Upload model weights

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+images/arch.jpg filter=lfs diff=lfs merge=lfs -text
+images/ls4s_qual.jpg filter=lfs diff=lfs merge=lfs -text
+images/spider_gb.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,179 @@
+# TerraFM: A Scalable Foundation Model for Unified Multisensor Earth Observation
+<p align="center">
+    <img src="https://i.imgur.com/waxVImv.png" alt="Oryx TerraFM">
+</p>
+
+#### [Muhammad Sohail Danish](https://www.linkedin.com/in/muhammad-sohail-danish/), [Muhammad Akhtar Munir](https://akhtarvision.github.io/), [Syed Roshaan Ali Shah](https://www.linkedin.com/in/syed-roshaan-ali-shah-b797b44a/), [Muhammad Haris Khan](https://www.linkedin.com/in/kartik-kuckreja-930531221/), [Rao Muhammad Anwer](https://research.ibm.com/people/paolo-fraccaro) , [Jorma Laaksonen](https://www.servicenow.com/research/author/alexandre-lacoste.html), [Fahad Shahbaz Khan](https://sites.google.com/view/fahadkhans/home), and [Salman Khan](https://salman-h-khan.github.io/)
+
+
+#### **Mohamed bin Zayed University of AI, University College London, Aalto University, Linköping University, Australian National University**
+
+[![paper](https://img.shields.io/badge/arXiv-Paper-<COLOR>.svg)]()
+[![Model Zoo](https://img.shields.io/badge/Model%20Zoo-HuggingFace-blue)](#🧠-model-zoo)
+
+---
+
+## 📢 Latest Updates
+- **Jun-09-25**: 🚀 Initial release of **TerraFM codebase** and **pretrained models**
+- **Jun-09-25**: 📄 Paper released on arXiv: [arxiv link](). 🔥🔥
+
+---
+
+## 🌍 Overview
+
+**TerraFM** is a scalable foundation model designed for unified processing of multisensor Earth Observation (EO) data. Built on a ViT backbone and trained over **18.7M tiles (~23T pixels)** from Sentinel-1 SAR and Sentinel-2 optical imagery, TerraFM unifies modality-specific inputs using:
+
+- 🧩 Modality-specific patch embeddings  
+- 🌀 Adaptive cross-attention fusion  
+- 🎯 Dual-centering regularization for long-tailed distributions
+
+TerraFM sets a new benchmark on **GEO-Bench** and **Copernicus-Bench**, demonstrating strong generalization across geographies, modalities, and tasks — including classification, segmentation, and landslide detection.
+
+---
+
+
+## 🔬 Key Features
+
+<p align="center">
+  <img src="images/spider_gb.jpg" alt="TerraFM Architecture" width="500"/>
+</p>
+
+- **Multimodal Pretraining**: Uses Sentinel-1 (SAR) and Sentinel-2 (L1C, L2A) as natural augmentations.
+- **Large-Scale Dataset**: Trained on 18.7M global tiles from the [Major-TOM](https://huggingface.co/Major-TOM) dataset.
+- **Cross-Attention Fusion**: Dynamically aggregates information across sensors at patch level.
+- **Dual-Centering**: Mitigates long-tailed land cover bias using ESA WorldCover statistics.
+- **Benchmark SOTA**: Outperforms prior FMs (Galileo, Prithvi, DOFA) across multiple EO tasks.
+
+---
+## 🧱 Architecture
+
+<p align="center">
+  <img src="images/arch.jpg" alt="TerraFM Architecture" width="700"/>
+</p>
+
+Overall architecture of TerraFM. It unifies student-teacher contrastive framework with modality augmentation with cross-attention fusion, and a new dual centering regularization. TerraFM is founded on ViT backbone and is trained on 18.7M globally distributed samples for pre-training and utilizes large-tile inputs for encoding broader spatial context. For illustration, RGB channels from S2-L2A and S2-L1C are selected, and S1 is visualized using a false-color RGB composite.
+
+---
+## 🧠 Model Zoo
+
+| Model | Modality | Input Size | Backbone | Link |
+|-------|----------|------------|--------|------|
+| TerraFM-B | Sentinel-1 RTC + Sentinel-2 Level 2A + Sentinel-2 Level 1C | 224×224 | ViT-Base | [Download](https://huggingface.co/MBZUAI/TerraFM) |
+| TerraFM-L | Sentinel-1 RTC + Sentinel-2 Level 2A + Sentinel-2 Level 1C | 224×224 | ViT-Large | [Download](https://huggingface.co/MBZUAI/TerraFM) |
+
+---
+
+## 🛠 Usage
+
+TerraFM can be used directly via the `terrafm.py` module, which provides standalone implementations of the TerraFM-Base and TerraFM-Large models for easy integration into any codebase.
