CAIM-Net: Change Area Inference Network

• The paper introduces CAIM-Net, which infers change area directly from change moment to ensure consistency between spatial and temporal change predictions.
• It employs a three-stage pipeline—difference extraction, coarse/fine change moment extraction, and CAM-based refinement—to capture detailed spatiotemporal correlations.
• Extensive experiments on DynamicEarthNet and SpaceNet7 demonstrate CAIM-Net's improved Kappa scores and efficiency compared to prior TSCD methods.

CAIM-Net, short for Change Area Inference from Moment Network, is a time series change detection network for remote sensing image sequences that is designed to predict both where change occurred and when it occurred, while enforcing consistency between those outputs by inferring change area from change moment rather than treating the two as fully separate tasks (Li et al., 3 Sep 2025). In the formulation used by the paper, the network comprises three stages—Difference Extraction and Enhancement, Coarse Change Moment Extraction, and Fine Change Moment Extraction and Change Area Inference—and combines a lightweight encoder with batch dimension stacking, boundary enhancement convolution, spatiotemporal correlation analysis, and multiscale temporal CAM (Li et al., 3 Sep 2025).

1. Problem setting and conceptual motivation

The paper studies time series change detection (TSCD) on remote sensing image sequences acquired over the same area at multiple times. TSCD is intended to determine both the change area, meaning the pixels whose state changed at some point in the sequence, and the change moment, meaning between which two consecutive images the change occurred (Li et al., 3 Sep 2025). The labeling convention follows prior work: if a pixel changes multiple times, the assigned label corresponds to the last change moment (Li et al., 3 Sep 2025).

The motivation for CAIM-Net is that prior deep learning methods either perform only change area detection or use multi-task learning to predict change area and change moment with separate branches. The paper argues that such treatment can yield inconsistent outputs, for example when some pixels are predicted as changed but have no corresponding change moment, or vice versa (Li et al., 3 Sep 2025). Its central observation is that pixels with an identified change moment must have undergone change, so change area can be inferred from change moment (Li et al., 3 Sep 2025).

A second motivation concerns the imaging regime. The paper emphasizes that many practical TSCD datasets use medium- or low-resolution time series images, where object boundaries are comparatively ambiguous. This motivates the use of boundary enhancement convolution to sharpen inter-temporal difference features around changed and unchanged regions (Li et al., 3 Sep 2025).

2. Three-stage architecture

CAIM-Net is organized into three stages: Difference Extraction and Enhancement, Coarse Change Moment Extraction, and Fine Change Moment Extraction and Change Area Inference. The raw input tensor is described as having shape

$[T, B, C, H, W],$

where $T$ is the number of time steps, $B$ the batch size, $C$ the channel dimension, and $H, W$ the spatial dimensions (Li et al., 3 Sep 2025).

In the first stage, the network applies batch dimension stacking, reshaping the input from

$[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$

so that all images from all times and all samples are processed by the same encoder in a single pass (Li et al., 3 Sep 2025). The encoder is explicitly described as lightweight and non-downsampling. It has two branches: one branch uses two $3 \times 3$ convolutions with padding 1 and stride 1, and the other uses two $1 \times 1$ convolutions with padding 0 and stride 1. The two branch outputs are summed and passed through another $3 \times 3$ convolution, with each convolution followed by GroupNorm and ReLU. The extracted feature size is

$C \times H \times W,\qquad C = 64,$

and the encoder is written as

$T$0

$T$1

$T$2

(Li et al., 3 Sep 2025)

After restoring the temporal structure, CAIM-Net computes adjacent-frame difference features. The paper writes

$T$3

while also stating that the absolute value is taken in the implementation, i.e. conceptually

$T$4

for $T$5 (Li et al., 3 Sep 2025).

These difference features are then processed by boundary enhancement convolution. The paper describes this operator as differing from ordinary convolution in that it computes the difference between the center pixel and surrounding pixels, in a manner said to be similar in spirit to Local Binary Pattern (LBP). The difference tensor is reshaped from

$T$6

and grouped convolution is used with output channels $T$7 and groups $T$8, so that each channel is processed independently and temporal order is preserved. The extracted boundary information is then added back to the original difference features (Li et al., 3 Sep 2025).

3. Spatiotemporal correlation and change-moment extraction

Before estimating change moments, CAIM-Net performs spatiotemporal correlation analysis. The enhanced difference features are reshaped from

$T$9

and a Transformer encoder is applied for the correlation modeling described in the paper as spatial correlation. This is followed by an LSTM for temporal correlation. The LSTM output has shape

$B$0

and is reshaped back to

$B$1

before being passed to two separate coarse change-moment extractors (Li et al., 3 Sep 2025).

The first extractor reasons from adjacent-frame change/no-change features. For each temporal difference, a convolution and GroupNorm produce a two-channel output:

• class 1: no-change,
• class 2: change.

