[Paper Review] Learning to forecast vegetation greenness at fine resolution over Africa with ConvLSTMs

Claire Nicolle Robin, Christian Requena‐Mesa+5 more|arXiv (Cornell University)|Oct 24, 2022

Remote Sensing in Agriculture17 citations

TL;DR

This paper proposes a ConvLSTM-based model to forecast fine-resolution vegetation greenness across Africa using Sentinel-2 NDVI, ERA5 weather reanalysis, SMAP soil moisture, and SRTM topography as inputs. The model successfully predicts seasonal NDVI dynamics and differential responses to weather anomalies at 30m resolution, demonstrating strong potential for anticipatory action in drought-prone regions.

ABSTRACT

Forecasting the state of vegetation in response to climate and weather events is a major challenge. Its implementation will prove crucial in predicting crop yield, forest damage, or more generally the impact on ecosystems services relevant for socio-economic functioning, which if absent can lead to humanitarian disasters. Vegetation status depends on weather and environmental conditions that modulate complex ecological processes taking place at several timescales. Interactions between vegetation and different environmental drivers express responses at instantaneous but also time-lagged effects, often showing an emerging spatial context at landscape and regional scales. We formulate the land surface forecasting task as a strongly guided video prediction task where the objective is to forecast the vegetation developing at very fine resolution using topography and weather variables to guide the prediction. We use a Convolutional LSTM (ConvLSTM) architecture to address this task and predict changes in the vegetation state in Africa using Sentinel-2 satellite NDVI, having ERA5 weather reanalysis, SMAP satellite measurements, and topography (DEM of SRTMv4.1) as variables to guide the prediction. Ours results highlight how ConvLSTM models can not only forecast the seasonal evolution of NDVI at high resolution, but also the differential impacts of weather anomalies over the baselines. The model is able to predict different vegetation types, even those with very high NDVI variability during target length, which is promising to support anticipatory actions in the context of drought-related disasters.

Motivation & Objective

• Address the challenge of forecasting vegetation greenness at fine spatial resolution to support anticipatory action in climate-vulnerable regions.
• Overcome limitations of coarse-resolution models by leveraging high-resolution spatio-temporal data for localized ecosystem response prediction.
• Enable early warning systems for drought and extreme weather impacts by modeling time-lagged vegetation responses to climate drivers.
• Develop a strongly guided video prediction framework using environmental and meteorological variables to improve forecasting accuracy.
• Support Forecast-based Financing and humanitarian planning by providing reliable, localized vegetation forecasts across diverse African ecosystems.

Proposed method

• Formulate land surface forecasting as a strongly guided video prediction task using spatio-temporal data cubes of satellite and reanalysis variables.
• Employ a Convolutional Long Short-Term Memory (ConvLSTM) architecture to model complex spatio-temporal dependencies in vegetation dynamics.
• Input multiple drivers: Sentinel-2 reflectance (red, blue, green, NIR), derived NDVI (target), ERA5 meteorological variables (temperature, precipitation, radiation, pressure, evaporation), SMAP soil moisture (surface and root zone), and SRTMv4.1 digital elevation model (upsampled to 30m).
• Apply spatial and temporal masking using the ESRI 2020 land cover map to exclude non-vegetated and missing pixels during training and evaluation.
• Train the model for 150 epochs with a batch size of 32 and a learning rate of 1e-6 using PyTorch Lightning, minimizing L2 loss on NDVI predictions.
• Use future weather inputs during inference, reflecting a strongly guided forecasting setup where future conditions are known at prediction time.

Experimental results

Research questions

• RQ1 Can a ConvLSTM model accurately forecast fine-resolution (30m) vegetation greenness (NDVI) across diverse African ecosystems?
• RQ2 How well does the model capture time-lagged responses of vegetation to weather anomalies such as drought or early rainfall?
• RQ3 To what extent do topography, soil moisture, and meteorological variables improve prediction accuracy compared to NDVI alone?
• RQ4 How does prediction performance vary across different vegetation types (e.g., croplands vs. savannas vs. forests) and seasonal conditions?
• RQ5 What is the impact of data quality issues—particularly undetected clouds and artifacts—on model prediction reliability?

Key findings

• The ConvLSTM model successfully forecasts seasonal NDVI evolution at 30m spatial resolution across Africa with high fidelity.
• The model captures differential vegetation responses to weather anomalies, including time-lagged effects such as soil moisture recharge and drought legacy impacts.
• Prediction accuracy is strongly correlated with target variance, with higher accuracy for low-variance periods (e.g., non-growing seasons) and lower accuracy during high-variability growing seasons.
• Samples with persistent cloud artifacts show degraded performance, highlighting the impact of data quality on model reliability despite pre-processing.
• The model generalizes across diverse vegetation types, including those with high NDVI variability, indicating robustness for use in heterogeneous ecosystems.
• The inclusion of future weather inputs significantly improves prediction quality, confirming the value of strongly guided forecasting in this context.