+
+```python
+from terrafm import terrafm_base, terrafm_large
+import torch
+
+# Simulated input: 1 sample, 12 channels, 224×224 resolution (e.g., Sentinel-2 L2A)
+x = torch.randn(1, 12, 224, 224)
+
+# Load TerraFM-Base model
+model = terrafm_base()
+
+# Load pretrained weights (e.g., TerraFM-B.pth)
+state_dict = torch.load("TerraFM-B.pth", map_location="cpu")
+msg = model.load_state_dict(state_dict, strict=False)
+
+# Forward pass
+y = model(x)
+print(f"Output shape: {y.shape}")
+```
+---
+
+
+## 📊 Results
+
+### 🔍 k-NN Classification Results
+
+We evaluate image classification using k-nearest neighbors (kNN) and report Top-1 accuracy for all single-label tasks. For the multilabel BigEarthNet benchmark, we report the F1 score.
+
+| Model           | Backbone   | m-EuroSat (100%) | m-EuroSat (1%) | m-BigEarthNet (100%) | m-BigEarthNet (1%) | m-So2Sat (100%) | m-So2Sat (1%) | m-Brick-Kiln (100%) | m-Brick-Kiln (1%) |
+|----------------|------------|------------------|----------------|------------------------|--------------------|------------------|----------------|----------------------|--------------------|
+| SatMAE         | ViT-Base   | 84.1             | 34.8           | 50.6                   | 29.0               | 36.0             | 23.1           | 86.1                 | 73.5               |
+| SatMAE++       | ViT-Large  | 82.7             | 48.5           | 50.8                   | 31.6               | 34.7             | 23.4           | 89.6                 | 76.7               |
+| CROMA          | ViT-Base   | 85.6             | 51.3           | 58.8                   | 44.7               | 48.8             | 33.8           | 92.6                 | 85.1               |
+| SoftCon        | ViT-Small  | 89.8             | 27.2           | 64.7                   | 43.3               | 51.1             | 31.4           | 89.2                 | 77.8               |
+| DOFA           | ViT-Base   | 82.8             | 49.6           | 49.4                   | 29.9               | 41.4             | 29.4           | 88.3                 | 78.3               |
+| Satlas         | Swin-Tiny  | 81.7             | 35.8           | 51.9                   | 29.6               | 36.6             | 27.1           | 88.2                 | 73.0               |
+| MMEarth        | CNN-atto   | 81.7             | 30.0           | 58.3                   | 39.6               | 39.8             | 25.1           | 89.4                 | 79.7               |
+| DeCUR          | ViT-Small  | 89.0             | 46.6           | 63.8                   | 49.6               | 45.8             | 30.9           | 83.7                 | 74.2               |
+| AnySat         | ViT-Base   | 82.2             | 47.1           | 54.9                   | 33.7               | 39.8             | 29.0           | 85.3                 | 72.0               |
+| Galileo        | ViT-Base   | 93.0             | 56.6           | 59.0                   | 36.5               | 54.8             | **43.2**       | 90.7                 | 78.0               |
+| Prithvi-2.0    | ViT-Large  | 80.2             | 48.0           | 49.4                   | 28.8               | 29.5             | 26.1           | 87.9                 | 80.6               |
+| Copernicus-FM  | ViT-Base   | 76.0             | 47.4           | 53.8                   | 33.3               | 38.4             | 23.3           | 93.0                 | 83.2               |
+| **TerraFM**    | ViT-Base   | _94.2_           | _59.3_         | _68.7_                 | 49.4               | _55.1_           | _41.6_         | **94.5**             | **85.6**           |
+|**TerraFM**| ViT-Large  | **95.1**         | **62.1**       | **69.4**               | **50.6**           | **55.9**         | 41.1           | _93.0_               | 82.2               |
+
+
+### 🛰 Copernicus-Bench
+
+Comparison of TerraFM with existing supervised and self-supervised methods on **Copernicus-Bench**.  
+Metrics include **OA** (Overall Accuracy), **mAP** (mean Average Precision), and **mIoU** (mean Intersection over Union).