The paper gives

$B$2

and then forms a $B$3-class coarse change moment by using the minimum no-change feature over time as the no-change class evidence and the per-step change features as the change-moment evidence:

$B$4

where $B$5 denotes SoftMax (Li et al., 3 Sep 2025).

The second extractor treats change-moment identification as a multi-class semantic segmentation problem over the full temporal stack. The temporal and channel dimensions are concatenated,

$B$6

and two $B$7 convolutions, with the first followed by GroupNorm, reduce the representation to $B$8 classes. The paper writes this as

$B$9

while also noting notation inconsistencies in this equation (Li et al., 3 Sep 2025).

Both extractors produce coarse change-moment tensors of shape

$C$0

where class $C$1 denotes no-change and the remaining $C$2 classes denote change between adjacent times (Li et al., 3 Sep 2025).

The refinement stage uses multiscale temporal Class Activation Mapping (CAM). The coarse change-moment outputs are spatially regrouped into coarser scales, for example

$C$3

and

$C$4

For a regrouped scale, the paper applies mean aggregation and a fully connected mapping,

$C$5

and computes CAM scores as

$C$6

followed by bilinear interpolation,

$C$7

Because there are two coarse change-moment outputs and two scales, CAIM-Net generates four CAM outputs, denoted $C$8, $C$9, $H, W$0, and $H, W$1, which are converted into supplementary moment predictions and fused as

$H, W$2

(Li et al., 3 Sep 2025)

4. Change-area inference and optimization

The defining step of CAIM-Net is the conversion from refined change moment to change area. The paper states that a pixel with a predicted change moment must belong to the changed area, so the binary area map is derived directly from the $H, W$3-class moment map rather than decoded independently (Li et al., 3 Sep 2025).

Operationally, the no-change probability in the area output is taken directly from the no-change class of the moment output, while the change probability in the area output is taken as the maximum over all change-moment classes, followed by SoftMax normalization (Li et al., 3 Sep 2025). This means that temporal-spatial consistency is enforced structurally: any pixel assigned substantial probability to any change moment also receives corresponding support in the change-area output.

For supervision, the paper uses Focal Weighted Cross-Entropy Loss (FWCL) for both change-moment identification and change-area detection. It gives

$H, W$4

where $H, W$5 is the total number of samples, $H, W$6 is the ratio of current class samples to all training samples, $H, W$7 is the predicted value, $H, W$8 is the label, and the focusing parameter is set to

$H, W$9

The rationale stated in the paper is that weighted cross-entropy addresses class imbalance and focal loss emphasizes hard samples over easy ones (Li et al., 3 Sep 2025).

The total objective combines supervision of the final fine change moment, the inferred fine change area, and the four supplementary moment outputs. The paper describes the loss as the sum of $[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$0, $[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$1, and the average of $[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$2 (Li et al., 3 Sep 2025).

5. Data, implementation, and empirical results

The evaluation uses two datasets: DynamicEarthNet and SpaceNet7. DynamicEarthNet consists of monthly Sentinel-2 images from January 2018 to December 2019, with 13 bands available, though the paper uses 4 bands at 10 m resolution—Blue, Green, Red, and Near-Infrared. The dataset contains 75 AOIs globally, each with 24 images of size $[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$3. After removing cloudy images, 30 TSI cubes were selected from 55 labeled AOIs, and each cube contains 6 images (Li et al., 3 Sep 2025).

SpaceNet7 consists of monthly Planet imagery from July 2017 to January 2020, with 4 bands—Red, Green, Blue, and Near-Infrared—at 4 m resolution. It contains 100 AOIs globally, with about 24 images per AOI of size around $[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$4. After removing cloudy images, 60 TSI cubes were selected, and each cube contains 9 images (Li et al., 3 Sep 2025).

Both datasets originally provide pixel-wise semantic labels. The paper derives TSCD labels by differencing adjacent semantic labels. A pixel is labeled unchanged in the area map if it is unchanged across all adjacent differences, and changed otherwise. For the moment map, a pixel receives label 0 if unchanged across all adjacent differences, label 1 if changed between image 1 and 2, label 2 if changed between image 2 and 3, and so forth; if multiple changes occur, the last change moment is used (Li et al., 3 Sep 2025).

The input patch size is

$[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$5

For both datasets, the train/validation/test split is

$[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$6

DynamicEarthNet yields 36,864 training images, 4,608 validation images, and 4,608 testing images. SpaceNet7 yields 110,592 training images, 13,824 validation images, and 13,824 testing images (Li et al., 3 Sep 2025).

Training uses Adam with initial learning rate

$[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$7

The batch size is 96 for DynamicEarthNet and 64 for SpaceNet7. Training lasts 100 epochs on DynamicEarthNet and 50 epochs on SpaceNet7. The implementation is in PyTorch and uses an NVIDIA GeForce RTX 3090Ti (Li et al., 3 Sep 2025).

