engineers on a case-by-case basis, requiring evaluation of geophysical conditions and often necessitating onsite assessments. This process is time-consuming and difficult to scale, particularly when considering the vast geographic areas that may need to be evaluated.

[0027] The current disclosure addresses these challenges by leveraging machine learning techniques, specifically convolutional neural networks (CNNs), to analyze complex geospatial data and regulatory decisions. This approach enables more efficient and scalable jurisdictional determinations compared to conventional case-by-case assessments. The systems and methods described herein utilize a multi-channel input image that combines diverse geospatial data layers with encoded regulatory information. This approach allows the trained model to learn relationships between physical landscape features and jurisdictional outcomes specific to particular regulatory contexts. The input data layers include aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data for the geographic location of interest. Incorporation of regulatory boundary data, such as state borders, Army Corps of Engineers (ACE) district boundaries, and applicable Clean Water Act rules, as input data layers alongside geophysical data allows the model to account for variations in Clean Water Act implementation across different jurisdictions and regulatory regimes.

[0028] Further, by learning a mapping from geospatial and regulatory data to a regulatory probability, rather than a simple binary classification, the trained CNN provides valuable uncertainty information that can be used to prioritize manual review of ambiguous cases or inform risk assessments in development projects. In some embodiments, the system determines output scores for a plurality of sub-regions within a larger geographical location by inputting multichannel input images for each sub-region into the trained convolutional neural network. The system then identifies one or more of the sub-regions having a highest probability of containing regulated waters based on the output scores. This allows for a more granular analysis of potential regulated waters within a larger area of interest.

[0029] In one embodiment, a data processing system 100 for determining regulatory status ofwater resources under the Clean Water Act is depicted in FIG. 1. The data processing system 100 includes a processor 102, non-transitory memory 104, and a convolutional neural network 106 for analyzing geospatial and regulatory data to predict Clean Water Act jurisdiction. FIG. 2 illustrates a method 200 for determining regulatory status of water resources using the convolutional neural network, which includes obtaining input data layers, preprocessing the data into a multi-channel image, and feeding the image into the trained convolutional neural network to obtain an output score indicating the probability of regulated waters. Input data layers and components of the convolutional neural network are further detailed in the block diagram of FIG. 3, which shows the relationships between input data layers 302, the convolutional neural network 330, and the output regulatory status 340, hydrologic type 342, and legal classification 344.

[0030] The process of generating a training dataset for the convolutional neural network is outlined in FIG. 4, which illustrates a method 400 for obtaining and processing Approved Jurisdictional Determinations (AJDs) to create labeled training data. FIG. 5 provides a detailed flowchart of a method 500 for generating multi-channel input images, including steps for extracting various geospatial and regulatory data layers and compiling these layers into a multichannel image. The training process for the convolutional neural network is depicted in FIG. 6, which shows a method 600 involving the generation of a training dataset, initialization of network parameters, and iterative updating of these parameters based on computed multi-tasking loss.

[0031] FIG. 7 illustrates a method 700 for incorporating updated regulatory rules into the convolutional neural network, allowing the system to adapt to changes in Clean Water Act interpretation. This method includes steps for updating the definition of waters covered by the CWA, generating an embedding vector for the updated definition, and creating a rule encoding layer from the embedding vector. FIG. 8 presents a method 800 for determining the importance of input data layers in the convolutional neural network, which involves permuting values of selected input data layers and calculating the impact on prediction accuracy. FIG. 9 shows two charts comparing feature importance (as may be determined using method 800) at national and state levels for various environmental and geographical factors, illustrating the relative importance of different input features used by the trained convolutional neural network to predict Clean Water Act jurisdiction. FIG. 10 provides a graphical depiction of estimated probabilities of CWA regulation across the United States under different regulatory frameworks, showcasing the output of the trained convolutional neural network for multiple prediction points and visualizing changesin regulatory coverage under various CWA rules.

[0032] Referring to FIG. 1, a data processing system 100 for automatically determining regulatory status of water resources under the Clean Water Act is shown, in accordance with an exemplary embodiment. The data processing system 100 includes a processor 102, non-transitory memory 104, a convolutional neural network 106, a training module 108, a multi-channel image generation module 110, input data layers 112, and an alert generation module 114. These components work together to process geospatial and regulatory data, generate multi-channel images, analyze them using the convolutional neural network, and produce regulatory status predictions and alerts. The data processing system 100 may also include a display 154 and an input device 156 for user interaction, and can communicate with external user devices 176 and data sources 158. In some embodiments, data processing system 100 is incorporated into a geographic information system (GIS) or environmental monitoring system. In some embodiments, at least a portion of data processing system 100 is disposed at a device (e.g., edge device, server, etc.) communicably coupled to a satellite imagery system via wired and / or wireless connections. In some embodiments, at least a portion of data processing system 100 is disposed at a separate device (e g., a workstation) which can receive geospatial data from various data sources or from a storage device which stores the data generated by environmental monitoring systems.

[0033] Data processing system 100 includes a processor 102 configured to execute machine readable instructions stored in non-transitory memory 104. Processor 102 may be single core or multi-core, and the programs executed thereon may be configured for parallel or distributed processing. In some embodiments, processor 102 may optionally include individual components that are distributed throughout two or more devices, which may be remotely located and / or configured for coordinated processing. In some embodiments, one or more aspects of processor 102 may be virtualized and executed by remotely-accessible networked computing devices configured in a cloud computing configuration. Processor 102 may be configured to implement various methods for preprocessing geospatial data, training and executing convolutional neural networks, and generating regulatory status predictions and alerts.

[0034] Non-transitory memory 104 may store a convolutional neural network 106, a training module 108, a multi-channel image generation module 110, input data layers 112, and an alert generation module 114. Convolutional neural network 106 may include one or more deep learning networks, comprising a plurality of weights and biases, activation functions, and instructions forimplementing the one or more machine learning models to process an input multi-channel image. For example, convolutional neural network 106 may store instructions for preprocessing a multichannel image, and mapping the preprocessed image to a regulatory status prediction, hydrologic type prediction, and legal resource type prediction.

[0035] Training module 108 may be configured to train the convolutional neural network 106 using a dataset of approved jurisdictional determinations (AJDs) issued by the U.S. Army Corps of Engineers. Training module 108 may implement methods for training the convolutional neural network 106, including gradient descent algorithms, loss functions, and other training routines to improve performance of convolutional neural network 106 in predicting jurisdictional determinations for water resources under the Clean Water Act. In one embodiment, training module 108 may implement a method for generating a training dataset comprising a plurality of AJDs, preprocessing the training dataset by generating multi-channel input images, and training the convolutional neural network 106 to map the multi-channel input images to predicted regulatory statuses, Cowardin wetland classification codes, and legal resource types.

[0036] Multi-channel image generation module 110 may be configured to preprocess a plurality of input data layers into a multi-channel image for a geographic location of interest. In one embodiment, multi-channel image generation module 110 may generate a multi-channel input image by combining various geospatial and regulatory data layers, such as aerial imagery data, wetlands data, streams data, soil data, groundwater data, elevation data, climate data, land cover data, and geographic data identifying applicable regulatory boundaries and rules.

[0037] Input data layers 112 may comprise a plurality of geospatial and regulatory datasets used to generate the multi-channel input images. These may include, but are not limited to, aerial imagery from the National Agriculture Imagery Program (NAIP), wetland location data from the National Wetlands Inventory (NWI), stream location data from the National Hydrography Dataset (NHD), soil data from the Gridded National Soil Survey Geographic Database (gNATSGO), elevation data from USGS's 3DEP model, climate data from the PRISM Climate Group's 30-year Normals, land cover classification data from the National Land Cover Database (NLCD), and regulatory boundary data such as state borders, Army Corps of Engineers (ACE) district boundaries, and applicable Clean Water Act rules.

[0038] Alert generation module 114 may be configured to generate alerts or visualizations based on the regulatory status determinations made by the convolutional neural network 106. Inone embodiment, alert generation module 114 may generate an alert indicating whether a geographical location contains waters regulated under the Clean Water Act. This alert may be in the form of a message sent to a user device, specifying one or more locations that have been classified as containing regulated waters. In another embodiment, alert generation module 114 may generate a visualization displaying the geographical location and an indication of the type of water resource classified by the convolutional neural network.

[0039] Data processing system 100 may further include a display 154 and an input device 156. Display 154 may be configured to present visualizations, alerts, and other outputs generated by the system to a user. Input device 156 may comprise one or more of a touchscreen, a keyboard, a mouse, a trackpad, or other device configured to enable a user to interact with and provide inputs to data processing system 100. In one example, input device 156 may enable a user to specify a geographical location for regulatory status determination or to select specific data layers for inclusion in the multi-channel input image. In some embodiments, the processor is configured to generate a map overlay highlighting sub-regions identified as having a high probability of containing regulated waters. This map overlay may be displayed superimposed on the aerial imagery data on a user interface to visually indicate areas within the geographical location most likely to contain waters regulated under the Clean Water Act. This visual representation allows users to quickly identify areas of potential regulatory concern within a larger geographical area.

