Illegal mining poses significant environmental and socio-political challenges in the Brazilian Amazon, particularly within protected areas and indigenous territories. While optical remote sensing has been widely employed for detecting mining activities, its effectiveness is often hindered by persistent cloud cover. The aim of this paper is to investigate the use of C-band Synthetic Aperture Radar (SAR) imagery from Sentinel-1 combined with a lightweight convolutional neural network (CNN) to detect illegal mining sites under challenging atmospheric conditions. The model was trained on seven Sentinel-1 scenes from the Tapajós basin (Pará) and evaluated within the training region as well as on an independent test set from the Yanomami Indigenous Territory (Roraima), using reference annotations sourced from the Amazon Mining Watch project. A total of 2,394 labelled patches supported the supervised training. The CNN achieved balanced classification performance in the Tapajós area (F1-score: 0.676 at 0.80 threshold) and demonstrated the generalization capabilities in the unseen Yanomami region (F1-score: 0.630 at 0.90 threshold). Detection errors were mainly related to peripheral mining structures and small-scale disturbances, indicating challenges in identifying low-density mining patterns. These findings highlight the promise of SAR-based deep learning methods for monitoring illegal mining in cloud-prone Amazonian regions. Future work could improve detection accuracy by integrating terrain variables—such as elevation and proximity to watercourses—given the common occurrence of mining activities along narrow streams (igarapés) closely tied to local topography.