Our solution follows a two-stage training strategy combining self-supervised learning and supervised fine-tuning. Since the competition dataset contains multispectral Sentinel imagery with limited labeled samples, we first learned general spectral representations using a SimMIM-style masked image modeling task and then fine-tuned the pretrained encoder for disease classification.
\nThe final pipeline consists of:
\nThe dataset consists of Sentinel imagery containing 12 spectral bands:
\nB1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, and B12.
\nEach band was resized to 256×256 using bilinear interpolation. Band-specific normalization was applied using predefined mean and standard deviation statistics.
\nFor robustness, missing or unreadable bands were replaced with zero-valued arrays.
\nDuring training, only spatial augmentations were used:
\nNo spectral perturbation was applied because preserving physically meaningful spectral information was considered important for multispectral imagery.
\nThe backbone of the solution is Swin Transformer V2 Tiny (swinv2_tiny_window16_256).
\nThe original ImageNet-pretrained model accepts RGB input (3 channels). Since Sentinel imagery contains 12 channels, the patch embedding layer was modified to accept 12-band input.
\nInstead of random initialization, pretrained RGB weights were replicated across all spectral channels. This provided a stable starting point and allowed the model to benefit from ImageNet pretraining despite the different input modality.
\nBefore supervised training, a SimMIM-inspired masked image modeling stage was performed.
\nA random mask ratio of 60% was used. Masked regions were removed from the input image while visible regions remained unchanged.
\nThe encoder was required to reconstruct the missing multispectral content.
\nA lightweight convolutional decoder was attached to the Swin encoder. PixelShuffle was used to recover full image resolution.
\nOnly masked regions contributed to the reconstruction objective.
\nAn L1 reconstruction loss was computed between the reconstructed image and the original image:
\nAfter pretraining, only the encoder weights were retained. The decoder was discarded.
\nThe pretrained encoder weights were loaded into a classification model.
\nA new classification head was attached for the four disease categories:
\nThe dataset exhibited noticeable class imbalance. To address this issue, weighted cross-entropy loss was used:
\nHigher weights were assigned to minority classes to improve their contribution during optimization.
\nThe entire labeled dataset was used for supervised training.
\nFor inference, Test Time Augmentation (TTA) was applied.
\nPredictions were generated from six transformed versions of each image:
\nSoftmax probabilities from all transformations were averaged, and the final class prediction was obtained using argmax.
\nSentinel-2 Multispectral Images
\n→ Band Normalization
\n→ SimMIM Self-Supervised Pretraining
\n→ Swin Transformer Encoder
\n→ Supervised Fine-Tuning
\n→ Class-Weighted Cross-Entropy
\n→ Test Time Augmentation
\n→ Final Prediction
\nThe most important component of the solution was the self-supervised pretraining stage. By leveraging all available imagery through masked image modeling, the Swin encoder learned useful spectral representations before receiving class labels.
\nCombined with class-weighted fine-tuning and test-time augmentation, this approach produced a robust classification pipeline for multispectral crop disease recognition.
", "messages": [ { "id": 3467709, "postDate": "2026-06-07T06:08:09.550Z", "content": "Our solution follows a two-stage training strategy combining self-supervised learning and supervised fine-tuning. Since the competition dataset contains multispectral Sentinel imagery with limited labeled samples, we first learned general spectral representations using a SimMIM-style masked image modeling task and then fine-tuned the pretrained encoder for disease classification.
\nThe final pipeline consists of:
\nThe dataset consists of Sentinel imagery containing 12 spectral bands:
\nB1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, and B12.
\nEach band was resized to 256×256 using bilinear interpolation. Band-specific normalization was applied using predefined mean and standard deviation statistics.
\nFor robustness, missing or unreadable bands were replaced with zero-valued arrays.
\nDuring training, only spatial augmentations were used:
\nNo spectral perturbation was applied because preserving physically meaningful spectral information was considered important for multispectral imagery.
\nThe backbone of the solution is Swin Transformer V2 Tiny (swinv2_tiny_window16_256).
\nThe original ImageNet-pretrained model accepts RGB input (3 channels). Since Sentinel imagery contains 12 channels, the patch embedding layer was modified to accept 12-band input.
\nInstead of random initialization, pretrained RGB weights were replicated across all spectral channels. This provided a stable starting point and allowed the model to benefit from ImageNet pretraining despite the different input modality.
\nBefore supervised training, a SimMIM-inspired masked image modeling stage was performed.
\nA random mask ratio of 60% was used. Masked regions were removed from the input image while visible regions remained unchanged.
\nThe encoder was required to reconstruct the missing multispectral content.
\nA lightweight convolutional decoder was attached to the Swin encoder. PixelShuffle was used to recover full image resolution.
\nOnly masked regions contributed to the reconstruction objective.
\nAn L1 reconstruction loss was computed between the reconstructed image and the original image:
\nAfter pretraining, only the encoder weights were retained. The decoder was discarded.
\nThe pretrained encoder weights were loaded into a classification model.
\nA new classification head was attached for the four disease categories:
\nThe dataset exhibited noticeable class imbalance. To address this issue, weighted cross-entropy loss was used:
\nHigher weights were assigned to minority classes to improve their contribution during optimization.