[0040] User device 176 may be communicatively coupled to data processing system 100 via one or more networks. User device 176 may be a mobile device, such as a smartphone or tablet, or a desktop computer used by environmental regulators, land developers, or other stakeholders to access the regulatory status predictions and alerts generated by data processing system 100. In some embodiments, user device 176 may host a web-based or local client application that interfaces with data processing system 100 to request regulatory status determinations for specific geographic locations.

[0041] Data sources 158 may comprise various external databases, servers, or other data providers that supply the input data layers 112 used by data processing system 100. These may include government agencies such as the U.S. Geological Survey, the Environmental Protection Agency, and the Army Corps of Engineers, as well as private data providers. Data sources 158 may be connected to data processing system 100 via APIs, database connections, or other data transfer mechanisms, allowing for regular updates to the input data layers used in regulatory statusdeterminations.

[0042] The data processing system 100 enables automatic determination of the regulatory status of water resources under the Clean Water Act. By integrating diverse geospatial and regulatory data sources, it enables more comprehensive and accurate predictions of jurisdictional status. The system's ability to process large volumes of data and generate rapid predictions can significantly streamline regulatory decision-making processes. Furthermore, the alert generation module allows for efficient communication of results to relevant stakeholders, potentially reducing the time and resources required for regulatory compliance. The system's flexibility in incorporating different regulatory rules also makes it adaptable to changes in Clean Water Act interpretation, ensuring its continued relevance in a dynamic regulatory environment.

[0043] Referring to FIG. 2, a flowchart of a method 200 for determining regulatory status of water resources using a convolutional neural network is shown. The method 200 may be implemented by the data processing system 100 shown in FIG. 1.

[0044] At operation 202, the system obtains a plurality of input data layers for a geographic location. The input data layers may include aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data for the geographic location. In one embodiment, the aerial imagery data may comprise red, green, blue, and near-infrared bands from aerial photographs obtained from the National Agricultural Imagery Program (NAIP). In another embodiment, the wetland location data may be obtained from the National Wetland Inventory (NWI) and the stream location data may be obtained from the National Hydrography Dataset (NHD). The soil data may be extracted from the Gridded National Soil Survey Geographic Database, while climate data may be obtained from the PRISM Climate Group's 30-year Normals. Additional input data layers may include elevation data from USGS's 3DEP model, land cover data from the National Land Cover Database (NLCD), and ecoregion classification from the US EPA Level IV Ecoregions.

[0045] At operation 204, the system preprocesses the input data layers into a multi-channel image. In one embodiment, this may involve rasterizing all vector layers into 512x512 pixel tiles, where each pixel is 0.6 to 1.1 meters, to match the resolution of the NAIP imagery. For categorical data, such as NLCD land cover classes or ACE districts, ordinal encoding may be used. In another embodiment, the system may create a rule encoding layer with the same dimensions as other image layers, populating each pixel with a rule embedding vector that encodes the applicable Clean WaterAct rule for the geographic location.

[0046] At operation 206, the multi-channel image is fed into a trained convolutional neural network. In one embodiment, the convolutional neural network may be a ResNet-18 architecture, comprising an input layer, multiple convolutional layers, pooling layers, and fully connected layers. The ResNet-18 architecture may begin with a convolutional block, followed by four residual blocks, an average pooling operation, and finally a fully connected layer. The convolutional neural network may be trained on a dataset of AJDs issued by the Army Corps of Engineers, with each AJD indicating whether a corresponding location contains waters regulated under the Clean Water Act.

[0047] At operation 208, the system obtains an output score from the neural network indicating the probability of regulated waters. In one embodiment, the output of the fully connected layer may be passed through a softmax function, producing a score in the range [0, 1], This score represents the model's confidence that a given location contains waters that are jurisdictional (regulated) under the Clean Water Act. In another embodiment, the system may obtain separate output scores for different Clean Water Act rules, such as Rapanos, the Clean Water Rule (CWR), and the Navigable Waters Protection Rule (NWPR). Additionally, the system may determine output scores for a plurality of sub-regions within the geographical location by inputting multichannel input images for each sub-region into the trained convolutional neural network. This allows for a more granular analysis of the geographical location, enabling the identification of specific areas with a higher likelihood of containing regulated waters.

[0048] At operation 210, the system determines the regulatory status based on the output score. In one embodiment, the system may predict that a site is jurisdictional if the score exceeds 0.5. In another embodiment, the system may use different thresholds for different levels of confidence. For example, scores below 0.17 or above 0.83 may be considered to have 90% accuracy, while scores below 0.07 or above 0.95 may be considered to have 95% accuracy. The system may flag locations with scores in these high-confidence ranges for potential automated decision-making. In another embodiment, output scores below a pre-determined confidence level may be flagged for manual human review. Furthermore, the system may identify one or more sub-regions having the highest probability of containing regulated waters based on the output scores determined for the plurality of sub-regions.

[0049] At operation 212, the system determines the water resource hydrologic type. In oneembodiment, this may involve classifying the water resource into categories such as ephemeral streams, intermittent streams, perennial streams, emergent wetlands, forested wetlands, isolated wetlands, wetlands adjacent to regulated waters, and traditional navigable waters. The classification may be based on additional output nodes in the convolutional neural network, trained to predict these hydrologic types.

[0050] At operation 214, the system determines the water resource legal classification. In one embodiment, this may involve categorizing the water resource according to legal definitions established by the Clean Water Act and related court decisions. For example, the system may classify waters as "waters of the United States" (WOTUS), "adjacent wetlands," "tributaries," or "isolated waters." This classification may be based on the output score from operation 208, as well as additional contextual information from the input data layers. In another embodiment, the system may produce separate classification scores for each of nine legal resource types, such as ephemeral feature, isolated water, non-relatively permanent water, relatively permanent water, traditional navigable water, and wetland adjacent to a regulated water.

[0051] At operation 216, the system generates an alert or visualization of the regulatory status. In one embodiment, this may involve creating a map displaying the predicted regulatory status across the geographic area of interest. The map may use color coding to indicate the probability of regulation, with darker colors representing higher probabilities. The system may also highlight areas where the regulatory status is particularly uncertain, prompting users to conduct further investigation or manual review in these areas. In some implementations, the alert may take the form of a message sent to a device associated with a user, specifying that potentially regulated waters have been detected in a particular geographic location. The message may include coordinates corresponding to the geographic locations which have been identified as containing regulated waters, or may contain a link to an interactive map that displays the areas where regulated waters have been detected. Additionally, the system may generate a map overlay highlighting the identified sub-regions with the highest probability of containing regulated waters, and display this map overlay superimposed on the aerial imagery data on a user interface. This visualization allows users to visually identify areas within the geographical location that are most likely to contain waters regulated under the Clean Water Act, facilitating more targeted analysis and decisionmaking.

[0052] By following this method 200, the system can provide rapid, consistent, andgeographically extensive predictions of Clean Water Act jurisdiction, potentially saving significant time and resources compared to traditional case-by-case determinations while also supporting more comprehensive analysis of regulatory impacts across large areas.

[0053] Referring to FIG. 3, a convolutional neural network system 300 for regulatory status determination is shown. The system 300 comprises input data layers 302, a convolutional neural network 330, and predictions including regulatory status 340, hydrologic type 342, and legal classification 344. The system 300 is configured to analyze geospatial and regulatory data to predict Clean Water Act jurisdiction for water resources at a given geographic location.

[0054] The input data layers 302 include multiple types of geospatial and regulatory data that serve as inputs to the convolutional neural network 330. These layers include aerial imagery data 304, which may comprise red, green, blue, and near-infrared bands from aerial photographs, such as those from the National Agricultural Imagery Program (NAIP). In some embodiments, the aerial imagery data 304 may be satellite imagery, while in other embodiments it may be imagery captured by aircraft or unmanned aerial vehicles. The wetland location data 306 and stream location data 308 provide information on the locations and characteristics of wetlands and streams proximate to the geographic location being analyzed. In one embodiment, this data may be derived from the National Wetlands Inventory (NWI) and the National Hydrography Dataset (NHD), respectively. In another embodiment, the wetland and stream location data may be obtained from state or local environmental agencies.

[0055] The soil data 310 includes information on soil characteristics for the location, which may be extracted from the Gridded National Soil Survey Geographic Database (gNATSGO) in one embodiment. Alternatively, soil data may be obtained from on-site soil surveys or other soil databases. Elevation data 312 provides information on the topography of the area, which may be sourced from the USGS 3DEP model or other digital elevation models. Climate data 314 includes local climate information, such as precipitation and temperature variables, which may be obtained from the PRISM Climate Group's 30-year Normals or other climate databases. Land cover classification data 316 provides information on the types of land cover present in the area, which may be sourced from the National Land Cover Database (NLCD) or other land cover classification systems.