The evaluation metrics are Overall Accuracy (OA), F1-score, Kappa, Precision (Pre), and Recall (Rec). For change-moment identification, F1, Precision, and Recall are computed per moment class and then averaged (Li et al., 3 Sep 2025).

On DynamicEarthNet, for change area detection, CAIM-Net reports:

• OA: 81.22
• F1: 57.58
• Kappa: 45.87
• Pre: 67.95
• Rec: 49.96

The strongest prior method among the compared methods is reported as Multi-RLD-Net, with:

• OA: 80.85
• F1: 56.65
• Kappa: 44.75
• Pre: 67.18
• Rec: 48.98

The Kappa improvement is stated as

$[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$8

For change moment identification on DynamicEarthNet, CAIM-Net reports:

• OA: 76.01
• F1: 43.83
• Kappa: 37.67
• Pre: 53.80
• Rec: 40.99

Multi-RLD-Net reports:

• OA: 74.87
• F1: 41.14
• Kappa: 37.31
• Pre: 45.89
• Rec: 41.87

The Kappa improvement is

$[T, B, C, H, W] \rightarrow [T \times B, C, H, W],$9

(Li et al., 3 Sep 2025)

On SpaceNet7, for change area detection, CAIM-Net reports:

• OA: 97.97
• F1: 52.53
• Kappa: 51.49
• Pre: 55.97
• Rec: 49.48

Multi-RLD-Net reports:

• OA: 97.68
• F1: 50.51
• Kappa: 49.33
• Pre: 49.15
• Rec: 51.96

The Kappa improvement is

$3 \times 3$0

For change moment identification, CAIM-Net reports:

• OA: 97.66
• F1: 40.73
• Kappa: 44.61
• Pre: 44.74
• Rec: 39.57

Multi-RLD-Net reports:

• OA: 97.63
• F1: 41.84
• Kappa: 43.64
• Pre: 47.02
• Rec: 39.83

The Kappa improvement is

$3 \times 3$1

The paper explicitly notes that although CAIM-Net’s moment F1 on SpaceNet7 is lower than Multi-RLD-Net’s, Kappa is more reliable under the extreme class imbalance of SpaceNet7, where changed samples are only about 1% (Li et al., 3 Sep 2025).

6. Ablation evidence, efficiency, limitations, and disambiguation

The ablation study attributes measurable gains to each major component. Comparing Encoder with Boundary, the boundary enhancement convolution improves Kappa from 31.77 to 32.35 for change area and from 23.23 to 24.91 for change moment on DynamicEarthNet, and from 41.05 to 43.52 for change area and from 36.67 to 39.28 for change moment on SpaceNet7 (Li et al., 3 Sep 2025). Adding the spatiotemporal correlation module further improves Kappa from 32.35 to 33.56 for area and from 24.91 to 27.90 for moment on DynamicEarthNet, and from 43.52 to 46.43 for area and from 39.28 to 40.36 for moment on SpaceNet7 (Li et al., 3 Sep 2025).

The two coarse change-moment strategies are presented as complementary. On DynamicEarthNet, Moment1 gives moment Kappa 33.52, Moment2 gives 34.96, and Moment1+2 gives 35.22; on SpaceNet7, the corresponding values are 40.50, 41.70, and 43.13 (Li et al., 3 Sep 2025). The multiscale temporal CAM branches are also individually weaker than the full system. On DynamicEarthNet, the best individual CAM branch gives area Kappa 42.47 and moment Kappa 35.96, whereas full CAIM-Net gives 45.87 and 37.67. On SpaceNet7, the best individual CAM branch gives area Kappa 49.48 and moment Kappa 43.30, whereas full CAIM-Net gives 51.49 and 44.61 (Li et al., 3 Sep 2025).

The encoder design is additionally justified by runtime measurements against a standard Siamese structure. On DynamicEarthNet, training time decreases from 292.83s/epoch to 46.49s/epoch, and inference time from 9.62s to 6.13s. On SpaceNet7, training time decreases from 1281.02s/epoch to 134.09s/epoch, and inference time from 26.80s to 14.40s (Li et al., 3 Sep 2025).

The principal limitation stated by the paper is that CAIM-Net focuses on the last change event and does not characterize all intermediate changes between adjacent time points. The authors accordingly suggest future work should detect and characterize changes between every pair of adjacent images (Li et al., 3 Sep 2025).

The name CAIM-Net also requires disambiguation. In the cited corpus, it refers specifically to the remote-sensing TSCD model Change Area Inference from Moment Network (Li et al., 3 Sep 2025). It should not be conflated with CIM-NET, a CIM-aware video denoising network (Gao et al., 23 May 2025), CiMNet, a joint architecture–hardware co-search framework for compute-in-memory hardware (Kundu et al., 2024), CAIM, a cooperative AoA estimation method using the Ising method (Akbari et al., 2021), or CAIM, a cognitive AI memory framework for long-term interaction with intelligent agents (Westhäußer et al., 19 May 2025).