cs.LGMar 13, 2026

DirPA: Addressing Prior Shift in Imbalanced Few-shot Crop-type Classification

Authors: Joana Reuss, Ekaterina Gikalo, Marco Körner

Organizations: Technical University of Munich (TUM), TUM School of Engineering and Design, Department of Aerospace and Geodesy, Chair of Remote Sensing Technology, Munich, 80333, Germany · Technical University of Munich (TUM), Munich Data Science Institute (MDSI), Garching, 85748, Germany · ELLIS Unit Jena, University of Jena, Jena, Germany, Jena, 07743, Germany

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Real-world agricultural monitoring is often hampered by severe class imbalance and high label acquisition costs, resulting in significant data scarcity. In few-shot learning (FSL) -- a framework specifically designed for data-scarce settings -- , training sets are often artificially balanced. However, this creates a disconnect from the long-tailed distributions observed in nature, leading to a distribution shift that undermines the model's ability to generalize to real-world agricultural tasks. We previously introduced Dirichlet Prior Augmentation (DirPA; Reuss et al., 2026a) to proactively mitigate the effects of such label distribution skews during model training. In this work, we extend the original study's geographical scope. Specifically, we evaluate this extended approach across multiple countries in the European Union (EU), moving beyond localized experiments to test the method's resilience across diverse agricultural environments. Our results demonstrate the effectiveness of DirPA across different geographical regions. We show that DirPA not only improves system robustness and stabilizes training under extreme long-tailed distributions, regardless of the target region, but also substantially improves individual class-specific performance by proactively simulating priors.