+
+| Dataset         | Metric | Supervised | Random | SoftCon | CROMA | DOFA | Copernicus-FM | **TerraFM** |
+|----------------|--------|------------|--------|---------|--------|------|----------------|-------------|
+| **Backbone**    | --     | ViT-B/16   | ViT-B/16 | ViT-B/14 | ViT-B/8 | ViT-B/16 | ViT-B/16      | ViT-B/16    |
+| **Cloud-S2**       | mIoU  | 59.4       | 60.4   | 66.9    | 65.0   | 65.0 | 66.7          | **67.9**    |
+| **EuroSAT-S1**     | OA    | 81.5       | 75.4   | 83.6    | 83.9   | 81.7 | 87.2          | **87.8**    |
+| **EuroSAT-S2**     | OA    | 97.6       | 92.5   | 96.7    | 97.0   | 97.2 | 97.9          | **99.1**    |
+| **BigEarthNet-S1** | mAP   | 70.6       | 63.8   | **78.7**| 70.8   | 70.5 | 77.9          | 76.9        |
+| **BigEarthNet-S2** | mAP   | 80.1       | 71.6   | 83.6    | 76.4   | 75.5 | 79.0          | **84.4**    |
+| **DFC2020-S1**     | mIoU  | 50.8       | 45.4   | 52.8    | 52.7   | 49.7 | 52.4          | **55.4**    |
+| **DFC2020-S2**     | mIoU  | 66.2       | 62.3   | 64.1    | **66.5**| 61.8 | 64.5          | 63.8        |
+| **LCZ-S2**         | OA    | 85.3       | 77.4   | 83.6    | 84.1   | 83.0 | 84.4          | **87.0**    |
+
+### 🧪 GEO-Bench Performance
+
+Performance comparison on GEO-Bench for both **classification** (Top-1 Accuracy), **segmentation** (mIoU), and **F1 score** (for m-BigEarthNet).  
+TerraFM achieves state-of-the-art results across multiple datasets, outperforming previous foundation models.
+
+| Method       | Backbone   | m-EuroSat | m-BigEarthNet | m-So2Sat | m-Brick-Kiln | m-Cashew-Plant | m-SA-Crop-Type |
+|--------------|------------|-----------|----------------|----------|----------------|------------------|------------------|
+| SatMAE       | ViT-Large  | 96.6      | 68.3           | 57.2     | 98.4           | 30.8             | 24.8             |
+| SatMAE++     | ViT-Large  | 96.5      | 67.9           | 56.0     | 98.6           | 29.6             | 25.7             |
+| CROMA        | ViT-Large  | 96.6      | 71.9           | 60.6     | 98.7           | 31.8             | 32.0             |
+| SoftCon      | ViT-Base   | 97.5      | 70.3           | 61.7     | 98.7           | 29.6             | 30.8             |
+| DOFA         | ViT-Large  | 96.9      | 68.0           | 58.7     | 98.6           | 27.7             | 25.4             |
+| Satlas       | Swin-Base  | 97.5      | 72.8           | 61.9     | **98.9**       | 25.1             | 23.4             |
+| MMEarth      | CNN-atto   | 95.7      | 70.0           | 57.2     | 98.9           | 24.2             | 22.2             |
+| DeCUR        | ViT-Small  | 97.9      | 70.9           | 61.7     | 98.7           | 26.2             | 21.5             |
+| Prithvi 2.0  | ViT-Large  | 96.5      | 69.0           | 54.6     | 98.6           | 26.7             | 22.9             |
+| AnySat       | ViT-Base   | 95.9      | 70.3           | 51.8     | 98.6           | 26.1             | 27.1             |
+| Galileo      | ViT-Base   | 97.7      | 70.7           | 63.3     | 98.7           | 33.0             | 30.1             |
+| **TerraFM**  | ViT-Base   | *98.1*    | 72.6           | *64.9*   | 98.7           | *34.1*           | *33.0*           |
+| **TerraFM**  | ViT-Large  | **98.6**  | **73.1**       | **66.6** | **99.0**       | **37.2**         | **34.5**         |
+
+
+### 🌋 Landslide Detection (Landslide4Sense)
+
+Landslide detection performance on the **Landslide4Sense** test set.  
+Despite having significantly fewer parameters (120M vs. 300M), **TerraFM** achieves higher overall segmentation performance, especially for landslide regions.
