STFNet—運用多源衛星影像進行森林變化偵測的時空融合網路STFNet: A Spatiotemporal Fusion Network for Forest Change Detection Using Multi-Source Satellite Images

How can high-spatial-resolution bi-temporal images and medium-to-low-resolution time-series multispectral images be integrated to improve forest change detection accuracy while suppressing pseudo changes caused by seasonal phenology?

The study area is Shenmu in northern Shaanxi Province, China (about 38°12′–39°27′N, 109°40′–110°55′E), a highland and mountainous region with high forest coverage. High-resolution data come from the Gaofen-1 (GF-1) satellite, fused into 2 m RGB imagery captured in 2021 and 2022. Time-series data come from ESA Sentinel-2 L2A; images with cloud cover below 30% over July 2021 to July 2022 were selected, with one image synthesized per month for a total of 13 phases. Manually annotated forest-change vectors were converted to binary raster maps, images were cropped into non-overlapping 256×256 patches, yielding 2310 patch-label pairs split into 70% training, 10% validation, and 20% testing.

The authors propose STFNet, which uses a double Siamese CNN (VGG16 backbone) to separately extract multi-level features from high-resolution bi-temporal images and time-series multispectral images. They introduce a cross-attention-based temporal differential feature fusion module (CATFF) that builds spatiotemporal dependencies across time-series images through a joint differential attention mechanism (a differential cross-attention DCA and a common cross-attention CA) to focus on genuine change areas, and a multiresolution contextual differential feature fusion module (MCDF) that fuses multi-source multiresolution differential features through global and local dual-branch channel attention. The decoder reconstructs the mask via transposed convolution and is trained with binary cross-entropy loss. Implemented in PyTorch 2.1.0 with the AdamW optimizer, an initial learning rate of 4×10⁻⁵, a batch size of 4, and 200 epochs. Metrics are precision, recall, F1-score, IoU, and OA.

On the self-built forest change detection dataset, STFNet achieved precision 86.97%, recall 88.34%, F1-score 87.65%, IoU 78.01%, and OA 97.67%, outperforming FC-EF, FC-Diff, FC-Conc, SNUNet, BIT, L-UNet, SGSLN, SEIFNet, and CDMamba, with the F1-score 2.02% higher and IoU 3.14% higher than the best competitor (Table 1). The ablation study (Table 2) shows that both CATFF and MCDF improve all metrics; adding time-series information raised the F1-score from 81.01% (GF only) to 87.65% (Table 3). The band-combination analysis (Table 4) shows the standard false-color combination (B8, B4, B3) performed best, highlighting the importance of the near-infrared band; the temporal-number analysis (Table 5) shows five phases were optimal, with overly dense time series slightly reducing accuracy. The complexity analysis (Table 6) notes the modules increase parameter count and runtime, with CATFF having the larger impact.

Feature extraction and aggregation of time-series multispectral images increase model complexity and parameter count (from a baseline of 42.04M to 62.72M), and computational efficiency needs improvement. When the temporal frequency becomes too dense (for example 7 or 13 phases), detection accuracy slightly declines, indicating a need to balance temporal resolution. The study was validated only in the single Shenmu study area in China with a specific GF-1 and Sentinel-2 data combination, and the GF data are not openly available. The conclusion proposes future work combining memory-and-forgetting mechanisms to improve temporal handling and reduce computation.