Figures & tables

Figure 1: Visualization of a label (prior) distribution shift between train dataset and a long-tailed test dataset.
Country# Crop classes# Parcels
Austria98982 589 1922\,589\,192
Belgium136136560 671560\,671
Flanders120120223 269223\,269
Wallonia9191337 402337\,402
Estonia127127175 905175\,905
Germany1981981 632 6201\,632\,620
Table 1: Dataset details of EuroCropsML 2.0 . The # Crop classes column refers to the number of unique crop types in HCATlevel 66 . The # Parcels column refers to unique parcel geometries after removing potential duplicates from the EuroCrops reference data and filtering out clouds, cirrus, snow, and ice.
Figure 2: Pre-training and fine-tuning country split. Countries colored in blue are used for pre-training, whereas those colored in green are used for fine-tuning. The darker the coloring, the denser the labeled agricultural fields.
Figure 3: EuroCropsML 2.0 binned class distribution for Germany (LSand NRW) and Austria, shown on a logarithmic scale, with the pasture meadow grassland grass class removed.
Figure 4: The number of annotated crop classes that are shared and distinct between the pre-training countries—Austria and Germany—and the respective fine-tuning country.
Figure 5: Biogeographical regions present in EuroCropsML 2.0 , as defined by the ( European Environment Agency (EEA), 2016 ) .
Country# Crop classes
 Train ∖\setminus (validation ∪\cup test)ValidationTestValidation ∩\cap test
Belgium22133133133133132132
Estonia1515107107108108103103
Latvia22106106105105105105
Lithuania00222222222222
Portugal1010727271716969
Table 2: Class distribution of the EuroCropsML 2.0 fine-tuning sets, showing the number of distinct crop classes. The total number of classes in the training set equals the number of available crop types and can be extracted from Table 1 .
CountryGini coefficient CGC_{\text{G}}
Belgium0.9390.939
Estonia0.9190.919
Latvia0.9100.910
Lithuania0.7610.761
Portugal0.8810.881
Slovakia0.9040.904
Table 3: Gini coefficient for the test sets of the EuroCropsML 2.0 fine-tuning countries, a measure of dataset skewness. A value of 00 indicates a perfectly balanced class distribution, while a value approaching 11 means almost all samples belong to a single majority class.
Countrykk -shot training set
 115510102020100100
Belgium0.8800.8800.8500.8500.8270.8270.7860.7860.6370.637
Estonia0.8570.8570.7400.7400.6910.6910.6290.6290.4500.450
Latvia0.7360.7360.7160.7160.6880.6880.6550.6550.5390.539
Lithuania0.3650.3650.3650.3650.3650.3650.3650.3650.3650.365
Portugal0.7050.7050.5880.5880.5520.5520.4950.4950.3700.370
Table 4: Bhattacharyya distance for fine-tuning countries in EuroCropsML 2.0 , calculated between the fixed test set and the few-shot training sets. The distance is 00 if the two distributions are identical and increases as they become more distinct.
Figure 6: Dirichlet density for K=3K=3 classes, defined over the (K−1)=2(K-1)=2 -simplex. The plots show different concentration parameters α=α⋅1\boldsymbol{\alpha}=\alpha\cdot\boldsymbol{1} , where 1=(1,1,1)⊤\boldsymbol{1}=(1,1,1)^{\top} ( Reuss et al., 2026a ) .
Figure 7: Visualization of four possible draws of the pseudo-prior π~(s)∼Dir⁡(α⋅1)\tilde{\boldsymbol{\pi}}^{(s)}\sim\operatorname{Dir}(\alpha\cdot\boldsymbol{1}) for K=5K=5 classes at consecutive training iterations ss . The orange sampling ( α=0.5\alpha=0.5 ) produces sparse draws, with a different class dominating at each iteration. With α=5\alpha=5 (teal), the draws remain closer to the uniform distribution, indicated by the dashed line. Resampling the pseudo-prior at each iteration exposes the model to a diverse set of imbalance patterns without assuming any fixed test distribution.