+| Model                  | mIoU | IoU (Landslide) |
+|------------------------|------|-----------------|
+| Prithvi-EO-2.0 (300M)  | 65.0 | 31.5            |
+| **TerraFM (120M)**     | **70.8** | **43.1**     |
+
+<p align="center">
+  <img src="images/ls4s_qual.jpg" alt="Landslide Detection" width="700"/>
+</p>
+---
+
+## 📜 Citation
+If you find our work and this repository useful, please consider giving our repo a star and citing our paper as follows:
+```bibtex
+@article{
+}
+```
+
+
+
+
+## 📨 Contact
+If you have any questions, please create an issue on this repository or contact at [email protected].

TerraFM-B.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bdb6a4cfd707a09d79f4790119daa9d76bb316b6ea756c6eb305f99d1797a06
+size 451862010

terrafm.py

@@ -0,0 +1,374 @@
+# ------------------------------------------------------------------------------
+# This file includes code copied and adapted from DINO:
+# - DINO (https://github.com/facebookresearch/dino)
+#
+# ------------------------------------------------------------------------------
+import random
+import math
+import torch
+import torch.nn as nn
+from torch import Tensor
+from functools import partial
+
+
+def make_2tuple(x):
+    if isinstance(x, tuple):
+        assert len(x) == 2
+        return x
+
+    assert isinstance(x, int)
+    return (x, x)
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                      "The distribution of values may be incorrect.",
+                      stacklevel=2)
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    # type: (Tensor, float, float, float, float) -> Tensor
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+def drop_path(x, drop_prob: float = 0., training: bool = False):
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn
+
+
+class Block(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x, return_attention=False):
+        y, attn = self.attn(self.norm1(x))
+        if return_attention:
+            return attn
+        x = x + self.drop_path(y)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+
+class PatchEmbed(nn.Module):
+    def __init__(
+            self, 
+            img_size: int,
+            embed_dim: int,
+            patch_size: int,
+            in_chans_s1: int,
+            in_chans_s2: int,
+        ):
+        super().__init__()
+        attn_dim = embed_dim*3 # from Panopticon design
+        self.img_size = img_size
+        self.patch_size = patch_size
+        num_patches = (img_size // patch_size) * (img_size // patch_size)
+        self.num_patches = num_patches
+        
+        self.conv2d_s2_l2a = nn.Conv2d(in_chans_s2, attn_dim, kernel_size=patch_size, stride=patch_size)
+        self.conv2d_s2_l1c = nn.Conv2d(in_chans_s2, attn_dim, kernel_size=patch_size, stride=patch_size)
+        self.conv2d_s1 = nn.Conv2d(in_chans_s1, attn_dim, kernel_size=patch_size, stride=patch_size)
+        
+        
+        self.projection = TokenProjection(embed_dim=embed_dim, attn_dim=attn_dim)
+        self.s2_l2a_embed = nn.Parameter(torch.zeros(1, attn_dim))
+        self.s2_l1c_embed = nn.Parameter(torch.zeros(1, attn_dim))
+        self.s1_embed = nn.Parameter(torch.zeros(1, attn_dim))
+        self.attn_dim = attn_dim
+
+    def forward(self, x12: Tensor, is_l2a: bool = False) -> Tensor:
+
+        B,C,W,H = x12.shape
+        device, dtype = x12.device, x12.dtype
+        B = len(x12)
+        if C == 2:
+            x = self.conv2d_s1(x12).flatten(2).transpose(1, 2)
+            x += self.s1_embed
+        elif is_l2a:
+            x = self.conv2d_s2_l2a(x12).flatten(2).transpose(1, 2)
+            x += self.s2_l2a_embed
+        else:  
+            x = self.conv2d_s2_l1c(x12).flatten(2).transpose(1, 2)
+            x += self.s2_l1c_embed   
+
+        x = self.projection(x) 
+        return x         
+        
+
+class TokenProjection(nn.Module):
+    def __init__(self, embed_dim: int, attn_dim: int):
+        super().__init__()
+        self.proj1 = nn.Linear(attn_dim, attn_dim, bias=False)
+        self.norm_input = nn.LayerNorm(attn_dim)
+        self.proj2 = nn.Linear(attn_dim, attn_dim)
+        self.proj3 = nn.Linear(attn_dim, embed_dim)
+    
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        Applies a sequence of linear projections used for Case 1 & N in modality augmentation.
+
+        Steps:
+        1. proj1 is shared between Case 1 and Case N (acts like value projection in attention).
+        2. Applies LayerNorm to stabilize training and normalize features.
+        3. In Case N, proj2 is applied after the weighted mean operation.
