Predicting Rangeland Herbaceous Biomass with Multi-Branch LSTMs

The problem#

How much forage a rangeland is carrying, and whether that stock is growing or shrinking, is the quantity a rancher and a land manager actually act on. Estimating it well means combining signals that live on completely different footings: what satellites see every sixteen days, how much rain fell each day, where cattle actually grazed, and what kind of vegetation sits underneath. Each is a time series with its own cadence, noise profile, and physical meaning, and the interesting behavior is in how they interact.

Two targets are modeled: the standing herbaceous biomass in a given period, and the change in that biomass from the previous period. The second is the harder and the more useful of the two, and it is the one reported here.

Data#

Two grazed pastures on a ranch near Animas, New Mexico, across the 2025 season. Four input groups feed the model:

• Satellite biomass history, from 16-day RAP herbaceous above-ground biomass composites.
• Precipitation, a daily on-ranch weather station series, supplied as the 32-day window ending the day before each prediction.
• Grazing pressure, derived from GPS collars combined with collar behavior classification, as animal grazing-days per pixel over the same 32-day window.
• Land cover, six static fractional-cover bands per pixel.

The full pixel-by-period grid is 1.27 million records, one record being a single roughly 30 m pixel on a single 16-day date. Because most of that grid is ground that never saw an animal, models are trained and evaluated on the 74,374 records with grazing activity in their trailing window. Every number below is on that filtered set.

Approach#

Rather than flattening every input into one feature vector, the model uses a multi-branch design: the three sequence inputs each get their own LSTM encoder, so a rainfall series and a grazing-intensity series are read on their own terms instead of being concatenated early and lossily. Land cover is static per pixel, so it carries no temporal structure to encode and is passed straight into the fusion step. The branches meet late, at a two-layer head, which is where the network learns how the signals combine.

Ten models in all: five input combinations, each trained against both targets, forming a ladder where each rung adds one input group.

Results#

On a held-out test split of 14,747 records, for the biomass-change target:

Reading the ladder:

• Rainfall does most of the work. Adding it to a history-only model moves R² from 0.364 to 0.888, by far the largest single jump.
• Land cover is the surprise. It is the second most valuable input despite being the one input with no encoder at all, lifting R² a further 0.044.
• Grazing pressure adds no measurable improvement. The best model, at R² 0.932, is the one without the collar-derived branch; adding it back moves R² to 0.921, a difference small enough to sit inside run-to-run noise.

What this shows#

The headline is methodological as much as ecological. Giving each sequence input its own temporal encoder and fusing late is what moved biomass-change R² from 0.36 to 0.93. But the two most interesting findings both cut against intuition: the input that needed no temporal encoder at all turned out to be the second most valuable, and cattle-location data, for all its obvious appeal in a grazing study, carried no signal that rainfall and land cover had not already explained.

Two caveats worth stating plainly. The test split is random by pixel within a single 2025 season, so every date appears in both training and test data; the results speak to spatial generalization, not to an unseen year. And the grazing-pressure finding rests on a difference of about 0.011 R² from a single training run, which makes it suggestive rather than settled.