Illegal activities such as drug trafficking, illegal mining, and deforestation pose serious threats to national security, environmental sustainability, and economic stability in developing countries like Ecuador. Due to human, economic, and logistical limitations, detecting such activities in sparsely populated and remote areas like the Amazon rain-forest remains highly challenging. This study presents a deep learning-based approach for detecting illicit activity hotspots by classifying multispectral satellite imagery. Leveraging convolutional neural networks (CNNs) and the AmazonCrime dataset, the proposed system identifies indicators such as illegal mining operations, illicit crops cultivations, clandestine airstrips, and land cover changes. A key contribution of this work lies in the comparative analysis of pretrained CNN architectures versus custom-trained models, using both traditional RGB composites and extended spectral bands from Sentinel-2 imagery. This study also investigates the relative importance of the spectral bands for classification accuracy, highlighting their role in enhancing detection capabilities. In addition, pixel-level segmentation techniques are employed to generate class-specific masks, enabling precise spatial quantification of illicit activities. The proposed framework delivers actionable geospatial intelligence to improve situational awareness, inform decision-making, and support strategic interventions by military and law enforcement agencies. The findings contribute to the development of more effective, data-driven strategies for protecting national territory and mitigating the environmental and social impacts of organized criminal activity in ecologically sensitive and geopolitically complex regions.