Deep Transfer Learning for Forest Classification Using Multispectral Satellite Imagery

Abstract. Accurate forest-cover mapping is essential for biodiversity conservation, carbon accounting, and sustainable land-use planning. This work presents a novel deep learning pipeline that fuses spectral and vegetation-index information to classify forest and non-forest areas using Sentinel-2 satellite imagery. The proposed approach achieves 98% accuracy and a 98% F₁ score on the EuroSAT benchmark. At the core of the method is a modified EfficientNet-B0 backbone, pretrained on ImageNet, with the Normalized Difference Vegetation Index (NDVI) integrated as a fourth input channel alongside the standard RGB bands. To facilitate this fusion, the first convolutional layer is adapted to accept four input channels, and the NDVI weights are initialized using the mean of RGB parameters. This configuration enables the model to jointly capture chlorophyll-related absorption characteristics and structural reflectance cues within a single, end-to-end architecture. To enhance generalization, class imbalance is addressed by sampling an equal number of non-forest patches from nine complementary land-cover classes. The training process incorporates extensive data augmentation—such as random flips, rotations, and color jitter—and employs mixed-precision training with a cosine-annealing learning-rate schedule, resulting in faster convergence and improved robustness. The model demonstrates effective delineation of forest boundaries across diverse landscapes, as confirmed by quantitative metrics and visual overlays. The complete pipeline is implemented using a single Python script, requiring no proprietary software or GIS platforms, and is organized using a simple folder- based structure. The approach offers a scalable, efficient, and reproducible solution for large-scale forest monitoring and classification tasks.