[0056] The rule encoding layer 318 encodes the specific Clean Water Act rule to be applied for determining regulatory status. This layer may comprise an embedding vector for each pixel ofthe multi-channel image, where the embedding vector encodes the text of the applicable rule. In one embodiment, the rule encoding may be generated using a pretrained language model to create an embedding of the rule text. In another embodiment, the rule encoding may be a one-hot encoded vector representing different rules (e.g., Rapemos, Clean Water Rule, Navigable Waters Protection Rule).

[0057] Regulatory boundary data 320 includes information on relevant regulatory jurisdictions, such as state boundaries, Army Corps of Engineers (ACE) district boundaries, and the distance to the nearest ACE district headquarters. This data helps the model account for variations in Clean Water Act implementation across different jurisdictions and regulatory regimes.

[0058] The convolutional neural network 330 receives the input data layers and generates the output predictions. The network architecture includes convolutional layers 332, which apply learned filters to the input data to extract relevant features. In one embodiment, the convolutional neural network 330 may comprise a ResNet-18 architecture, while in another embodiment, it may use another convolutional neural network architecture, such as a U-net architecture. In other embodiments, different types of machine learning models may be used, such as decision tree-based models (e.g., random forests, gradient boosting machines), support vector machines, or other suitable machine learning models. Further, a hybrid machine learning model may be used, combining multiple machine learning methods. Additionally, the machine learning model 330 may utilize transformer architectures or foundation models adapted for image segmentation tasks. The specific choice of machine learning model may depend on factors such as the characteristics of the input data, available computational resources, and desired trade-offs between model performance and interpretability.

[0059] The pooling layers 334 downsample the feature maps produced by the convolutional layers 332, reducing the spatial dimensions of the feature maps while retaining salient information. In some embodiments, max pooling may be used, while in others, average pooling or adaptive pooling techniques may be employed. The fully connected layers 336 take the flattened output from the previous layers and perform high-level reasoning to generate the output predictions. These layers may vary in number and size depending on the complexity of the classification task and the amount of training data available.

[0060] The softmax layers 338 are the final layers of the network, which normalize the output predictions of the fully connected layers 336 into probability distributions for each classificationtask. In one embodiment, there may be separate softmax layers for regulatory status, hydrologic type, and legal classification. In another embodiment, a single softmax layer may be used with multiple output nodes corresponding to the different classification tasks.

[0061] The output of the convolutional neural network 330 includes the regulatory status 340, which indicates the probability that the analyzed location contains waters regulated under the Clean Water Act. This output may be a binary classification (regulated or not regulated) or a continuous probability score. The hydrologic type 342 provides a classification of the water resource into categories such as ephemeral streams, intermittent streams, perennial streams, emergent wetlands, forested wetlands, or isolated wetlands. This classification is based on the hydrological characteristics of the water resource as inferred from the input data layers 302.

[0062] The legal classification 344 provides a categorization of the water resource according to legal definitions under the Clean Water Act. This may include classifications such as traditional navigable waters, relatively permanent waters, non-relatively permanent waters, wetlands adjacent to regulated waters, isolated waters, etc.

[0063] In operation, the input data layers 302 are preprocessed into a multi-channel image, where each pixel corresponds to a specific geographic location and comprises values from the various input data layers. This multi-channel image is then fed into the convolutional neural network 330, which processes the data through its layers to generate the output predictions. The system can be trained on a dataset of approved jurisdictional determinations (AJDs) issued by the Army Corps of Engineers, allowing it to learn the complex relationships between geospatial features, regulatory context, and jurisdictional outcomes. The convolutional neural network system 300 may be used in various applications, such as supporting permitting decisions, informing policy design, or providing rapid assessments of potential Clean Water Act jurisdiction over large geographic areas. By integrating diverse geospatial data with regulatory information, the system offers a powerful tool for streamlining the process of determining Clean Water Act jurisdiction while providing a quantitative basis for decision-making in water resource management and protection.

[0064] Referring to FIG. 4, a flowchart of a method 400 for generating a training dataset for water resource classification is shown. The method 400 may be executed by the data processing system 100 and comprises steps for obtaining and processing AJDs to create a comprehensive training dataset for use in machine learning models that predict CWA jurisdiction for waterresources.

[0065] At operation 402, the data processing system obtains AJDs for water resources at multiple geographic locations. This operation involves retrieving AJD records from the U.S. ACE database, which contains legally-binding case-by-case decisions made by ACE engineers regarding the jurisdictional status of water resources under the CWA. The AJDs may be obtained for a specific time period, such as between August 2015 and May 2022, to ensure relevance to current regulatory frameworks. In some embodiments, the operation may include filtering the AJDs to focus on specific geographic regions or regulatory districts to create a more targeted training dataset.

[0066] At operation 404, the data processing system extracts information from each AJD, including geographic coordinates, ground truth regulatory status, ground truth Cowardin wetland code, ground truth legal classification, and the applicable Clean Water Act rule. This extraction process comprise using natural language processing techniques to parse the AJD documents and extract the relevant fields. The geographic coordinates identify the precise location of the water resource, while the ground truth regulatory status indicates whether the resource is regulated under the CWA. The Cowardin wetland code provides a standardized classification of the water resource type, such as palustrine, riverine, or lacustrine. The ground truth legal classification may include categories such as adjacent wetlands, isolated wetlands, or relatively permanent waters. The applicable CWA rule (e.g., Rapanos, Clean Water Rule, or Navigable Waters Protection Rule) is also extracted and recorded to account for variations in regulatory interpretation over time.

[0067] At operation 406, the data processing system generates multi-channel input images centered on each AJD location. This operation involves compiling various geospatial data layers into a single multi-channel image for each AJD site. This process may include extracting data from multiple sources, such as red, green, blue, and near-infrared bands from National Agricultural Imagery Program (NAIP) aerial photographs; soil, groundwater, and elevation data from the Gridded National Soil Survey Geographic Database; hydrological information from the National Hydrography Dataset (NHD); wetland coverage from the National Wetland Inventory (NWI); land cover data from the National Land Cover Database (NLCD); climate data from PRISM Climate Group's 30-year Normals; elevation data from USGS's 3DEP model; and ecoregion classification from US EPA Level IV Ecoregions. In some embodiments, the multi-channel image generation may also incorporate regulatory boundary data, such as state borders, ACE district boundaries, andencoded information about the applicable CWA rule. This additional information allows the model to account for variations in CWA implementation across different jurisdictions and regulatory regimes.

[0068] At operation 408, the data processing system associates each multi-channel input image with its corresponding ground truth labels extracted in operation 404. This step creates a direct link between the geospatial data representation of each site and its known regulatory status, wetland classification, and legal classification. This association may be stored as a structured data object, with the multi-channel image data linked to a metadata object containing the ground truth labels. Alternatively, the association could be maintained in a relational database, with separate tables for image data and label information, linked by a unique identifier for each AJD site.

[0069] Finally, at operation 410, the data processing system stores the training dataset of multichannel images and ground truth labels. This may involve saving the dataset to a non-transitory computer-readable medium, such as a hard drive or solid-state storage device, in a format optimized for efficient access during model training. The storage format may be chosen to balance factors such as data compression, read / write speed, and compatibility with machine learning frameworks. In some embodiments, the training dataset may be stored in a distributed file system or cloud storage solution to facilitate parallel processing and scalability for large-scale model training operations.

[0070] The resulting training dataset created by method 400 may subsequently be used for training machine learning models, particularly convolutional neural networks, to predict CWA jurisdiction for water resources. By incorporating diverse geospatial data layers and accurate ground truth labels from AJDs, the dataset enables the development of models that can learn relationships between physical landscape features, regulatory contexts, and jurisdictional outcomes.

[0071] Referring to FIG. 5, a flowchart of a method 500 for compiling multi-channel image data for geographic coordinates of interest is shown. The method 500 may be executed by a data processing system as part of method 200 for determining the regulatory status of one or more water resources located at the geographic coordinates of interest (e.g., at operation 202) or as part of a method for generating a training dataset (e.g., at operation 406 of method 400) for training a machine learning model to predict regulatory status and other information for one or mor water resources located at the geographic coordinates.

[0072] At operation 502, the data processing system extracts red, green, blue, and near-infrared bands from National Agricultural Imagery Program (NAIP) aerial photographs centered on the geographic coordinates. In one embodiment, this involves retrieving aerial imagery data covering an area surrounding the coordinates of interest. The NAIP imagery provides visual information about the land cover, vegetation, and potential water features at the site. In another embodiment, the data processing system may apply pre-processing techniques to the NAIP imagery, such as atmospheric correction or image enhancement. The extracted NAIP bands serve as a component for the multi-channel image, providing spectral information for identifying and characterizing potential water resources.

