Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates

This paper proposes a probabilistic transformer-based framework that integrates historical NDVI observations with meteorological covariates and a temporal-distance weighted quantile loss to accurately forecast field-level vegetation dynamics despite sparse, cloud-affected satellite data.

The Gist

Imagine you are a farmer trying to decide when to water your crops or add fertilizer. You need to know: "How healthy will my plants be in the next two weeks?"

To answer this, you look at satellite images of your fields. These images show a "greenness score" called NDVI. The greener the number, the healthier the plants.

The Problem:
Satellites are like helpful but unreliable friends. They want to take a picture of your farm every few days, but clouds often block their view. Sometimes they miss a week; sometimes they miss two. This leaves you with a messy, broken timeline of greenness scores. Plus, the weather (rain, heat, cold) changes constantly, and it's hard to tell exactly how a rainy day last week will affect your crops two weeks from now.

The Solution (The Paper's Idea):
The researchers built a smart AI "weather forecaster" for plants. Instead of just guessing, they created a system that acts like a super-organized chef preparing a meal for the future.

Here is how their "recipe" works, broken down into simple parts:

1. The Two-Track Kitchen (The Architecture)

Imagine a kitchen with two separate counters:

• Counter A (The Past): This is where the chef looks at the history. "What did the plants look like last month? Did they get wet? Did they get hot?" The AI studies the broken, cloudy satellite photos to understand the plant's story.
• Counter B (The Future): This is where the chef looks at the weather forecast. "The meteorologist says it will rain on Tuesday and be scorching hot on Friday." The AI takes this future weather data and prepares for it.

The magic happens when the chef mixes the two counters together. The AI doesn't just look at the past or the future; it combines the plant's history with the upcoming weather to predict what will happen next.

2. The "Missing Puzzle Piece" Fix (Handling Clouds)

Since the satellite photos are full of holes (clouds), the AI has to be smart about filling them in.

• The Analogy: Imagine a timeline of your plant's health. If you have a photo from Monday and the next one is from Friday, but the clouds hid Tuesday, Wednesday, and Thursday, the AI doesn't just guess randomly. It draws a smooth, logical line between Monday and Friday, knowing exactly how much time passed. It respects the real time gaps, rather than pretending the days were evenly spaced.

3. The "Weather Diary" (Feature Engineering)

The AI doesn't just look at "It rained 1 inch today." It keeps a weather diary that counts things that matter to plants:

• "How many cold days (below 10°C) happened since the last photo?"
• "How many hot days (above 30°C) happened?"
• "What was the total rain between the last two photos?"

This helps the AI understand that a single rainy day might not matter, but a week of cold snaps could stunt the plant's growth.

4. The "Confidence Meter" (Probabilistic Forecasting)

Most weather apps just say, "It will be 75°F." This AI is smarter. It says, "It will likely be 75°F, but there's a chance it could be as low as 65°F or as high as 85°F."

• The Analogy: Instead of giving you one single answer, it gives you a range of possibilities (like a safety net). It tells the farmer: "We are 90% sure the plants will be this green, but here is the worst-case and best-case scenario." This helps farmers plan better because they know the risks.

5. The "Time-Weighted" Score (The Loss Function)

When the AI learns from its mistakes, it treats errors differently based on time.

• The Analogy: If the AI guesses wrong about what happens tomorrow, that's a big mistake. If it guesses wrong about what happens two weeks from now, it's expected to be a bit less accurate because the future is fuzzier. The AI's training system gives a "gentler" penalty for long-term guesses and a "stricter" penalty for short-term ones, so it learns to be very sharp for the immediate future.

The Result

The researchers tested this "Super Chef" AI against other methods (like old-school math formulas and other AI models) using data from farms all over Europe.

• The Winner: Their model was the most accurate at predicting plant health, even when the satellite photos were messy and the weather was weird.
• The Takeaway: By combining the plant's history with future weather forecasts and being honest about uncertainty, this AI helps farmers make better decisions, saving water, money, and ensuring a good harvest.

In short: They built a smart system that turns messy, cloudy satellite photos and weather forecasts into a clear, reliable roadmap for the future of your crops.

Technical Summary

1. Problem Statement

The paper addresses the challenge of short-term forecasting of the Normalized Difference Vegetation Index (NDVI) at the field level for precision agriculture. While NDVI is a critical proxy for vegetation health and canopy development, operational forecasting faces two major hurdles:

• Sparse and Irregular Sampling: Satellite observations (specifically Sentinel-2) are frequently obscured by cloud cover, resulting in time series with missing values and irregular temporal intervals between valid acquisitions.
• Heterogeneous Conditions: Forecasting models often struggle to generalize across diverse European climatic zones and varying growing seasons due to the complex interplay between vegetation dynamics and meteorological factors.

Existing methods often rely on regularized or densified time series, which may introduce artifacts, or focus on pixel-level dense maps rather than field-level trajectories required for actionable decision support (e.g., irrigation scheduling).

2. Methodology

The authors propose a probabilistic forecasting framework based on a Transformer architecture designed to handle irregular sampling and integrate multimodal data.

