Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mind the Gap: Bridging Prior Shift in Realistic Few-Shot Crop-Type Classification

Joana Reuss, Ekaterina Gikalo, Marco Körner

TL;DR

This work tackles prior-shift in realistic few-shot crop-type classification by introducing Dirichlet Prior Augmentation (DirPA/DiPA), which proactively simulates unknown target priors during training through Dirichlet-distributed pseudo-priors and logit adjustment. The method acts as a dynamic feature regularizer, stabilizing training and improving generalization, particularly in low-shot regimes. Evaluated on Estonia/EuroCropsML time-series Sentinel-2 data with a Transformer backbone, DirPA yields higher accuracy and Cohen's kappa across most few-shot scenarios, while remaining effective as the data scale grows. The approach is model-agnostic and can be extended beyond crop-type classification to any few-shot task afflicted by training/test prior distribution mismatch.

Abstract

Real-world agricultural distributions often suffer from severe class imbalance, typically following a long-tailed distribution. Labeled datasets for crop-type classification are inherently scarce and remain costly to obtain. When working with such limited data, training sets are frequently constructed to be artificially balanced -- in particular in the case of few-shot learning -- failing to reflect real-world conditions. This mismatch induces a shift between training and test label distributions, degrading real-world generalization. To address this, we propose Dirichlet Prior Augmentation (DirPA), a novel method that simulates an unknown label distribution skew of the target domain proactively during model training. Specifically, we model the real-world distribution as Dirichlet-distributed random variables, effectively performing a prior augmentation during few-shot learning. Our experiments show that DirPA successfully shifts the decision boundary and stabilizes the training process by acting as a dynamic feature regularizer.

Mind the Gap: Bridging Prior Shift in Realistic Few-Shot Crop-Type Classification

TL;DR

This work tackles prior-shift in realistic few-shot crop-type classification by introducing Dirichlet Prior Augmentation (DirPA/DiPA), which proactively simulates unknown target priors during training through Dirichlet-distributed pseudo-priors and logit adjustment. The method acts as a dynamic feature regularizer, stabilizing training and improving generalization, particularly in low-shot regimes. Evaluated on Estonia/EuroCropsML time-series Sentinel-2 data with a Transformer backbone, DirPA yields higher accuracy and Cohen's kappa across most few-shot scenarios, while remaining effective as the data scale grows. The approach is model-agnostic and can be extended beyond crop-type classification to any few-shot task afflicted by training/test prior distribution mismatch.

Abstract

Real-world agricultural distributions often suffer from severe class imbalance, typically following a long-tailed distribution. Labeled datasets for crop-type classification are inherently scarce and remain costly to obtain. When working with such limited data, training sets are frequently constructed to be artificially balanced -- in particular in the case of few-shot learning -- failing to reflect real-world conditions. This mismatch induces a shift between training and test label distributions, degrading real-world generalization. To address this, we propose Dirichlet Prior Augmentation (DirPA), a novel method that simulates an unknown label distribution skew of the target domain proactively during model training. Specifically, we model the real-world distribution as Dirichlet-distributed random variables, effectively performing a prior augmentation during few-shot learning. Our experiments show that DirPA successfully shifts the decision boundary and stabilizes the training process by acting as a dynamic feature regularizer.

Paper Structure

This paper contains 19 sections, 5 equations, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Few-shot learning paradigm
  4. FSL as transfer vs. meta-learning
  5. FSL in remote sensing
  6. Class imbalance and prior shift in FSL
  7. Prior distribution shift correct at inference-time
  8. Dirichlet priors and distribution augmentation
  9. Dirichlet for FSL evaluation
  10. Summary and relation to our work
  11. Dataset
  12. Dataset split
  13. Methodology
  14. Dirichlet distribution
  15. DiPA: Dirichlet prior augmentation
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Abundances of crop types in Estonia. Histograms showing the binned distribution of crop-type abundances in Estonia for 1000.0 randomly sampled data points of the validation set and the full test set.
  • Figure 2: Spatial distribution of crop classes in the test set of Estonia. Map of Estonia showing the location and distribution of crop types within the final test set. Each data point marks the central coordinate of an agricultural parcel and is color-coded by its corresponding crop class.
  • Figure 3: Dirichlet density for $K=3$ (defined over the $(K-1)=2$-simplex) and different concentration parameters $\boldsymbol{\alpha}$.
  • Figure 4: Visualization of test metrics (including macro-averaged F1 score) across the $k$-shot benchmark tasks. The $x$-axis is plotted on a logarithmic scale. Metrics are shown as mean $\pm$ standard deviation over five runs, cf.\ref{['sec:experiments']}. The postfix indicates the use of the method. Due to the highly imbalanced nature of the 102.0-class classification task, the macro-F1 scores remain numerically low, reflecting the inherent challenge in achieving high performance on the numerous low-resource classes.

Theorems & Definitions (1)

  • Definition 1