[0073] At operation 504, the data processing system extracts soil, groundwater, and elevation data from the Gridded National Soil Survey Geographic Database (gNATSGO) centered on the geographic coordinates. In one embodiment, this data includes soil taxonomic class, soil hydric rating, flooding frequency, ponding frequency, and water table depth variables. The gNATSGO data may be resampled to match the spatial resolution of the NAIP imagery. In another embodiment, the soil property data may be obtained from a survey of the field conducted by the user or a private contractor, if available.

[0074] At operation 506, the data processing system extracts hydrological information from the National Hydrography Dataset (NHD) centered on the geographic coordinates. In one embodiment, this includes information on the locations and characteristics of streams and other water bodies within a certain radius (e g., 10 meters) of the coordinates. The NHD data may distinguish between perennial, intermittent, and ephemeral streams. In another embodiment, the data processing system may use the high-resolution NHD files where available, while noting potential incompleteness or measurement errors in the high-resolution data.

[0075] At operation 508, the data processing system extracts wetland coverage from the National Wetland Inventory (NWI) centered on the geographic coordinates. In one embodiment, this includes information on the locations and types of wetlands within a certain radius (e.g., 10 meters) of the coordinates. The NWI data may be used to identify potential wetland areas for training the convolutional neural network to output classification scores for various wetland categories.

[0076] At operation 510, the data processing system extracts land cover data from the National Land Cover Database (NLCD) centered on the geographic coordinates. In one embodiment, thisincludes information on the type of land cover (e g., developed, forest, cropland, wetland) for each pixel in a pre-determined region surrounding and including the geographic coordinates of interest. The NLCD data may be used to identify potential agricultural areas or other land uses relevant to Clean Water Act jurisdiction, as well as to inform the classification of stream categories.

[0077] At operation 512, the data processing system extracts climate data from the PRISM Climate Group's 30-year normals centered on the geographic coordinates. In one embodiment, this includes average annual total precipitation, daily minimum and maximum temperatures, mean temperature, mean dew point temperature, minimum and maximum vapor pressure deficit, solar radiation, and cloudiness data. These climate variables may be used to assess the likelihood of water accumulation and persistence in the area of interest. In another embodiment, the data processing system may also incorporate more recent climate data to account for potential changes in precipitation patterns.

[0078] At operation 514, the data processing system extracts elevation data from the USGS 3D Elevation Program (3DEP) model centered on the geographic coordinates. In one embodiment, this provides high-resolution elevation data that can be used to assess topography and potential water flow patterns. The 3DEP data may be resampled to match the spatial resolution of the NAIP imagery. In another embodiment, the data processing system may use the elevation data to calculate additional topographic indices, such as slope and flow accumulation, which can be relevant for assessing potential jurisdictional waters.

[0079] At operation 516, the data processing system extracts ecoregion classification from the US EPA Level IV Ecoregions dataset centered on the geographic coordinates. In one embodiment, this provides information on the ecological context of the area, which can be relevant for assessing the potential for certain types of water resources. The ecoregion data may be used to inform the interpretation of other environmental variables. In another embodiment, the data processing system may use the ecoregion classification to adjust thresholds or parameters in the jurisdictional determination model based on regional characteristics.

[0080] At operation 518, the data processing system determines the distance to the nearest U.S. ACE district headquarters from the geographic coordinates. In one embodiment, this distance is calculated as a straight-line distance between the coordinates and the nearest headquarters location. The distance to ACE headquarters may be relevant because the probability of in-person visits decreases with this distance, and sites with in-person visits have a higher probability of beingregulated. In another embodiment, the data processing system may calculate travel time or road distance instead of straight-line distance to better represent the accessibility of the site to ACE personnel.

[0081] At operation 520, the data processing system encodes the state, ACE regulatory district, and applicable Clean Water Act rule for the geographic coordinates. In one embodiment, operation 520 includes assigning numeric values to categorical data such that similar classes or adjacent geographic areas are close together in numeric space. For example, sequential numbers may be assigned to neighboring states rather than using federal information processing standards (FIPS) codes. In another embodiment, the data processing system may use one-hot encoding for the applicable Clean Water Act rule, creating separate input data layers for each rule (e.g., Rapanos, Clean Water Rule, Navigable Waters Protection Rule). In another embodiment, a definition of “waters of the United States” may be extracted from a Clean Water Act rule or supreme court decision, and encoded as an embedding vector.

[0082] At operation 522, the data processing system compiles the extracted and encoded data into a multi-channel image with a pre-determined pixel resolution. In one embodiment, this involves creating a 512x512 pixel image where each pixel corresponds to a specific geographic location proximal to and centered on the geographic coordinates of interest, and contains values from all the input data layers. For input data layers that are not natively in a raster or vector format (e g., distance to ACE headquarters, state and district encodings), the data processing system fills a 512x512 pixel raster with one value to make these layers conform to the convolutional neural network architecture. In another embodiment, the data processing system may use a different image size or resolution depending on the specific requirements of the neural network model or the desired level of detail for the analysis.

[0083] The resulting multi-channel image serves as a machine readable representation of the environmental, regulatory, and geographic context of the geographic coordinates of interest. This enables the convolutional neural network to learn complex relationships between physical landscape features, regulatory boundaries, and jurisdictional outcomes under various Clean Water Act rules.

[0084] Referring to FIG. 6, a flowchart of a method 600 for training a convolutional neural network to predict regulatory status, hydrologic type, and legal classification of water bodies from multi-channel images is shown. The method 600 is executed by a data processing system, such asthe system described in FIG. 1.

[0085] At operation 602, the data processing system generates a training dataset comprising multi-channel images and associated ground truth labels for regulatory status, hydrologic type, and legal classification. In one embodiment, the training dataset is generated by obtaining a plurality of AJDs for water resources located at various geographic locations. Each AJD includes geographic coordinates identifying the location of a water resource, a ground truth regulatory status indicating whether the water resource is regulated under the Clean Water Act, a ground truth Cowardin wetland classification code, a ground truth legal resource type, and the rule under which the AJD was decided.

[0086] The data processing system then pre-processes the training dataset by generating a plurality of multi-channel input images, each centered on a respective geographic location of an AJD. Each multi-channel input image comprises multiple input data layers, including aerial imagery data, vector data representing locations and characteristics of wetlands and streams, soil characteristic data, climate data, elevation data, land cover classification data, ecoregion classification data, data identifying distances to regulatory agency headquarters, and data encoding the applicable state, regulatory district, and Clean Water Act rule.

[0087] At operation 604, the data processing system initializes the parameters of the convolutional neural network. In one embodiment, this involves setting initial values for the weights and biases of the convolutional layers, pooling layers, and fully connected layers of the network. The initial values may be randomly generated or set based on pre-trained models. In another embodiment, the initialization may involve setting up the network architecture, including defining the number and arrangement of convolutional layers, the kernel sizes for each layer, and the number of neurons in the fully connected layers.

[0088] At operation 606, the data processing system inputs the multi-channel images from the training dataset into the convolutional neural network. In one embodiment, this involves feeding each multi-channel image through the series of convolutional layers, where each layer applies a set of learnable filters to the input. The output of each convolutional layer is then passed through an activation function, such as a rectified linear unit (ReLU), to introduce non-linearity. In another embodiment, the input operation may also include data augmentation techniques, such as random rotations or flips of the input images, to increase the diversity of the training data and improve the model's generalization capabilities.

[0089] At operation 608, the data processing system obtains predicted regulatory status, hydrologic type, and legal classification predictions from the convolutional neural network for each of the input multi-channel images. In one embodiment, this involves passing the output of the final convolutional layer through one or more fully connected layers, followed by a softmax activation function to produce probability distributions for each of the three output predictions. The regulatory status output may be a binary classification (regulated or not regulated), while the hydrologic type and legal classification outputs may be multi-class predictions corresponding to Cowardin codes and legal resource types, respectively.

