The Harmful Algal Blooms (HABs) forecast is crucial for the mitigation of health hazards and to inform actions for the protection of ecosystems and fisheries in the Gulf of Mexico (GoM). For the sake of simplicity of our application we assume ocean color satellite imagery from the National Oceanic and Atmospheric Administration as a proxy for HABs.
In this study we use a deep neural network trained on the 2-Dimensional time series proxy data to provide a forecast of the HABs’ manifestations in the GoM.Our approach analyzes between both spatial and temporal features simultaneously. In addition, the network also helps to fill in the gaps of the time series data along the way. We use Long Short Term Memory (LSTM) layers to learn the underlying trends in the time series data and Convolutional layers to decode the spatial trends in the 2-Dimensional gridded data.
Our unique contribution is an iterative, bidirectional training scheme, where we train two models: for forward and backward prediction. The intention is that if there is a functional dependence within the data in the forward time direction, then such a dependence may also exist in the backward time direction, which may be leveraged for predictions to fill the gaps in the data. We train each model to predict the next data point in their respective time-direction, based on an LSTM recurrence over the “lookback” data points. Since there are missing cells in the grid within each data point, we use a custom loss function that ignores prediction errors on missing cells. Thus the loss function critiques the models based on known cells alone, while the models act with (forward/backward) predictions that are spatiotemporally consistent across both missing and visible cells, thus updating the input training data, and consequently changing the object of critique. This actor-critic training scheme progresses iteratively, leading to the iterative improvement of the models/actors.
Several models are developed with varying combinations of convolutional layers and max pooling layers to enable the model to learn the spatial and temporal trends within the month-long training data. The most effective model performs reasonably well with prediction of chlorophyll intensities.
","abstract_html":"<p>The Harmful Algal Blooms (HABs) forecast is crucial for the mitigation of health hazards and to inform actions for the protection of ecosystems and fisheries in the Gulf of Mexico (GoM). For the sake of simplicity of our application we assume ocean color satellite imagery from the National Oceanic and Atmospheric Administration as a proxy for HABs.</p> <p>In this study we use a deep neural network trained on the 2-Dimensional time series proxy data to provide a forecast of the HABs’ manifestations in the GoM.Our approach analyzes between both spatial and temporal features simultaneously. In addition, the network also helps to fill in the gaps of the time series data along the way. We use Long Short Term Memory (LSTM) layers to learn the underlying trends in the time series data and Convolutional layers to decode the spatial trends in the 2-Dimensional gridded data.</p> <p>Our unique contribution is an iterative, bidirectional training scheme, where we train two models: for forward and backward prediction. The intention is that if there is a functional dependence within the data in the forward time direction, then such a dependence may also exist in the backward time direction, which may be leveraged for predictions to fill the gaps in the data. We train each model to predict the next data point in their respective time-direction, based on an LSTM recurrence over the “lookback” data points. Since there are missing cells in the grid within each data point, we use a custom loss function that ignores prediction errors on missing cells. Thus the loss function critiques the models based on known cells alone, while the models act with (forward/backward) predictions that are spatiotemporally consistent across both missing and visible cells, thus updating the input training data, and consequently changing the object of critique. This actor-critic training scheme progresses iteratively, leading to the iterative improvement of the models/actors.</p> <p>Several models are developed with varying combinations of convolutional layers and max pooling layers to enable the model to learn the spatial and temporal trends within the month-long training data. The most effective model performs reasonably well with prediction of chlorophyll intensities.</p>","abstract_has_math":false,"creators":["Yerrapothu, Bala Tripura Sundari"],"institution":null,"degree_name":"Master of Science (MS)","degree_level":"Masters Thesis","degree_discipline":null,"degree_department":null,"school":null,"contributors":["Dr. Bikramjit Banerjee","Dr. Diana Bernstein","Dr. Ramakalavathi Marapareddy"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2021,"date_issued":"2021-05-01T07:00:00Z","date_published":"2021-05-01T07:00:00Z","updated_at":"2026-07-24T05:45:19Z","subjects":["Neural Network","Forecasting","LSTM","Prediction","Machine Learning","Artificial Intelligence and Robotics","Computer Sciences","Data Science","Environmental Sciences","Oceanography and Atmospheric Sciences and Meteorology"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://aquila.usm.edu/masters_theses/809","outbound_label":"Repository record","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Dr. Bikramjit Banerjee","Dr. Diana Bernstein","Dr. Ramakalavathi Marapareddy"]},{"key":"dc:creator","label":"Author","values":["Yerrapothu, Bala Tripura Sundari"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.available","label":"Dc Date Available","values":["2021-03-26T07:00:00Z"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Masters Thesis"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master of Science (MS)"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Neural Network","Forecasting","LSTM","Prediction","Machine Learning","Artificial Intelligence and Robotics","Computer Sciences","Data Science","Environmental Sciences","Oceanography and Atmospheric Sciences and Meteorology"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://aquila.usm.edu/masters_theses/809"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["The Harmful Algal Blooms (HABs) forecast is crucial for the mitigation of health hazards and to inform actions for the protection of ecosystems and fisheries in the Gulf of Mexico (GoM). For the sake of simplicity of our application we assume ocean color satellite imagery from the National Oceanic and Atmospheric Administration as a proxy for HABs.