\nThe entire labeled dataset was used for supervised training.
\nFor inference, Test Time Augmentation (TTA) was applied.
\nPredictions were generated from six transformed versions of each image:
\nSoftmax probabilities from all transformations were averaged, and the final class prediction was obtained using argmax.
\nSentinel-2 Multispectral Images
\n→ Band Normalization
\n→ SimMIM Self-Supervised Pretraining
\n→ Swin Transformer Encoder
\n→ Supervised Fine-Tuning
\n→ Class-Weighted Cross-Entropy
\n→ Test Time Augmentation
\n→ Final Prediction
\nThe most important component of the solution was the self-supervised pretraining stage. By leveraging all available imagery through masked image modeling, the Swin encoder learned useful spectral representations before receiving class labels.
\nCombined with class-weighted fine-tuning and test-time augmentation, this approach produced a robust classification pipeline for multispectral crop disease recognition.
", "rawMarkdown": "## Overview\n\nOur solution follows a two-stage training strategy combining self-supervised learning and supervised fine-tuning. Since the competition dataset contains multispectral Sentinel imagery with limited labeled samples, we first learned general spectral representations using a SimMIM-style masked image modeling task and then fine-tuned the pretrained encoder for disease classification.\n\nThe final pipeline consists of:\n\n1. Multispectral preprocessing and normalization\n2. Self-supervised pretraining with SimMIM\n3. Supervised fine-tuning using class-weighted cross-entropy\n4. Test-time augmentation (TTA) during inference\n\n---\n\n## Data Preparation\n\nThe dataset consists of Sentinel imagery containing 12 spectral bands:\n\nB1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, and B12.\n\nEach band was resized to 256×256 using bilinear interpolation. Band-specific normalization was applied using predefined mean and standard deviation statistics.\n\nFor robustness, missing or unreadable bands were replaced with zero-valued arrays.\n\nDuring training, only spatial augmentations were used:\n\n* Horizontal Flip\n* Vertical Flip\n* Random Rotation\n\nNo spectral perturbation was applied because preserving physically meaningful spectral information was considered important for multispectral imagery.\n\n---\n\n## Model Architecture\n\nThe backbone of the solution is Swin Transformer V2 Tiny (swinv2_tiny_window16_256).\n\nThe original ImageNet-pretrained model accepts RGB input (3 channels). Since Sentinel imagery contains 12 channels, the patch embedding layer was modified to accept 12-band input.\n\nInstead of random initialization, pretrained RGB weights were replicated across all spectral channels. This provided a stable starting point and allowed the model to benefit from ImageNet pretraining despite the different input modality.\n\n---\n\n## Self-Supervised Learning (SimMIM)\n\nBefore supervised training, a SimMIM-inspired masked image modeling stage was performed.\n\n### Masking Strategy\n\nA random mask ratio of 60% was used. Masked regions were removed from the input image while visible regions remained unchanged.\n\nThe encoder was required to reconstruct the missing multispectral content.\n\n### Reconstruction Head\n\nA lightweight convolutional decoder was attached to the Swin encoder. PixelShuffle was used to recover full image resolution.\n\n### Loss Function\n\nOnly masked regions contributed to the reconstruction objective.\n\nAn L1 reconstruction loss was computed between the reconstructed image and the original image:\n\n* Loss evaluated only on masked pixels\n* Visible pixels ignored during optimization\n\n### Training Configuration\n\n* Optimizer: AdamW\n* Learning Rate: 5e-4\n* Epochs: 30\n\nAfter pretraining, only the encoder weights were retained. The decoder was discarded.\n\n---\n\n## Supervised Fine-Tuning\n\nThe pretrained encoder weights were loaded into a classification model.\n\nA new classification head was attached for the four disease categories:\n\n* Aphid\n* Blast\n* RPH\n* Rust\n\n### Class Imbalance Handling\n\nThe dataset exhibited noticeable class imbalance. To address this issue, weighted cross-entropy loss was used:\n\n* Aphid: 1.0\n* Blast: 4.0\n* RPH: 0.6\n* Rust: 12.0\n\nHigher weights were assigned to minority classes to improve their contribution during optimization.\n\n### Training Configuration\n\n* Optimizer: AdamW\n* Learning Rate: 1e-4\n* Scheduler: Cosine Annealing\n* Epochs: 20\n\nThe entire labeled dataset was used for supervised training.\n\n---\n\n## Inference\n\nFor inference, Test Time Augmentation (TTA) was applied.\n\nPredictions were generated from six transformed versions of each image:\n\n1. Original image\n2. Horizontal flip\n3. Vertical flip\n4. 90° rotation\n5. 180° rotation\n6. 270° rotation\n\nSoftmax probabilities from all transformations were averaged, and the final class prediction was obtained using argmax.