+        4. proj3 projects to the final embedding dimension.
+        Args:
+            tokens (Tensor): Input tensor of shape [B, N, input_dim], where
+                             B = batch size, N = number of tokens.
+
+        Returns:
+            Tensor: Projected output of shape [B, N, final_dim].
+        """
+        x = self.proj1(x) #V in corss attn
+        x = self.norm_input(x) 
+        x = self.proj2(x)
+        x = self.proj3(x) #final projection
+        return x
+
+class TerraFM(nn.Module):
+    def __init__(self, img_size=[224], patch_size=16, in_chans=3, num_classes=0, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., norm_layer=nn.LayerNorm, **kwargs):
+        super().__init__()
+        self.num_features = self.embed_dim = embed_dim
+        
+        self.patch_embed = PatchEmbed(
+                img_size=img_size[0], patch_size=patch_size, in_chans_s1=2, in_chans_s2=12, embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
+            for i in range(depth)])
+        self.norm = norm_layer(embed_dim)
+
+        # Classifier head
+        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+        trunc_normal_(self.pos_embed, std=.02)
+        trunc_normal_(self.cls_token, std=.02)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def interpolate_pos_encoding(self, x, w, h):
+        npatch = x.shape[1] - 1
+        N = self.pos_embed.shape[1] - 1
+        if npatch == N and w == h:
+            return self.pos_embed
+        class_pos_embed = self.pos_embed[:, 0]
+        patch_pos_embed = self.pos_embed[:, 1:]
+        dim = x.shape[-1]
+        w0 = w // self.patch_embed.patch_size
+        h0 = h // self.patch_embed.patch_size
+        # we add a small number to avoid floating point error in the interpolation
+        # see discussion at https://github.com/facebookresearch/dino/issues/8
+        w0, h0 = w0 + 0.1, h0 + 0.1
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),
+            scale_factor=(w0 / math.sqrt(N), h0 / math.sqrt(N)),
+            mode='bicubic',
+        )
+        assert int(w0) == patch_pos_embed.shape[-2] and int(h0) == patch_pos_embed.shape[-1]
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+
+    def prepare_tokens(self, x):
+        B, nc, w, h = x.shape
+        x = self.patch_embed(x)  # patch linear embedding
+
+        # add the [CLS] token to the embed patch tokens
+        cls_tokens = self.cls_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        # add positional encoding to each token
+        x = x + self.interpolate_pos_encoding(x, w, h)
+
+        return self.pos_drop(x)
+
+    def forward_features(self, x):
+        return self.forward(x)
+
+    def forward(self, x):
+        x = self.prepare_tokens(x)
+        for blk in self.blocks:
+            x = blk(x)
+        x = self.norm(x)
+        return x[:, 0]
+
+    def get_last_selfattention(self, x):
+        x = self.prepare_tokens(x)
+        for i, blk in enumerate(self.blocks):
+            if i < len(self.blocks) - 1:
+                x = blk(x)
+            else:
+                # return attention of the last block
+                return blk(x, return_attention=True)
+
+    def get_intermediate_layers(self, x, n=1,
+        return_class_token = False,
+        norm=False,                        
+        ):
+        x = self.prepare_tokens(x)
+        # we return the output tokens from the `n` last blocks
+        output = []
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if len(self.blocks) - i <= n:
+                output.append(x)
+                # output.append(self.norm(x))
+        if norm:
+            output = [self.norm(out) for out in output]
+        class_tokens = [out[:, 0] for out in output]
+        output = [out[:, 1:] for out in output]
+        if return_class_token:
+            return tuple(zip(output, class_tokens))
+        return output
+    
+    def extract_feature(self, images, return_h_w=True, out_indices=[3, 5, 7, 11]):
+        x = self.prepare_tokens(images)
+        output = []
+        h, w = int(images.shape[2] / self.patch_embed.patch_size), int(images.shape[3] / self.patch_embed.patch_size)
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if i in out_indices:
+                out = x[:, 1:]
+                out = self.norm(out)
+                B, _, C = out.shape
+                out = (
+                    out.reshape(B, h, w, C)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                output.append(out)
+
+        return output
+
+
+
+
+def terrafm_base(patch_size=16, **kwargs):
+    model = TerraFM(
+        patch_size=patch_size, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4,
+        qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
+    return model
+
+def terrafm_large(patch_size=16, **kwargs):
+    model = TerraFM(
+        patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True,
+         norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
+    return model