In this study we use a deep neural network trained on the 2-Dimensional time series proxy data to provide a forecast of the HABs’ manifestations in the GoM.Our approach analyzes between both spatial and temporal features simultaneously. In addition, the network also helps to fill in the gaps of the time series data along the way. We use Long Short Term Memory (LSTM) layers to learn the underlying trends in the time series data and Convolutional layers to decode the spatial trends in the 2-Dimensional gridded data.
Our unique contribution is an iterative, bidirectional training scheme, where we train two models: for forward and backward prediction. The intention is that if there is a functional dependence within the data in the forward time direction, then such a dependence may also exist in the backward time direction, which may be leveraged for predictions to fill the gaps in the data. We train each model to predict the next data point in their respective time-direction, based on an LSTM recurrence over the “lookback” data points. Since there are missing cells in the grid within each data point, we use a custom loss function that ignores prediction errors on missing cells. Thus the loss function critiques the models based on known cells alone, while the models act with (forward/backward) predictions that are spatiotemporally consistent across both missing and visible cells, thus updating the input training data, and consequently changing the object of critique. This actor-critic training scheme progresses iteratively, leading to the iterative improvement of the models/actors.
Several models are developed with varying combinations of convolutional layers and max pooling layers to enable the model to learn the spatial and temporal trends within the month-long training data. The most effective model performs reasonably well with prediction of chlorophyll intensities.
"]},{"key":"dc:title","label":"Title","values":["Application of Machine Learning Techniques to Forecast Harmful Algal Blooms in Gulf of Mexico"]}]}],"canonical_facts":{"dc:contributor":["Dr. Bikramjit Banerjee","Dr. Diana Bernstein","Dr. Ramakalavathi Marapareddy"],"dc:creator":["Yerrapothu, Bala Tripura Sundari"],"dc:date.available":["2021-03-26T07:00:00Z"],"dc:description.abstract":["The Harmful Algal Blooms (HABs) forecast is crucial for the mitigation of health hazards and to inform actions for the protection of ecosystems and fisheries in the Gulf of Mexico (GoM). For the sake of simplicity of our application we assume ocean color satellite imagery from the National Oceanic and Atmospheric Administration as a proxy for HABs.
In this study we use a deep neural network trained on the 2-Dimensional time series proxy data to provide a forecast of the HABs’ manifestations in the GoM.Our approach analyzes between both spatial and temporal features simultaneously. In addition, the network also helps to fill in the gaps of the time series data along the way. We use Long Short Term Memory (LSTM) layers to learn the underlying trends in the time series data and Convolutional layers to decode the spatial trends in the 2-Dimensional gridded data.
Our unique contribution is an iterative, bidirectional training scheme, where we train two models: for forward and backward prediction. The intention is that if there is a functional dependence within the data in the forward time direction, then such a dependence may also exist in the backward time direction, which may be leveraged for predictions to fill the gaps in the data. We train each model to predict the next data point in their respective time-direction, based on an LSTM recurrence over the “lookback” data points. Since there are missing cells in the grid within each data point, we use a custom loss function that ignores prediction errors on missing cells. Thus the loss function critiques the models based on known cells alone, while the models act with (forward/backward) predictions that are spatiotemporally consistent across both missing and visible cells, thus updating the input training data, and consequently changing the object of critique. This actor-critic training scheme progresses iteratively, leading to the iterative improvement of the models/actors.
Several models are developed with varying combinations of convolutional layers and max pooling layers to enable the model to learn the spatial and temporal trends within the month-long training data. The most effective model performs reasonably well with prediction of chlorophyll intensities.
"],"dc:identifier":["https://aquila.usm.edu/masters_theses/809"],"dc:subject":["Neural Network","Forecasting","LSTM","Prediction","Machine Learning","Artificial Intelligence and Robotics","Computer Sciences","Data Science","Environmental Sciences","Oceanography and Atmospheric Sciences and Meteorology"],"dc:title":["Application of Machine Learning Techniques to Forecast Harmful Algal Blooms in Gulf of Mexico"],"thesis:degree_level":["Masters Thesis"],"thesis:degree_name":["Master of Science (MS)"]},"updated_at":"2026-07-24T05:45:19Z"}