Table 5: Few-shot classification results for Belgium . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black . The best-performing PRESTOmodel is highlighted in bold teal .
Table 6: Few-shot classification results for Estonia . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black . The best-performing PRESTOmodel is highlighted in bold teal .
Table 7: Few-shot classification results for Latvia . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black .
Table 8: Few-shot classification results for Lithuania . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black .
Table 9: Few-shot classification results for Portugal . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black .
Table 10: Few-shot classification results for Slovakia . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black .
Table 11: Few-shot classification results for Slovenia . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black . The best-performing PRESTOmodel is highlighted in bold teal .
Table 12: Few-shot classification results for Spain . We show the results for the most-granular crop types, as presented in Table 2 , as well as for the broader land cover classes (HCATlevel 33 ). We report the metrics’ means ±\pm standard deviations on the test set for each variant and few-shot task, computed over five runs, cf. Section 5 . The best result for each few-shot scenario is highlighted in bold blue . Depending on which loss function achieves the best result, the top result of the other loss function is shown in bold black . The best-performing PRESTOmodel is highlighted in bold teal .
Figure 8: Performance gains achieved by applying DirPA, visualized relative to the Bhattacharyya distance ( cf. Table 4 ) and Gini coefficient ( cf. Table 3 ). Circular markers denote performance gains, while crosses represent losses of the DirPAmethod. The marker sizes indicate the magnitude of gains or losses compared to the respective baseline. An increasing DBD_{\text{B}} is equivalent to a decreasing number of kk in the few-shot scenario and represents a bigger prior shift. For Lithuania and Spain, DBD_{\text{B}} stays constant since the relative class distribution remains the same in each few-shot task.
Figure 9: Performance gains on HCATlevel 33 achieved by applying DirPA, visualized relative to the Bhattacharyya distance ( cf. Table 4 ) and Gini coefficient ( cf. Table 3 ). Circular markers denote performance gains, while crosses represent losses of the DirPAmethod. The marker sizes indicate the magnitude of gains or losses in metrics compared to the respective baseline. An increasing DBD_{\text{B}} is equivalent to a decreasing number of kk in the few-shot scenario and represents a bigger prior shift. For Lithuania and Spain, DBD_{\text{B}} stays constant since the relative class distribution remains the same in each few-shot task.
Figure 10: Comparison of predictive output space on the 22 -simplex Δ2\Delta^{2} for a three-class classification (c1,c2,c3)(c_{1},c_{2},c_{3}) . The Softmax function reduces feature information to a single coordinate on the simplex, producing a point estimate. In contrast, the Dirichlet distributionmaps features to a probability density over the simplex, thereby capturing epistemic uncertainty.
Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Node(s)GPUVRAMCPUvCPUsArchitecture
1–3NVIDIA RTX A60004848 GBIntel Xeon1212Ampere
4NVIDIA A100-SXM44040 GBAMD EPYC2424Ampere
Appendix
Table 13: Hardware specifications.
hyperparametervalue / value range
training/tuning epochs100100
tuning trials per setup2424
tuning warmup steps (batches)50005000
early stopping patience (epochs)1515
batch size bb256256
learning rate β\beta\left[{10}^{-4},,{10}^{-2}\right]
Appendix
Table 14: Pre-training hyperparameters. The hyperparameters utilized during pre-training, along with their corresponding valid values.