[0090] At operation 610, the data processing system determines a loss by comparing the predicted regulatory status, hydrologic type, and legal type to the ground truth labels for each of the AJDs in the training dataset. In one embodiment, separate models are trained for each prediction task (regulatory status, hydrologic type, and legal type), with each model using a single type of loss function specific to its prediction category. For the regulatory status model, a binary cross-entropy loss may be calculated by comparing the predicted regulatory status to the ground truth regulatory status for each of the plurality of AJDs. For the hydrologic type model, a multiclass cross-entropy loss may be determined by comparing the predicted Cowardin wetland classification code to the ground truth Cowardin wetland classification code for each of the plurality of AJDs. For the legal resource type model, another multi-class cross-entropy loss may be computed by comparing the predicted legal resource type to the ground truth legal resource type for each of the plurality of AJDs. Each model is updated independently based on its respective loss, allowing for specialized optimization of each prediction task. In another embodiment, a multitasking machine learning model is trained by calculating a combined loss function comprising multiple components: a first loss determined by comparing the predicted regulatory status to the ground truth regulatory status for each of the plurality of AJDs, a second loss determined by comparing the predicted Cowardin wetland classification code (corresponding to the hydrologic type) to the ground truth Cowardin wetland classification code for each of the plurality of AJDs, and a third loss determined by comparing the predicted legal resource type to the ground truth legal resource type for each of the plurality of AJDs. The combined loss function comprises the first loss, the second loss, and the third loss. In another embodiment, the loss function may also include regularization terms to prevent overfitting, such as LI or L2 regularization on the network weights.

[0091] At operation 612, the data processing system updates the parameters of the convolutional neural network based on the determined loss to minimize future losses on the training dataset. In one embodiment, this involves using a gradient descent optimization algorithm, such as stochastic gradient descent (SGD) or Adam, to adjust the weights and biases of the network in the direction that minimizes the loss. The learning rate and other hyperparameters of the optimization algorithm may be tuned to balance the speed of convergence with the stability of the training process. In another embodiment, techniques such as batch normalization or dropout may be applied during the update step to improve the network's generalization and prevent overfitting.

[0092] At operation 614, the data processing system determines whether convergence criteria have been satisfied. In one embodiment, this involves checking if the loss has decreased below a predetermined threshold, or if the change in loss between consecutive epochs has decreased below a second predetermined threshold. In another embodiment, the convergence check may also include evaluating the model's performance on a separate validation dataset to ensure that the model is not overfitting to the training data. If the convergence criteria are not satisfied, the process returns to operation 606 to continue training with another batch of input images.

[0093] If the convergence criteria are satisfied at operation 614, the method proceeds to operation 616, where the data processing system stores the trained convolutional neural network model. In one embodiment, this involves saving the final weights and biases of the network, along with the network architecture and any necessary preprocessing steps, to a persistent storage medium. The stored model can then be used for making predictions on new, unseen multi-channel images of water resources. In another embodiment, the storage operation may also include generating metadata about the training process, such as the final loss values, training duration, and performance metrics on a test dataset, to facilitate future model comparisons and iterations.

[0094] The trained convolutional neural network model resulting from method 600 can be used to efficiently and accurately predict the regulatory status, hydrologic type, and legal type (also referred to herein as a legal classification) of water resources across diverse geographic regions. This model may provide a valuable tool for regulatory agencies, environmental planners, and policymakers to assess the scope of Clean Water Act jurisdiction and streamline the process of making j uri sdi cti onal determi nati ons .

[0095] Referring to FIG. 7, a flowchart of a method 700 for updating and encoding Clean Water Act rule definitions into a machine readable format is shown. This method is executed by adata processing system and enables flexible adaptation of predictions by a trained convolutional neural network to accommodate changes in interpretation of what constitutes "waters of the United States" without needing to retrain the convolutional neural network.

[0096] At operation 702, the data processing system receives text of a new Clean Water Act rule or Supreme Court decision. This may involve the system obtaining the full text of a newly published rule from the Federal Register or the text of a Supreme Court opinion from the Court's website. In one embodiment, the text is automatically retrieved by the data processing system from official government sources using web scraping techniques. In another embodiment, the text is manually input by a user through a graphical user interface provided by the system.

[0097] At operation 704, the data processing system updates the definition of waters covered by the Clean Water Act based on the new rule or court decision. This may involve the system parsing the received text to extract key phrases or concepts that modify the existing definition of "waters of the United States" (WOTUS). For example, if a new rule excludes certain types of wetlands from CWA jurisdiction, the system would update the WOTUS definition to reflect this change. In one embodiment, the data processing system uses natural language processing techniques to automatically identify and incorporate definitional changes. In another embodiment, legal experts may manually review the text and update the WOTUS definition in a structured format within the system.

[0098] At operation 706, the data processing system generates an embedding vector for the updated definition using a pretrained language model. This step involves the system converting the textual WOTUS definition into a dense vector representation that captures semantic meaning. In one embodiment, the system uses a transformer-based language model like BERT to generate fixed-length embedding vectors. In another embodiment, the system employs a recurrent neural network to create variable-length embeddings that capture more nuanced aspects of the definition.

[0099] At operation 708, the data processing system receives a multi-channel image with geospatial data layers for a geographic location. These layers may include aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data. In one embodiment, the system generates the multi-channel image on-demand for a specific location. In another embodiment, the system retrieves pre-generated multi-channel images from a database.

[0100] At operation 710, the data processing system creates a rule encoding layer with thesame dimensions as the other image layers. This new layer will be used to incorporate the updated WOTUS definition into the input data for the convolutional neural network. In one embodiment, the system generates a rule encoding layer with the same spatial dimensions (height and width) as the other layers but may have a different number of channels. In another embodiment, the system creates a rule encoding layer that matches all dimensions of the other layers exactly.

[0101] At operation 712, the data processing system populates each pixel of the rule encoding layer with the rule embedding vector generated in operation 706. This step ensures that the updated WOTUS definition is associated with every spatial location in the input data. In one embodiment, the system replicates the same embedding vector across all pixels. In another embodiment, the system slightly perturbs the embedding vector for each pixel to introduce controlled noise, which may improve the model's robustness.

[0102] At operation 714, the data processing system concatenates the rule encoding layer with the other input data layers to create the final multi-channel input image. This combined input allows the convolutional neural network to consider both geospatial features and the current regulatory definition when making predictions about CWA jurisdiction. In one embodiment, the system appends the rule encoding layer as additional channels to the existing multi-channel image. In another embodiment, the system interleaves the rule encoding layer with the existing layers to promote better integration of the regulatory information.

[0103] the method 700 provides a flexible and efficient approach to incorporating evolving Clean Water Act regulations into a machine learning-based system for predicting CWA jurisdiction. By encoding regulatory definitions as part of the input data, the system can adapt to changes without extensive retraining, enabling more responsive and accurate predictions of which waters are protected under the Clean Water Act. This approach allows for rapid updates to the prediction system as regulatory interpretations evolve. Furthermore, the use of embedding vectors to represent the WOTUS definition enables the system to capture complex semantic relationships within the regulatory text. This nuanced representation allows the convolutional neural network to make more informed decisions about CWA jurisdiction based on the specific language and intent of the current rule. Additionally, the method 700 supports the analysis of hypothetical scenarios. By generating embedding vectors for proposed rule changes, the data processing system allows policymakers to predict the potential impact of new regulations on CWA jurisdiction across different geographic areas. This capability can inform decision-making and help assess theconsequences of various policy options.

[0104] Referring to FIG. 8, a flowchart of a method 800 for determining feature impact on regulatory status predictions by a trained convolutional neural network is shown. Method 800 is executed by a data processing system to evaluate the importance of different input data layers in predicting Clean Water Act jurisdiction.

[0105] At operation 802, the data processing system accesses a trained convolutional neural network. In one embodiment, this involves loading a previously trained convolutional neural network. The trained model may have been stored after completing the training process outlined in FIG. 6. In another embodiment, accessing the trained network may include retrieving model weights and architecture information from a model repository or database.

[0106] At operation 804, the data processing system selects an input data layer for analysis. The input data layers may include aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data. In one embodiment, the system may select a single input layer, such as the climate data from the PRISM Climate Group's 30-year Normals. In another embodiment, the system may select a group of related layers, such as all soil characteristic layers from the Gridded National Soil Survey Geographic Database.

[0107] At operation 806, the data processing system randomly permutes values of the selected input data layer across samples. This permutation breaks the link between the selected feature and its label, allowing the data processing system to assess the feature's importance. In one embodiment, the permutation may be performed across all samples nationally. In another embodiment, the permutation may be performed within-state, shuffling layers solely within state samples to indicate which features help the model replicate AJDs within a state.

[0108] At operation 808, the data processing system feeds the permuted data into the trained convolutional neural network. This involves inputting the multi-channel images with the permuted layer(s) an input layer of the trained convolutional neural network. In one embodiment, the system may process a subset of the validation dataset to reduce computational time. In another embodiment, the entire validation dataset may be processed to ensure a comprehensive analysis.

[0109] At operation 810, the data processing system obtains regulatory status predictions from the trained neural network. These predictions represent the model's output when the selected input layer has been randomized. In one embodiment, the system may collect binary predictions ofwhether each sample is regulated under the Clean Water Act. In another embodiment, the system may also collect probability scores for each sample, representing the model's confidence in its predictions.

