ConvFormer-Lite: A Lightweight CNN–Transformer Hybrid for Haze/Nonhaze Classification on a Pan-Asian Sentinel-2 True Color Haze Benchmark Dataset

Earth observation plays a crucial role in environmental monitoring, disaster damage assessment, and sustainable resource management, among other applications. These tasks may be affected by atmospheric haze, which reduces the contrast in Sentinel-2 multispectral imagery (MSI) as well as other satellite MSI. Hence, haze detection is essential as an early-stage screening step to further enhance the performance of satellite imagery analysis. This paper presents a fast and reliable network called CONVFORMER-LITE, a lightweight Convolutional–Transformer for binary haze/non-haze image classification on true colour bands of Sentinel-2 MSI. The proposed model combines efficient ConvMixer-based spatial integration with a Safe Multi-Head Self-Attention 2D module, i.e., a token-budgeted multi-head self-attention block that applies full attention when the token count is within a preset budget, and otherwise reduces tokens via pooling (factor$p$) before attention and upsamples back to the original spatial size. To enable robust training and evaluation of the proposed network under strong geographic diversity, we introduce a Pan-Asian Sentinel-2 haze/non-haze benchmarking dataset, constructed from 100 locations using a standard physics-guided atmospheric-scattering model with cirrus-band-guided haze thickness control. The proposedConvFormer-Liteachieves near-perfect performance (Accuracy/F1-score$\approx 0.9998$, Area Under the ROC Curve (AUC)$\approx 1.0$) with 4.68 M parameters and 0.712 G multiply-accumulate operations, which outperforms several recent lightweight baseline models (e.g.,CoAtNet-LiteandEConvNeXt-Lite). The complexity and runtime analyses reveal thatConvFormer-Liteoffers an excellent trade-off between accuracy and efficiency, which makes it a practical front-end classifier for haze-aware preprocessing in large Sentinel-2 processing pipelines. The implementation of our work is provided athttps://github.com/kumarballapavan/CFL-H-NH-class-PAS2DB.