hyperparametervalue / value range
training/tuning epochs200200
tuning trials per setup †{32,40,72,80}\{32,40,72,80\}
tuning warmup steps (batches)300300
early stopping patience (epochs)2020
batch size bb1616
head learning rate βhead\beta_{\text{head}}\left[{10}^{-6},,{10}^{-2}\right]
Appendix
Table 15: General fine-tuning hyperparameters. The hyperparameters utilized during fine-tuning, along with their corresponding valid values.
hyperparametervalue / value range
DirPAconcentration α\alpha{0.01,0.1,0.5,1.0,2.0,3.0,5.0}\{0.01,0.1,0.5,1.0,2.0,3.0,5.0\}
DirPAscaling factor τ\tau{0.5,1.0,2.0,5.0,10.0}\{0.5,1.0,2.0,5.0,10.0\}
Appendix
Table 16: DirPAfine-tuning hyperparameters. The DirPAhyperparameters utilized during fine-tuning, along with their corresponding valid values.
Table 17: Median of selected optimal values of DirPAhyperparameters with cross-entropy loss, calculated based on finding the optimal hyperparameter values to maximize Cohen’s kappa. The median is calculated over five Optuna studies with different random seeds.
Table 18: Median of selected optimal values of DirPAhyperparameters with focal loss, calculated based on finding the optimal hyperparameter values to maximize Cohen’s kappa. The median is calculated over five Optuna studies with different random seeds.
Table 19: Relative importance of DirPAhyperparameters with cross-entropy loss, calculated based on finding the optimal hyperparameter values to maximize Cohen’s kappa. The values represent the median over five Optuna studies with different random seeds. The relative importance is calculated based on a total of five tunable hyperparameters.
Table 20: Relative importance of DirPAhyperparameters with focal loss, calculated based on finding the optimal hyperparameter values to maximize Cohen’s kappa. The values represent the median over five Optuna studies with different random seeds. Please note that an additional hyperparameter (FLfocus parameter γ\gamma ) was taken into account in these studies. Thus, the relative importance of α\alpha and τ\tau is calculated against a broader set of parameters (six in total) and, therefore, not directly comparable to the ones in Table 19 .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE13.289 857(±8.427 037)13.289\,857(\pm 8.427\,037)11.469 213(±4.363 450)11.469\,213(\pm 4.363\,450)13.612 350(±3.053 609)13.612\,350(\pm 3.053\,609)12.309 223(±2.472 891)12.309\,223(\pm 2.472\,891)19.970 963(±3.569 147)19.970\,963(\pm 3.569\,147)
CEDirPA9.616 047(±0.581 898)9.616\,047(\pm 0.581\,898)10.417 910(±1.413 838)10.417\,910(\pm 1.413\,838)10.119 823(±0.993 733)10.119\,823(\pm 0.993\,733)12.615 223(±2.866 580)12.615\,223(\pm 2.866\,580)17.508 493(±1.727 554)17.508\,493(\pm 1.727\,554)
FL6.726 383(±2.291 953)6.726\,383(\pm 2.291\,953)6.857 050(±2.732 554)6.857\,050(\pm 2.732\,554)10.840 830(±4.093 952)10.840\,830(\pm 4.093\,952)10.562 220(±2.915 440)10.562\,220(\pm 2.915\,440)13.426 667(±5.082 452)13.426\,667(\pm 5.082\,452)
FLDirPA6.672 557(±1.450 298)6.672\,557(\pm 1.450\,298)7.142 610(±3.073 632)7.142\,610(\pm 3.073\,632)9.461 057(±1.558 183)9.461\,057(\pm 1.558\,183)8.670 053(±1.461 380)8.670\,053(\pm 1.461\,380)12.663 447(±1.661 755)12.663\,447(\pm 1.661\,755)
Appendix
Table 21: Fine-tuning runtimes for final training on Belgium . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE15.114 293(±13.133 837)15.114\,293(\pm 13.133\,837)7.352 643(±0.798 790)7.352\,643(\pm 0.798\,790)8.896 243(±1.606 300)8.896\,243(\pm 1.606\,300)9.896 347(±1.570 817)9.896\,347(\pm 1.570\,817)17.153 410(±5.647 102)17.153\,410(\pm 5.647\,102)
CEDirPA9.110 967(±2.644 530)9.110\,967(\pm 2.644\,530)7.040 727(±1.124 475)7.040\,727(\pm 1.124\,475)8.513 653(±1.809 453)8.513\,653(\pm 1.809\,453)8.628 690(±1.207 103)8.628\,690(\pm 1.207\,103)12.403 610(±2.363 247)12.403\,610(\pm 2.363\,247)