[0110] At operation 812, the data processing system calculates the accuracy of predictions with the permuted data. This involves comparing the model's predictions to the ground truth labels from the validation dataset. In one embodiment, the system may calculate overall accuracy as the percentage of correctly classified samples. In another embodiment, the system may compute additional metrics such as precision, recall, or Fl score to provide a more comprehensive evaluation of the model's performance with the permuted data.[0U1] At operation 814, the data processing system determines the permutation importance as the difference between the original accuracy and the accuracy with permuted data. This quantifies how much the model's performance degrades when the selected input data layer or layers is / are randomized, indicating the importance of the input data layer(s) to the model's predictions. In one embodiment, the system may calculate a single importance score for the selected layer or group of layers. In another embodiment, the system may compute separate importance scores for different geographic regions or regulatory contexts to identify variations in feature importance across different scenarios.

[0112] Through the above described permutation importance analysis, method 800 enables a deeper understanding of how the convolutional neural networks taught herein make predictions, and which input features are most relevant for determination of Clean Water Act jurisdiction. This information may be valuable for refining the model, prioritizing data collection efforts, and providing interpretability for regulatory decision-making processes. Further, by quantifying input data layer correlation with accurate regulatory status predictions, less impactful or correlated features may be discarded, enabling an improvement in the computational efficiency of regulatory status prediction with little or no impact on prediction accuracy. For example, the inventors have determined that climate data, stream and wetland vector data, and regulatory boundary information had the largest influence on the accuracy of regulatory status predictions. Additionally, understanding the relative importance of different input data layers may guide the development of more targeted data collection strategies, focusing resources on obtaining high-quality data for the most influential features.

[0113] Referring to FIG. 9, two charts comparing feature importance at national and statelevels for various environmental and geographical factors are shown. The charts illustrate the relative importance of different input features used by the trained convolutional neural network (taught herein) to predict Clean Water Act jurisdiction for water resources.

[0114] Chart 902 depicts the national feature importance, using permutation importance as the measure of feature importance. At the national level, the PRISM (Parameter-elevation Regressions on Independent Slopes Model) climate data shows the highest importance, indicating that local climate patterns, such as precipitation and temperature, may have a strong influence on regulatory status across the country. This suggests that climate variables may play a role in determining the presence and characteristics of water resources that may fall under Clean Water Act jurisdiction. The second most important feature at the national level is the State, highlighting the potential significance of state-specific factors in determining jurisdiction. This may reflect variations in state-level implementation of Clean Water Act regulations or differences in state-specific environmental conditions. The third most important feature is the combined National Wetlands Inventory (NWI) and National Hydrography Dataset (NHD), which provide information on the locations and types of wetlands and water bodies. The high importance of this feature may indicate the role of existing mapped water resources in predicting Clean Water Act jurisdiction at a national scale.

[0115] Chart 904 presents the state-level feature importance, also using permutation importance. At the state level, the most important feature is the combined NWI and NHD data. This indicates that within individual states, the presence and characteristics of mapped wetlands and water bodies may be strong predictors of Clean Water Act jurisdiction. The second most important feature at the state level is the gNATSGO (Gridded National Soil Survey Geographic Database) data, which includes soil characteristics such as taxonomic class, hydric rating, and water table depth. This suggests that local soil conditions may play a role in determining Clean Water Act jurisdiction within individual states. The third most important feature at the state level is the Rule, which encodes the applicable Clean Water Act rule (e.g., Rapanos, Clean Water Rule, or Navigable Waters Protection Rule). The high importance of the Rule at the state level indicates that the specific regulatory framework being applied may have an impact on jurisdictional determinations within states.

[0116] The differences between national and state-level feature importance highlight the interplay of factors potentially influencing Clean Water Act jurisdiction across different scales.While climate data and state-specific factors are most important at the national level, the presence of mapped water resources and local soil conditions become more important when considering variations within individual states. This may indicate the need for a flexible approach to predicting regulatory status, as implemented in the trained convolutional neural network described herein.

[0117] Referring to FIG. 10, a graphical depiction of estimated probabilities of CWA regulation across the United States under different regulatory frameworks is shown. The figure includes five visualizations (1002, 1004, 1006, 1008, 1010) representing regulatory probabilities predicted by the convolutional neural networks taught herein and changes under various CWA rules.

[0118] Visualization 1002 shows the estimated probability of CWA regulation under the Rapanos framework. Visualization 1002 depicts the output of the trained convolutional neural network for four million prediction points across the USA, with darker colors indicating higher probabilities of regulation. The map reveals geophysical and political patterns of jurisdiction, with higher regulation probabilities in coastal regions, wetlands in the coastal South and mid-Atlantic, and streams near the Pacific. Visualization 1002 highlights areas of lower regulation probability in the arid West, though some ephemeral streams are still predicted to be regulated. The Fall Line separating the Coastal Plain in the mid-Atlantic and South is visible as a discrete change in regulation probability.

[0119] Visualization 1004 presents the estimated regulatory probability under the Navigable Waters Protection Rule (NWPR). Visualization 1004 shows a significant reduction in overall regulation probability compared to Rapanos, with most predicted probabilities below 20%. Major waterways remain regulated, though with narrower channels, potentially due to decreased jurisdiction of nearby wetlands. In another embodiment, visualization 1004 emphasizes the least jurisdiction in the arid West, where ephemeral streams are common and largely excluded from regulation under NWPR.

[0120] Visualization 1006 illustrates the estimated regulation changes from Rapanos to NWPR. Visualization 1006 uses color coding to show areas of deregulation (brown) and new regulation (green). The visualization reveals that regulation decreases most around isolated wetlands in the mid-Atlantic and Gulf coast. Visualization 1006 demonstrates that NWPR shows no substantial areas of increased regulation compared to Rapanos, with limited areas of decreased regulation in the arid West.

[0121] Visualization 1008 depicts the estimated regulatory probability under the Clean Water Rule (CWR). Visualization 1008 shows broadly similar jurisdiction patterns to Rapanos, with some notable differences. Visualization 1008 indicates less coverage of ephemeral streams under CWR in the arid West compared to Rapanos. Visualization 1008 highlights the persistence of higher coverage areas across the country under CWR, similar to Rapanos but distinct from NWPR.

[0122] Visualization 1010 shows the estimated regulation changes from CWR to NWPR. Visualization 1010 uses color coding to illustrate areas of deregulation (brown) and new regulation (green) when transitioning from CWR to NWPR. The visualization emphasizes the widespread deregulation under NWPR compared to CWR across the country. Visualization 1010 reveals patterns within the 38 Army Corps of Engineers (ACE) districts, demonstrating some persistence in regulatory patterns across different rules.

[0123] The visualizations of FIG. 10 collectively provide a comprehensive view of how CWA jurisdiction varies across different regulatory frameworks and geographic regions. The maps are created using a 247x576 grid, displaying the mean model score in each bin, with approximately 28 prediction points per bin. This approach allows for a detailed yet manageable representation of the vast amount of data generated by the trained convolutional neural network in predicting regulatory status of water resources across the United States.

[0124] The systems and methods described herein provide several technical improvements and effects in the field of automated regulatory status determination for water resources. By leveraging convolutional neural networks trained on diverse geospatial and regulatory data, the invention enables more efficient, accurate, and consistent predictions of Clean Water Act jurisdiction compared to traditional case-by-case manual assessments. The multi-channel input image approach, which combines various geospatial layers with encoded regulatory information, allows the model to learn complex relationships between physical landscape features and jurisdictional outcomes specific to particular regulatory contexts. This integration of environmental and regulatory data enables the system to account for variations in Clean Water Act implementation across different jurisdictions and regulatory regimes, providing more nuanced and locally-relevant predictions. The scalability of the approach allows for rapid evaluation of Clean Water Act jurisdiction over large geographic areas, supporting more comprehensive policy analysis and environmental planning. Furthermore, the system's ability to output probability scores rather than binary classifications provides valuable uncertainty information that can be used to prioritizemanual review of ambiguous cases or inform risk assessments in development projects. The method's flexibility in incorporating different regulatory rules through appropriate training data and rule encoding layers also enables adaptation to evolving legal interpretations of Clean Water Act jurisdiction without requiring extensive retraining of the entire model. These technical improvements result in a more responsive, accurate, and efficient system for determining Clean Water Act jurisdiction, potentially reducing the time and resources required for regulatory compliance while supporting more informed decision-making in water resource management and protection.

