Skip to content
Rupak Dey

Navigate

Work

Links

Theme

← Research

Research

Predicting Net Primary Productivity with Multi-Branch LSTMs

Role
Machine Learning Research Assistant (Senior Thesis)
Lab
KDD Lab, New Mexico State University
Period
Sep 2025 – Present
Methods
  • LSTM
  • multi-modal fusion
  • temporal encoding
  • remote sensing
  • time-series regression

The problem#

Net primary productivity (NPP), how much plant biomass an ecosystem produces over time, is a central quantity for managing rangeland and understanding how forage responds to weather and grazing. Estimating it well means combining signals that live on completely different footings: what satellites see, how much rain falls, where cattle actually graze, and what kind of land cover sits underneath. Each of these is a time series with its own cadence, noise profile, and physical meaning, and the interesting behavior is in how they interact.

Data#

The work draws on four modalities across 71,000+ pixels:

  • Satellite imagery.
  • Precipitation records.
  • Cattle GPS tracks, which localize grazing pressure in space and time.
  • Land cover classification.

Approach#

Rather than flattening every input into one feature vector, the model uses a multi-branch LSTM: a separate temporal encoder per modality, so each data source is read on its own terms before the branches are fused for prediction. This preserves the distinct temporal structure of, say, a precipitation series versus a grazing-intensity series, and lets the network learn how they combine rather than forcing an early, lossy concatenation.

In total, seven LSTM models were built and compared across configurations.

Multi-branch LSTM architecture: four input modalities, one temporal encoder each, fused late into a single NPP prediction.SatelliteLSTM encoderPrecipitationLSTM encoderCattle GPSLSTM encoderLand coverLSTM encoderLate fusionNPPprediction

Architecture sketch: each modality is read by its own temporal encoder; branches fuse late, so the network learns how the signals combine instead of forcing an early concatenation.

Results#

  • The multi-branch architecture reaches R² 0.98 on the NPP prediction task.
  • Benchmarked against a multilayer perceptron (MLP) and a random forest (RF), the LSTM approach comes out ahead.
  • Adding cattle-GPS data cut prediction error by 36% relative to a precipitation-only model, direct evidence that where animals graze carries real signal about productivity, not just how much it rains.

What this shows / next steps#

The headline finding is methodological as much as ecological: giving each modality its own encoder and fusing late beats both classical baselines and a naive single-stream network, and the grazing signal matters.