FL9.097 793(±2.744 415)9.097\,793(\pm 2.744\,415)47.793 257(±36.073 263)47.793\,257(\pm 36.073\,263)37.295 497(±36.027 621)37.295\,497(\pm 36.027\,621)53.842 527(±34.654 026)53.842\,527(\pm 34.654\,026)31.593 223(±30.398 930)31.593\,223(\pm 30.398\,930)
FLDirPA51.352 183(±32.016 492)51.352\,183(\pm 32.016\,492)52.421 357(±56.770 861)52.421\,357(\pm 56.770\,861)36.577 673(±34.753 552)36.577\,673(\pm 34.753\,552)21.357 673(±31.188 609)21.357\,673(\pm 31.188\,609)19.374 983(±9.001 035)19.374\,983(\pm 9.001\,035)
Appendix
Table 22: Fine-tuning runtimes for final training on Estonia . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE8.044 280(±2.712 212)8.044\,280(\pm 2.712\,212)11.501 870(±6.405 188)11.501\,870(\pm 6.405\,188)8.765 760(±1.987 778)8.765\,760(\pm 1.987\,778)7.654 137(±3.540 995)7.654\,137(\pm 3.540\,995)10.395 847(±1.682 871)10.395\,847(\pm 1.682\,871)
CEDirPA15.541 147(±3.231 581)15.541\,147(\pm 3.231\,581)12.152 577(±3.396 038)12.152\,577(\pm 3.396\,038)13.179 780(±0.185 752)13.179\,780(\pm 0.185\,752)11.130 960(±2.382 089)11.130\,960(\pm 2.382\,089)10.203 390(±5.444 159)10.203\,390(\pm 5.444\,159)
FL14.149 937(±4.741 221)14.149\,937(\pm 4.741\,221)11.469 900(±3.070 765)11.469\,900(\pm 3.070\,765)12.451 517(±4.235 579)12.451\,517(\pm 4.235\,579)13.019 090(±3.546 511)13.019\,090(\pm 3.546\,511)11.540 210(±2.920 380)11.540\,210(\pm 2.920\,380)
FLDirPA12.818 767(±3.097 913)12.818\,767(\pm 3.097\,913)11.385 537(±2.469 036)11.385\,537(\pm 2.469\,036)9.637 880(±2.430 279)9.637\,880(\pm 2.430\,279)12.292 803(±1.662 080)12.292\,803(\pm 1.662\,080)11.855 353(±0.876 824)11.855\,353(\pm 0.876\,824)
Appendix
Table 23: Fine-tuning runtimes for final training on Lithuania . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE9.260 703(±1.906 495)9.260\,703(\pm 1.906\,495)12.164 123(±5.219 105)12.164\,123(\pm 5.219\,105)11.136 137(±1.216 544)11.136\,137(\pm 1.216\,544)13.004 923(±2.372 812)13.004\,923(\pm 2.372\,812)19.428 480(±6.522 352)19.428\,480(\pm 6.522\,352)
CEDirPA9.881 603(±1.103 402)9.881\,603(\pm 1.103\,402)10.352 373(±2.155 403)10.352\,373(\pm 2.155\,403)8.529 727(±2.712 491)8.529\,727(\pm 2.712\,491)10.799 650(±2.958 940)10.799\,650(\pm 2.958\,940)13.654 000(±5.134 761)13.654\,000(\pm 5.134\,761)
FL9.108 660(±0.733 540)9.108\,660(\pm 0.733\,540)8.990 780(±1.093 702)8.990\,780(\pm 1.093\,702)9.304 413(±0.295 043)9.304\,413(\pm 0.295\,043)11.506 287(±4.570 783)11.506\,287(\pm 4.570\,783)17.666 677(±3.567 740)17.666\,677(\pm 3.567\,740)
FLDirPA8.983 157(±1.382 304)8.983\,157(\pm 1.382\,304)8.673 490(±0.917 893)8.673\,490(\pm 0.917\,893)11.096 097(±1.634 672)11.096\,097(\pm 1.634\,672)9.940 173(±2.497 455)9.940\,173(\pm 2.497\,455)15.615 883(±1.249 053)15.615\,883(\pm 1.249\,053)
Appendix
Table 24: Fine-tuning runtimes for final training on Latvia . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE6.106 120(±2.279 087)6.106\,120(\pm 2.279\,087)7.078 273(±1.320 107)7.078\,273(\pm 1.320\,107)7.638 620(±2.143 625)7.638\,620(\pm 2.143\,625)7.073 157(±2.193 797)7.073\,157(\pm 2.193\,797)11.256 137(±6.112 456)11.256\,137(\pm 6.112\,456)
CEDirPA4.985 393(±2.996 504)4.985\,393(\pm 2.996\,504)7.376 027(±2.614 819)7.376\,027(\pm 2.614\,819)6.348 770(±2.135 994)6.348\,770(\pm 2.135\,994)7.485 227(±1.915 216)7.485\,227(\pm 1.915\,216)9.924 320(±4.455 447)9.924\,320(\pm 4.455\,447)
FL21.017 800(±28.339 115)21.017\,800(\pm 28.339\,115)27.894 35(±29.843 895)27.894\,35(\pm 29.843\,895)35.990 027(±38.109 397)35.990\,027(\pm 38.109\,397)49.005 560(±58.671 489)49.005\,560(\pm 58.671\,489)14.716 073(±5.491 213)14.716\,073(\pm 5.491\,213)