\n\n---\n\n## Final Pipeline\n\nSentinel-2 Multispectral Images\n\n→ Band Normalization\n\n→ SimMIM Self-Supervised Pretraining\n\n→ Swin Transformer Encoder\n\n→ Supervised Fine-Tuning\n\n→ Class-Weighted Cross-Entropy\n\n→ Test Time Augmentation\n\n→ Final Prediction\n\n---\n\n## Conclusion\n\nThe most important component of the solution was the self-supervised pretraining stage. By leveraging all available imagery through masked image modeling, the Swin encoder learned useful spectral representations before receiving class labels.\n\nCombined with class-weighted fine-tuning and test-time augmentation, this approach produced a robust classification pipeline for multispectral crop disease recognition.\n", "votes": 1 } ], "comments": [], "raw_markdown_by_id": { "3467709": "## Overview\n\nOur solution follows a two-stage training strategy combining self-supervised learning and supervised fine-tuning. Since the competition dataset contains multispectral Sentinel imagery with limited labeled samples, we first learned general spectral representations using a SimMIM-style masked image modeling task and then fine-tuned the pretrained encoder for disease classification.\n\nThe final pipeline consists of:\n\n1. Multispectral preprocessing and normalization\n2. Self-supervised pretraining with SimMIM\n3. Supervised fine-tuning using class-weighted cross-entropy\n4. Test-time augmentation (TTA) during inference\n\n---\n\n## Data Preparation\n\nThe dataset consists of Sentinel imagery containing 12 spectral bands:\n\nB1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, and B12.\n\nEach band was resized to 256×256 using bilinear interpolation. Band-specific normalization was applied using predefined mean and standard deviation statistics.\n\nFor robustness, missing or unreadable bands were replaced with zero-valued arrays.\n\nDuring training, only spatial augmentations were used:\n\n* Horizontal Flip\n* Vertical Flip\n* Random Rotation\n\nNo spectral perturbation was applied because preserving physically meaningful spectral information was considered important for multispectral imagery.\n\n---\n\n## Model Architecture\n\nThe backbone of the solution is Swin Transformer V2 Tiny (swinv2_tiny_window16_256).\n\nThe original ImageNet-pretrained model accepts RGB input (3 channels). Since Sentinel imagery contains 12 channels, the patch embedding layer was modified to accept 12-band input.\n\nInstead of random initialization, pretrained RGB weights were replicated across all spectral channels. This provided a stable starting point and allowed the model to benefit from ImageNet pretraining despite the different input modality.\n\n---\n\n## Self-Supervised Learning (SimMIM)\n\nBefore supervised training, a SimMIM-inspired masked image modeling stage was performed.\n\n### Masking Strategy\n\nA random mask ratio of 60% was used. Masked regions were removed from the input image while visible regions remained unchanged.\n\nThe encoder was required to reconstruct the missing multispectral content.\n\n### Reconstruction Head\n\nA lightweight convolutional decoder was attached to the Swin encoder. PixelShuffle was used to recover full image resolution.\n\n### Loss Function\n\nOnly masked regions contributed to the reconstruction objective.\n\nAn L1 reconstruction loss was computed between the reconstructed image and the original image:\n\n* Loss evaluated only on masked pixels\n* Visible pixels ignored during optimization\n\n### Training Configuration\n\n* Optimizer: AdamW\n* Learning Rate: 5e-4\n* Epochs: 30\n\nAfter pretraining, only the encoder weights were retained. The decoder was discarded.\n\n---\n\n## Supervised Fine-Tuning\n\nThe pretrained encoder weights were loaded into a classification model.\n\nA new classification head was attached for the four disease categories:\n\n* Aphid\n* Blast\n* RPH\n* Rust\n\n### Class Imbalance Handling\n\nThe dataset exhibited noticeable class imbalance. To address this issue, weighted cross-entropy loss was used:\n\n* Aphid: 1.0\n* Blast: 4.0\n* RPH: 0.6\n* Rust: 12.0\n\nHigher weights were assigned to minority classes to improve their contribution during optimization.\n\n### Training Configuration\n\n* Optimizer: AdamW\n* Learning Rate: 1e-4\n* Scheduler: Cosine Annealing\n* Epochs: 20\n\nThe entire labeled dataset was used for supervised training.\n\n---\n\n## Inference\n\nFor inference, Test Time Augmentation (TTA) was applied.\n\nPredictions were generated from six transformed versions of each image:\n\n1. Original image\n2. Horizontal flip\n3. Vertical flip\n4. 90° rotation\n5. 180° rotation\n6. 270° rotation\n\nSoftmax probabilities from all transformations were averaged, and the final class prediction was obtained using argmax.\n\n---\n\n## Final Pipeline\n\nSentinel-2 Multispectral Images\n\n→ Band Normalization\n\n→ SimMIM Self-Supervised Pretraining\n\n→ Swin Transformer Encoder\n\n→ Supervised Fine-Tuning\n\n→ Class-Weighted Cross-Entropy\n\n→ Test Time Augmentation\n\n→ Final Prediction\n\n---\n\n## Conclusion\n\nThe most important component of the solution was the self-supervised pretraining stage. By leveraging all available imagery through masked image modeling, the Swin encoder learned useful spectral representations before receiving class labels.\n\nCombined with class-weighted fine-tuning and test-time augmentation, this approach produced a robust classification pipeline for multispectral crop disease recognition.\n" } }