[0125] The disclosure also provides support for a method for determining regulatory status of water resources under the Clean Water act, the method comprising: receiving a plurality of input data layers for a geographic location, the plurality of input data layers comprising aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data for the geographic location, preprocessing the plurality of input data layers into a multi-channel image, wherein each pixel of the multi-channel image corresponds to the geographic location and comprises values from the plurality of input data layers, feeding the multi-channel image into a trained convolutional neural network comprising a plurality of convolutional layers, obtaining an output score from the trained convolutional neural network, the output score indicating a probability that the geographic location contains waters regulated under the Clean Water act, and determining a regulatory status for the geographic location based on the output score, wherein the geographic location is determined to contain regulated waters when the output score exceeds a threshold value. In a first example of the method, the plurality of input data layers further comprises a rule encoding layer encoding a rule to be applied for determining regulatory status under the Clean Water Act, the rule encoding layer comprising a rule encoding vector for each pixel of the multi-channel image. In a second example of the method, optionally including the first example, the rule encoding layer comprises an embedding vector for each pixel of the multi-channel image, wherein the embedding vector encodes text of a rule to be applied for determining regulatory status under the Clean Water Act for the geographic location corresponding to the pixel. In a third example of the method, optionally including one or both of the first and second examples, the trained convolutional neural network is further configured to output a classification score for each of a plurality of hydrologic types, the classification score indicating a probability that the geographic location corresponds to anassociated hydrologic type. In a fourth example of the method, optionally including one or more or each of the first through third examples, the plurality of hydrologic types comprises a Cowardin wetland classification code. In a fifth example of the method, optionally including one or more or each of the first through fourth examples, feeding the multi-channel image into the trained convolutional neural network comprises: inputting the multi-channel image into an input layer of the trained convolutional neural network, processing the multi-channel image through a plurality of convolutional layers of the trained convolutional neural network, performing an average pooling operation on an output of the plurality of convolutional layers, and passing an output of the average pooling operation through a fully connected layer to produce the output score. In a sixth example of the method, optionally including one or more or each of the first through fifth examples, the trained convolutional neural network is further configured to output a classification score for each of a plurality of legal resource types defined by the Clean Water Act, the classification score indicating a probability that the geographic location corresponds to an associated legal resource type. In a seventh example of the method, optionally including one or more or each of the first through sixth examples, the plurality of legal resource types comprise one or more of an ephemeral feature, an isolated water, a non-relatively permanent water, a relatively permanent water, a traditional navigable water, and a wetland adjacent to a regulated water.

[0126] The disclosure also provides support for a data processing system for classifying a geographical location as containing waters regulated under the Clean Water act, comprising: a non-transitory memory storing instructions, a trained machine learning model, and a plurality of input data layers comprising geospatial data for the geographical location, wherein the plurality of input data layers include aerial imagery data, wetlands data, streams data, soil data, groundwater data, elevation data, climate data, land cover data, and geographic data identifying a state, an army Corps of Engineers (ACE) district, and a Clean Water act rule applicable to the geographical location, a processor coupled to the non-transitory memory and configured to execute the instructions to: generate a multi-channel input image for the geographical location by combining the plurality of input data layers, input the multi-channel input image into the trained machine learning model, wherein the trained machine learning model is trained on a dataset of approved jurisdictional determinations issued by the army Corps of Engineers, each approved jurisdictional determination indicating whether a corresponding location contains waters regulated under the Clean Water act, receive, from the trained machine learning model, an output score indicating aprobability that the geographical location contains waters regulated under the Clean Water act for the Clean Water act rule, determine, based on the output score, whether the geographical location contains waters regulated under the Clean Water act for the Clean Water act rule, and generate an alert indicating the determination of whether the geographical location contains waters regulated under the Clean Water act. In a first example of the system, the alert comprises sending a message to a user device, the message specifying one or more locations corresponding to the geographical location that have been classified as containing waters regulated under the Clean Water Act. In a second example of the system, optionally including the first example, the trained machine learning model is further configured to classify the geographical location into one or more categories of water resource types based on the multi-channel input image, wherein the categories of water resource types include ephemeral streams, intermittent streams, perennial streams, emergent wetlands, forested wetlands, isolated wetlands, wetlands adjacent to regulated waters, and traditional navigable waters. In a third example of the system, optionally including one or both of the first and second examples, the processor is further configured to generate a visualization displaying the geographical location and an indication of a type of water resource classified by the trained machine learning model. In a fourth example of the system, optionally including one or more or each of the first through third examples, the plurality of input data layers further comprise vector data identifying locations of wetlands and streams proximal to the geographical location. In a fifth example of the system, optionally including one or more or each of the first through fourth examples, the system further comprises: a user interface configured to receive a user input specifying the geographical location, wherein the processor generates the multi-channel input image for the geographical location specified by the user input. In a sixth example of the system, optionally including one or more or each of the first through fifth examples, the processor is further configured to: determine output scores for a plurality of sub-regions within the geographical location by inputting multi-channel input images for each sub-region into the trained machine learning model, identify one or more of the sub-regions having a highest probability of containing regulated waters based on the output scores, generate a map overlay highlighting the identified sub-regions, and display the map overlay superimposed on the aerial imagery data on a user interface to visually indicate areas within the geographical location most likely to contain waters regulated under the Clean Water Act.

[0127] The disclosure also provides support for a method for training a convolutional neuralnetwork to predict jurisdictional determinations for water resources under the Clean Water act, the method comprising: generating a training dataset comprising a plurality of approved jurisdictional determinations (AJDs) for water resources located at a plurality of geographic locations, wherein each AJD of the plurality of AJDs includes: geographic coordinates identifying a location of a water resource, a ground truth regulatory status indicating whether the water resource is regulated under the Clean Water act, a ground truth Cowardin wetland classification code assigned to the water resource, a ground truth legal resource type assigned to the water resource, and a rule under which the AJD was decided from among a plurality of rules for determining regulatory status, preprocessing the training dataset by generating a plurality of multi-channel input images, each multichannel input image centered on a respective geographic location of the plurality of geographic locations, and wherein each multi-channel input image comprises a plurality of input data layers, training the convolutional neural network to map the plurality of multi-channel input images to a predicted regulatory status, a predicted Cowardin wetland classification code, and a predicted legal resource type, wherein the training comprises: inputting the plurality of multi-channel input images into the convolutional neural network, obtaining the predicted regulatory status, the predicted Cowardin wetland classification code, and the predicted legal resource type from the convolutional neural network for each of the plurality of multi-channel input images, computing a loss by comparing the predicted regulatory status, the predicted Cowardin wetland classification code, and the predicted legal resource type to the ground truth regulatory status, the ground truth Cowardin wetland classification code, and the ground truth legal resource type, respectively, for each of the plurality of AJDs, and updating parameters of the convolutional neural network based on the computed loss to minimize future losses on the training dataset. In a first example of the method, the plurality of input data layers include: aerial imagery data for the location, vector data representing locations and characteristics of wetlands and streams proximate to the location, soil characteristic data for the location, climate data for the location including precipitation and temperature variables, elevation data for the location, land cover classification data for the location, data identifying an ecoregion classification for the location, data identifying a distance of the location to a nearest headquarters of an environmental regulatory agency, data encoding a state encompassing the location, data encoding a regulatory district encompassing the location, and data encoding the rule under which the AJD was decided. In a second example of the method, optionally including the first example, the vector data representing locations and characteristics of wetlandsand streams proximate to the location comprises data from the National Wetlands Inventory and the National Hydrography Dataset, and wherein the soil characteristic data for the location comprises data extracted from the Gridded National Soil Survey Geographic Database. In a third example of the method, optionally including one or both of the first and second examples, the convolutional neural network is trained to simultaneously predict the regulatory status, the Cowardin wetland classification code, and the legal resource type for each of the plurality of AJDs. In a fourth example of the method, optionally including one or more or each of the first through third examples, training the convolutional neural network comprises: determining a first loss by comparing the predicted regulatory status to the ground truth regulatory status for each of the plurality of AJDs, determining a second loss by comparing the predicted Cowardin wetland classification code to the ground truth Cowardin wetland classification code for each of the plurality of AJDs, determining a third loss by comparing the predicted legal resource type to the ground truth legal resource type for each of the plurality of AJDs, and updating parameters of the convolutional neural network based on the first loss, the second loss, and the third loss.

[0128] The description of a range herein should be considered to have specifically disclosed all the possible subranges as well of any individual numerical values within that range to the tenth of the unit of the lower limit unless the context clearly dictates otherwise. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well of any dividual values within that range, for example, 1.1, 2, 2.3, 5, and 5.9. This applies regardless of the breadth of the range. The upper and lower limits of these intervening ranges may independently be included in the smaller ranges, and are also encompassed within the invention, subject to any specifically excluded limit in the stated range. Where the stated range includes one or both of the limits, ranges excluding either or both of those included limits are also included in the invention, unless the context clearly dictates otherwise.

[0129] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of any embodiment. As used herein, the singular forms “a,” “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and / or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more otherfeatures, integers, steps, operations, elements, components, and / or groups thereof. As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items. .

[0130] The previously described versions of the disclosed subject matter have many advantages that were either described or would be apparent to a person of ordinary skill. Even so, these advantages or features are not required in all versions of the disclosed apparatus, systems, or methods.

