FOREST-GC: A conFOrmable Rendering Engine for Synthetic Tree Generation and Counting

Accurate tree counting from satellite imagery remains a critical challenge for agricultural monitoring, climate analysis, and land-use management. While recent advances in deep learning, particularly large pre-trained Vision-Language Models (VLMs), have achieved remarkable results in general visual tasks, we show that they fail to reliably address tree counting in high-resolution satellite imagery. This highlights the need for specialized datasets to support the development and evaluation of models capable of solving this task, especially in plantation environments. To this end, we introduce a configurable synthetic image generator that produces realistic satellite imagery of olive plantations along with precise annotations in terms of the number and positions of trees, providing valuable resources for advancing tree counting research in agricultural settings. The generator simulates terrain textures, spatial distributions of trees, and tree shadows to approximate real-world conditions and exposes over 100 customizable parameters to control appearance and layout. It also includes a simulated Near-Infrared (NIR) layer to support vegetation-specific spectral analysis. We compare deep learning methods based on object detection or segmentation models trained on the generated data with pre-trained VLMs and an unsupervised computer vision baseline, and show that synthetic training achieves superior accuracy on a manually annotated real-world dataset, demonstrating effective sim-to-real performance transfer specifically in olive plantation imagery. In particular, SAM-ViT-Large achieves an average relative error of 4.82% on the real data, which is comparable to the performance of human labelers. The generator source code is publicly available, supporting further research in data-scarce agricultural monitoring with a focus on plantation analysis.