FLDirPA7.043 510(±3.451 348)7.043\,510(\pm 3.451\,348)20.804 973(±31.671 793)20.804\,973(\pm 31.671\,793)23.449 157(±26.096 266)23.449\,157(\pm 26.096\,266)34.064 837(±39.974 590)34.064\,837(\pm 39.974\,590)12.887 087(±8.249 299)12.887\,087(\pm 8.249\,299)
Appendix
Table 25: Fine-tuning runtimes for final training on Portugal . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE7.892 167(±0.820 982)7.892\,167(\pm 0.820\,982)9.969 560(±3.554 896)9.969\,560(\pm 3.554\,896)9.700 100(±1.685 056)9.700\,100(\pm 1.685\,056)10.950 930(±2.928 832)10.950\,930(\pm 2.928\,832)16.145 813(±5.758 777)16.145\,813(\pm 5.758\,777)
CEDirPA7.176 353(±1.424 698)7.176\,353(\pm 1.424\,698)7.148 257(±3.316 162)7.148\,257(\pm 3.316\,162)8.040 400(±2.045 451)8.040\,400(\pm 2.045\,451)7.555 813(±2.496 166)7.555\,813(\pm 2.496\,166)14.083 130(±3.668 841)14.083\,130(\pm 3.668\,841)
FL9.434 570(±2.473 063)9.434\,570(\pm 2.473\,063)11.301 977(±2.598 286)11.301\,977(\pm 2.598\,286)12.092 953(±6.063 345)12.092\,953(\pm 6.063\,345)12.717 100(±2.652 926)12.717\,100(\pm 2.652\,926)22.179 613(±15.588 062)22.179\,613(\pm 15.588\,062)
FLDirPA6.637 677(±2.609 201)6.637\,677(\pm 2.609\,201)8.825 837(±4.043 449)8.825\,837(\pm 4.043\,449)9.456 963(±3.426 530)9.456\,963(\pm 3.426\,530)11.444 237(±1.362 571)11.444\,237(\pm 1.362\,571)12.155 097(±2.278 300)12.155\,097(\pm 2.278\,300)
Appendix
Table 26: Fine-tuning runtimes for final training on Slovakia . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE6.782 167(±2.156 214)6.782\,167(\pm 2.156\,214)9.548 483(±0.938 359)9.548\,483(\pm 0.938\,359)11.378 630(±4.238 550)11.378\,630(\pm 4.238\,550)10.645 473(±5.584 847)10.645\,473(\pm 5.584\,847)21.102 253(±6.188 302)21.102\,253(\pm 6.188\,302)
CEDirPA12.463 270(±2.008 205)12.463\,270(\pm 2.008\,205)12.086 047(±1.238 249)12.086\,047(\pm 1.238\,249)11.507 993(±1.089 948)11.507\,993(\pm 1.089\,948)11.931 230(±4.793 913)11.931\,230(\pm 4.793\,913)14.319 293(±4.398 825)14.319\,293(\pm 4.398\,825)
FL8.585 217(±4.065 608)8.585\,217(\pm 4.065\,608)10.547 860(±2.004 268)10.547\,860(\pm 2.004\,268)11.688 130(±2.862 822)11.688\,130(\pm 2.862\,822)14.868 950(±8.515 425)14.868\,950(\pm 8.515\,425)19.446 147(±9.068 916)19.446\,147(\pm 9.068\,916)
FLDirPA10.181 313(±2.305 979)10.181\,313(\pm 2.305\,979)12.295 833(±3.925 255)12.295\,833(\pm 3.925\,255)13.253 137(±2.696 096)13.253\,137(\pm 2.696\,096)12.952 990(±1.425 347)12.952\,990(\pm 1.425\,347)17.279 127(±3.316 301)17.279\,127(\pm 3.316\,301)
Appendix
Table 27: Fine-tuning runtimes for final training on Slovenia . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
algorithm/lossbenchmark task ( kk -shot)
 151020100
CE6.555 870(±0.842 287)6.555\,870(\pm 0.842\,287)6.420 410(±0.587 695)6.420\,410(\pm 0.587\,695)7.218 750(±1.291 122)7.218\,750(\pm 1.291\,122)6.710 627(±1.368 406)6.710\,627(\pm 1.368\,406)8.282 793(±3.141 645)8.282\,793(\pm 3.141\,645)
CEDirPA16.003 783(±6.129 671)16.003\,783(\pm 6.129\,671)31.394 197(±28.310 336)31.394\,197(\pm 28.310\,336)41.817 613(±38.153 626)41.817\,613(\pm 38.153\,626)39.423 680(±32.378 595)39.423\,680(\pm 32.378\,595)52.329 827(±31.291 299)52.329\,827(\pm 31.291\,299)
FL8.398 957(±2.644 772)8.398\,957(\pm 2.644\,772)8.238 637(±2.344 832)8.238\,637(\pm 2.344\,832)6.637 683(±2.244 511)6.637\,683(\pm 2.244\,511)7.226 997(±1.730 941)7.226\,997(\pm 1.730\,941)7.079 583(±1.902 185)7.079\,583(\pm 1.902\,185)
FLDirPA10.218 090(±3.601 519)10.218\,090(\pm 3.601\,519)10.818 533(±3.932 724)10.818\,533(\pm 3.932\,724)11.390 987(±5.333 990)11.390\,987(\pm 5.333\,990)9.201 033(±4.338 430)9.201\,033(\pm 4.338\,430)10.061 707(±4.423 198)10.061\,707(\pm 4.423\,198)
Appendix
Table 28: Fine-tuning runtimes for final training on Spain . We present the average training time (in minutes) as mean ±\pm standard deviation over five runs (seeds). The fastest run time per few-shot scenario is marked in bold .
More in this topic