[0131] Additionally, this written description makes reference to particular features. It is to be understood that the disclosure in this specification includes all possible combinations of those particular features. Where a particular feature is disclosed in the context of a particular aspect or example, that feature can also be used, to the extent possible, in the context of other aspects and examples.

[0132] Also, when reference is made in this application to a method having two or more defined steps or operations, the defined steps or operations can be carried out in any order or simultaneously, unless the context excludes those possibilities.

[0133] Although specific examples of the invention have been illustrated and described for purposes of illustration, it will be understood that various modifications may be made without departing from the spirit and scope of the invention.

Claims

CLAIMS:

1. A method for determining regulatory status of water resources under the Clean Water Act, the method comprising: receiving a plurality of input data layers for a geographic location, the plurality of input data layers comprising aerial imagery data, wetland location data, stream location data, soil data, elevation data, climate data, land cover classification data, and regulatory boundary data for the geographic location; preprocessing the plurality of input data layers into a multi-channel image, wherein each pixel of the multi-channel image corresponds to the geographic location and comprises values from the plurality of input data layers; feeding the multi-channel image into a trained convolutional neural network comprising a plurality of convolutional layers; obtaining an output score from the trained convolutional neural network, the output score indicating a probability that the geographic location contains waters regulated under the Clean Water Act; and determining a regulatory status for the geographic location based on the output score, wherein the geographic location is determined to contain regulated waters when the output score exceeds a threshold value.

2. The method of claim 1, wherein the plurality of input data layers further comprises a rule encoding layer encoding a rule to be applied for determining regulatory status under the Clean Water Act, the rule encoding layer comprising a rule encoding vector for each pixel of the multichannel image.

3. The method of claim 2, wherein the rule encoding layer comprises an embedding vector for each pixel of the multi-channel image, wherein the embedding vector encodes text of a rule to be applied for determining regulatory status under the Clean Water Act for the geographic location corresponding to the pixel.

4. The method of claim 1, wherein the trained convolutional neural network is further configured to output a classification score for each of a plurality of hydrologic types, the classification score indicating a probability that the geographic location corresponds to an associated hydrologic type.

5. The method of claim 4, wherein the plurality of hydrologic types comprises a Cowardin wetland classification code.

6. The method of claim 1, wherein feeding the multi-channel image into the trained convolutional neural network comprises: inputting the multi-channel image into an input layer of the trained convolutional neural network; processing the multi-channel image through a plurality of convolutional layers of the trained convolutional neural network; performing an average pooling operation on an output of the plurality of convolutional layers; and passing an output of the average pooling operation through a fully connected layer to produce the output score.

7. The method of claim 1, wherein the trained convolutional neural network is further configured to output a classification score for each of a plurality of legal resource types defined by the Clean Water Act, the classification score indicating a probability that the geographic location corresponds to an associated legal resource type.

8. The method of claim 7, wherein the plurality of legal resource types comprise one or more of an ephemeral feature, an isolated water, a non-relatively permanent water, a relatively permanent water, a traditional navigable water, and a wetland adjacent to a regulated water.

9. A data processing system for classifying a geographical location as containing waters regulated under the Clean Water Act, comprising: a non-transitory memory storing instructions, a trained machine learning model, and aplurality of input data layers comprising geospatial data for the geographical location, wherein the plurality of input data layers include aerial imagery data, wetlands data, streams data, soil data, groundwater data, elevation data, climate data, land cover data, and geographic data identifying a state, an Army Corps of Engineers (ACE) district, and a Clean Water Act rule applicable to the geographical location; a processor coupled to the non-transitory memory and configured to execute the instructions to: generate a multi-channel input image for the geographical location by combining the plurality of input data layers; input the multi-channel input image into the trained machine learning model, wherein the trained machine learning model is trained on a dataset of approved jurisdictional determinations issued by the Army Corps of Engineers, each approved jurisdictional determination indicating whether a corresponding location contains waters regulated under the Clean Water Act; receive, from the trained machine learning model, an output score indicating a probability that the geographical location contains waters regulated under the Clean Water Act for the Clean Water Act rule; determine, based on the output score, whether the geographical location contains waters regulated under the Clean Water Act for the Clean Water Act rule; and generate an alert indicating the determination of whether the geographical location contains waters regulated under the Clean Water Act.

10. The data processing system of claim 9, wherein the alert comprises sending a message to a user device, the message specifying one or more locations corresponding to the geographical location that have been classified as containing waters regulated under the Clean Water Act.

11. The data processing system of claim 9, wherein the trained machine learning model is further configured to classify the geographical location into one or more categories of water resource types based on the multi-channel input image, wherein the categories of water resource types include ephemeral streams, intermittent streams, perennial streams, emergent wetlands, forested wetlands, isolated wetlands, wetlands adjacent to regulated waters, and traditional navigable waters.

12. The data processing system of claim 11, wherein the processor is further configured to generate a visualization displaying the geographical location and an indication of a type of water resource classified by the trained machine learning model.

13. The data processing system of claim 9, wherein the plurality of input data layers further comprise vector data identifying locations of wetlands and streams proximal to the geographical location.

14. The data processing system of claim 9, further comprising a user interface configured to receive a user input specifying the geographical location, wherein the processor generates the multi-channel input image for the geographical location specified by the user input.

15. The data processing system of claim 9, wherein the processor is further configured to: determine output scores for a plurality of sub-regions within the geographical location by inputting multi-channel input images for each sub-region into the trained machine learning model; identify one or more of the sub-regions having a highest probability of containing regulated waters based on the output scores; generate a map overlay highlighting the identified sub-regions; and display the map overlay superimposed on the aerial imagery data on a user interface to visually indicate areas within the geographical location most likely to contain waters regulated under the Clean Water Act.

16. A method for training a convolutional neural network to predict jurisdictional determinations for water resources under the Clean Water Act, the method comprising: generating a training dataset comprising a plurality of approved jurisdictional determinations (AJDs) for water resources located at a plurality of geographic locations, wherein each AJD of the plurality of AJDs includes: geographic coordinates identifying a location of a water resource; a ground truth regulatory status indicating whether the water resource is regulated under the Clean Water Act;a ground truth Cowardin wetland classification code assigned to the water resource; a ground truth legal resource type assigned to the water resource; and a rule under which the AJD was decided from among a plurality of rules for determining regulatory status; pre-processing the training dataset by generating a plurality of multi-channel input images, each multi-channel input image centered on a respective geographic location of the plurality of geographic locations, and wherein each multi-channel input image comprises a plurality of input data layers; training the convolutional neural network to map the plurality of multi-channel input images to a predicted regulatory status, a predicted Cowardin wetland classification code, and a predicted legal resource type, wherein the training comprises: inputting the plurality of multi-channel input images into the convolutional neural network; obtaining the predicted regulatory status, the predicted Cowardin wetland classification code, and the predicted legal resource type from the convolutional neural network for each of the plurality of multi-channel input images; computing a loss by comparing the predicted regulatory status, the predicted Cowardin wetland classification code, and the predicted legal resource type to the ground truth regulatory status, the ground truth Cowardin wetland classification code, and the ground truth legal resource type, respectively, for each of the plurality of AJDs; and updating parameters of the convolutional neural network based on the computed loss to minimize future losses on the training dataset.

17. The method of claim 16, wherein the plurality of input data layers include: aerial imagery data for the location; vector data representing locations and characteristics of wetlands and streams proximate to the location; soil characteristic data for the location; climate data for the location including precipitation and temperature variables; elevation data for the location; land cover classification data for the location;data identifying an ecoregion classification for the location; data identifying a distance of the location to a nearest headquarters of an environmental regulatory agency; data encoding a state encompassing the location; data encoding a regulatory district encompassing the location; and data encoding the rule under which the AJD was decided.

18. The method of claim 17, wherein the vector data representing locations and characteristics of wetlands and streams proximate to the location comprises data from the National Wetlands Inventory and the National Hydrography Dataset, and wherein the soil characteristic data for the location comprises data extracted from the Gridded National Soil Survey Geographic Database.

19. The method of claim 16, wherein the convolutional neural network is trained to simultaneously predict the regulatory status, the Cowardin wetland classification code, and the legal resource type for each of the plurality of AJDs.

20. The method of claim 19, wherein training the convolutional neural network comprises: determining a first loss by comparing the predicted regulatory status to the ground truth regulatory status for each of the plurality of AJDs; determining a second loss by comparing the predicted Cowardin wetland classification code to the ground truth Cowardin wetland classification code for each of the plurality of AJDs; determining a third loss by comparing the predicted legal resource type to the ground truth legal resource type for each of the plurality of AJDs; and updating parameters of the convolutional neural network based on the first loss, the second loss, and the third loss.

Images