Illegal mining represents a major environmental and socio-political threat in the Brazilian Amazon, particularly within protected areas and indigenous territories. While optical remote sensing has been widely used to detect mining activity, its utility is limited by persistent cloud cover. This paper explores the potential of C-band Synthetic Aperture Radar (SAR) imagery from Sentinel-1, combined with a lightweight convolutional neural network (CNN), to identify illegal mining under such challenging conditions. The model was trained on seven Sentinel-1 scenes from the Tapajós basin (state of Pará) and evaluated both within this training region and on an independent test set from the Yanomami Indigenous Territory (state of Roraima), using reference data from the Amazon Mining Watch (AMW) project. A total of 2,394 labelled patches supported supervised training. Results show balanced performance in the Tapajós region (F1-score = 0.676) and robust generalization to the Yanomami region (F1-score = 0.630. Most errors were associated with peripheral mining structures and small-scale disturbances, reflecting difficulties in capturing low-density mining patterns. Overall, the findings demonstrate the full potential of SAR-based deep learning approaches for monitoring illegal mining in persistently cloud-covered Amazonian landscapes. Future improvements may come from integrating terrain variables such as elevation and hydrological proximity, as mining often follows narrow streams (<em>igarapés</em>) closely tied to local topography.