Class-Imbalanced Learning

Class-imbalanced learning develops classifiers that recognize rare classes without sacrificing performance on common ones.

See all 256 papers

Explore similar work

CardsList
Sep 9, 2026

Are We Really Doing Few-Shot Learning? A Critical Examination of Pre-Training Assumptions

Few-shot learning is commonly evaluated under protocols that pre-train a model on a large auxiliary set whose classes are disjoint from the target episodes yet drawn from the same visual domain. This paper examines whether such protocols truly reflect low-data learning. We systematically compare no pre-training, class-disjoint in-domain pre-training, supervised out-of-domain pre-training, and label-free out-of-domain pre-training across eight datasets, three few-shot architectures, and multiple way-shot settings. Our results show that class disjointness alone is insufficient to remove the influence of target-domain data. In-domain pre-training improves over no pre-training by 33.41 percentage points on average, whereas supervised out-of-domain pre-training yields 23.75 percentage points, revealing a 9.66-point optimistic bias associated with domain overlap. Although out-of-domain pre-training is more realistic in applications where target-domain data are scarce, its effectiveness depends strongly on the compatibility between source and target domains. We further show that labeled source data are not strictly required, with an augmentation-based label-free strategy reaching an average gain of 27.71 percentage points and closely matching supervised out-of-domain pre-training at 27.97 percentage points. Finally, we introduce a descriptor-based source-selection strategy that estimates source-domain suitability before pre-training, reaching a median gap of only 1.37 percentage points to oracle selection. These findings highlight the need to move beyond in-domain pre-training as the default few-shot evaluation protocol, since it can overestimate performance in realistic scenarios where target-domain data are scarce.

Alejandro Galan-Cuenca, Marcelo Saval-Calvo, Antonio Javier Gallego

University Institute for Computer Research, University of Alicante, Alicante, Spain

Domain AdaptationFew-Shot LearningFew-Shot Image Classification
AskPDF
Oct 4, 2026

Fast Convergence through Distributed Augmentation for Class-Imbalanced Federated Learning

In federated learning, mitigating class imbalance is essential to improve minority-class performance. A common approach to address this problem is to augment minority-class samples to achieve local class balance. Existing approaches treat augmentation as a heuristic and do not establish how the amount of augmentation influences the convergence of federated learning, leading to excessive augmentation and increased training time. To address this limitation, we first establish the relationship between augmentation and the convergence behavior of federated learning. Leveraging this insight, we propose DAFL, a distributed augmentation framework that determines the minimum augmentation required for each client-class pair by jointly minimizing augmentation and training time while constraining global class imbalance, thereby improving minority-class F1-score. Experimental results demonstrate that DAFL consistently improves minority-class F1-score while substantially reducing training time, particularly under severe global class imbalance and high label proportion imbalance.

Arathi Nair M, J. Harshan, Anwitaman Datta

Amar Nath and Shashi Khosla School of Information Technology, Indian Institute of Technology Delhi, India · Department of Electrical Engineering, Indian Institute of Technology Delhi, India · College of Computing and Data Science, Nanyang Technological University, Singapore

Class-Imbalanced LearningData AugmentationFederated Learning
AskPDF
Jul 10, 2026

A Strong Balanced-Softmax Classifier-Retraining Baseline for Long-Tailed Recognition

Long-tailed recognition methods often modify losses, margins, or representations to reduce the dominance of frequent classes. We ask whether, after Balanced Softmax training, the remaining tail error can be reduced by retraining only the classifier. We evaluate BS-cRT, a two-stage procedure that trains a backbone and cosine classifier with Balanced Softmax, freezes the backbone, and updates only the classifier on balanced episodic batches. The second stage keeps the empirical-prior Balanced Softmax objective and uses raw cosine logits at inference. Across CIFAR-100-LT, CIFAR-10-LT, ImageNet-LT, and Places-LT, this classifier-only step consistently improves Few-shot accuracy over the matched Balanced Softmax checkpoint. At imbalance factor 100, Few-shot gains are +5.15 points on CIFAR-100-LT and +5.83 on CIFAR-10-LT; on ImageNet-LT and Places-LT, gains are +6.92 and +9.78 points, respectively, with a Top-1/Few-shot trade-off on ImageNet-LT. We also analyze Counterfactual Boundary Risk Minimization (CBRM), a boundary-probe extension using prototype-based features near decision boundaries. CBRM identifies two failure modes: scaled-logit cosine margins destabilize training, and corrected hardest-negative probes remain head-class anchored. The results support BS-cRT as a practical classifier-side baseline and indicate that boundary supervision must account for class frequency.

Juan Terven, Diana Margarita Córdova Esparza, Julio Alejandro Romero Gonzalez +4

CICATA Querétaro Instituto Politecnico Nacional Mexico · Facultad de Informática Universidad Autónoma de Querétaro Mexico

Class-Imbalanced LearningLong-Tailed Classification
AskPDF