Deep Learning Based Mars Safe Landing Analysis and Terrain Classification

Abstract

The exponential growth in high-resolution imagery returned by Mars missions—Perseverance and Curiosity together transmit thousands of images daily—creates a classification bottleneck that manual analysis cannot sustain. Automated terrain classification is therefore essential not only for accelerating scientific discovery but, more critically, for identifying safe landing zones prior to crewed or robotic descent. This paper presents a full-stack, web-deployable deep learning system for Mars terrain classification and operational safety interpretation. A MobileNetV2 backbone, fine-tuned via transfer learning from ImageNet pre-trained weights, serves as the classification engine. Its inverted residual architecture and depthwise separable convolutions deliver competitive accuracy with minimal computational overhead, making the model deployable in resourceconstrained environments without GPU acceleration. The classifier recognises eight terrain categories—smooth plains, sandy dunes, rocky plains, rock fields, impact craters, valleys, bedrock outcrops, and polar terrain—and maps each predicted class to a three-tier operational safety designation (safe, caution, hazardous) based on quantified landing risk criteria. The inference pipeline is integrated into a Flask RESTful web application exposing four endpoints: /upload for validated image ingestion, /classify for inference with safety enrichment, /model-status for real-time readiness reporting, and / and /about for user interaction. Comprehensive input validation rejects non-image file types, oversized submissions, and malformed requests with structured JSON error responses, while a gracefuldegradation architecture keeps the application fully functional and informative even when model loading fails. A 31-case test suite spanning unit, integration, system, and performance categories achieves a 100% pass rate. Endpoint latency measurements confirm sub-200 ms page loads, 0.54 ms model-status checks, and 118 ms upload completion—all suitable for interactive local deployment. The addition of an operational safety layer uniquely bridges the gap between scientific classification output and mission-planning actionability, a capability absent from all reviewed prior systems.