A. Data Preprocessing and Feature Engineering

• Data Source: The study uses the GreenEarthNet dataset, comprising cloud-masked Sentinel-2 images and meteorological time series across Europe (2017–2022).
• Handling Irregularity:
  • Reconstruction: Missing NDVI values caused by clouds are reconstructed using time-aware linear interpolation between valid clear-sky observations, preserving the true temporal distance between acquisitions.
  • Sampling Strategy: The model operates on a sliding window over the sequence of available (observed or reconstructed) target values, rather than fixed calendar days.
• Weather Feature Engineering: To capture delayed and extreme weather effects, the authors derive 9 features from raw meteorological data (precipitation, temperature):
  • Between-target cumulative features: Aggregations (rainfall, cold/hot days) over the variable-length intervals between satellite overpasses.
  • Rolling-window features: Aggregations over fixed windows (7 and 14 days) to capture short-to-medium-term trends independent of sampling irregularity.
• Uncertainty Modeling: Future weather covariates are perturbed with noise that scales linearly with the forecasting horizon to simulate increasing forecast uncertainty.

B. Model Architecture

The core is a Transformer-based probabilistic model with decoupled branches:

1. History Encoder: Processes past NDVI targets and historical weather covariates. It uses a standard Transformer encoder with positional encoding and binary masking for missing values. The output is aggregated via temporal average pooling to create a compact context representation.
2. Future Encoder: Processes future weather covariates (which are known or forecasted). It uses a separate Transformer encoder. Crucially, it employs a sparse temporal selection mechanism that retains only the embeddings corresponding to actual future satellite acquisition dates, ignoring intermediate time steps where no observation is expected.
3. Fusion and Prediction: The pooled history representation is concatenated with the selected future embeddings. This joint representation is fed into a Quantile Head that predicts three quantiles ($q \in \{0.1, 0.5, 0.9\}$) for the next $h$ time steps in parallel (non-autoregressive).
   • $q=0.5$ serves as the point forecast.
   • $q=0.1$ and $q=0.9$ provide calibrated uncertainty bounds.

C. Loss Function

The model is trained using a Temporal-Distance Weighted Quantile Loss (Pinball Loss).

• Standard quantile loss is modified by a weight $w_k$ that decays based on the temporal distance ($\Delta t$) between the last historical observation and the future target.
• This weighting scheme aligns the training objective with the reality that uncertainty increases with the forecasting horizon and irregular revisit patterns.

3. Key Contributions

1. Transformer for Irregular Sampling: A novel architecture that explicitly separates historical vegetation dynamics from future exogenous information, utilizing sparse temporal selection to handle irregular satellite revisit schedules without requiring signal densification.
2. Temporal-Distance Weighted Loss: A custom loss function that accounts for the variable time gaps between observations and the horizon-dependent nature of forecasting uncertainty.
3. Advanced Weather Feature Engineering: A strategy to compute cumulative and extreme-weather features over variable-length intervals, better aligning daily meteorology with irregular NDVI observations.
4. Probabilistic Framework: The ability to output calibrated uncertainty estimates (quantiles) alongside point forecasts, which is essential for risk-aware agricultural decision-making.

4. Experimental Results

The model was evaluated on European data against a diverse set of baselines, including statistical models (AutoARIMA), recurrent networks (LSTM, RNN), convolutional models (InceptionTime), and recent Transformers (PatchTST, TimeLLM).

• Performance: The proposed model achieved state-of-the-art results across all metrics:
  • Point Forecasting: Lowest RMSE (0.0821), MAE (0.0509), and WMAPE.
  • Probabilistic Forecasting: Lowest CRPS (0.0315) and Pinball loss (0.0177), indicating superior uncertainty calibration.
  • Statistical significance (Diebold-Mariano test) confirmed improvements over all baselines ($p < 0.001$).
• Ablation Studies:
  • Target History: The most critical input; removing it caused the largest performance degradation.
  • Meteorological Covariates: Historical and future weather data provided complementary gains, improving both accuracy and uncertainty estimation.
  • Components: Both the temporal loss weighting and feature engineering contributed to performance, with loss weighting having the largest impact.
• Robustness: The model performed consistently across different Köppen–Geiger climate zones, though performance slightly degraded in Continental climates due to higher phenological variability.
• Efficiency: The model offers a favorable accuracy-complexity trade-off, requiring significantly fewer parameters and computational resources (MFLOPs) than PatchTST while achieving higher accuracy.

5. Significance

This work bridges the gap between theoretical deep learning and practical agricultural applications by addressing the real-world constraints of satellite data (cloud cover, irregularity).

• Decision Support: By providing reliable short-term (up to 14 days) forecasts with uncertainty quantification, the system enables farmers to make proactive decisions regarding irrigation, fertilization, and stress mitigation.
• Generalizability: The framework demonstrates robustness across diverse European ecozones, suggesting potential for broader application in precision agriculture.
• Methodological Advancement: It establishes a new paradigm for time-series forecasting where exogenous future information is integrated without assuming regular time steps, a common limitation in existing literature.

The code and pre-trained models are made publicly available to facilitate further research and deployment.