Chronological feed of everything captured from Rob Miles.
ELDOR is a large-scale UAV-derived benchmark dataset for detecting illegal gold mining activity in the Amazon, covering over 2,500 hectares with pixel-level semantic annotations. It defines four benchmark tasks spanning semantic segmentation, multi-label classification, and vision-language model evaluation under a closed-set protocol. Evaluation across generic, remote-sensing-specific, and foundation model-based methods reveals that all current approaches struggle with rare small-scale mining structures and fine-grained ecological recovery classes, indicating that satellite imagery and existing CV models are insufficient for fine-scale environmental monitoring in this domain.
This paper positions Mars as a high-fidelity analog for characterizing small, rocky exoplanets at the margins of habitability, leveraging decades of in-situ mission data (MRO, MAVEN, MSL, Mars2020) alongside JWST atmospheric observations. The authors synthesize how Mars's planetary evolution—spanning volatile loss, CO₂ condensation, magnetosphere collapse, and obliquity-driven climate swings—provides direct diagnostic frameworks applicable to exoplanet atmosphere characterization. The convergence of Solar System planetary science and exoplanet observational astronomy is framed as a dual-pathway methodology for identifying and interpreting Mars analogs beyond the Solar System.
Liu, Wang, and Lopes introduce the first goodness-of-fit (GoF) test for Independent Component (IC) models with formal theoretical validity guarantees in the high-dimensional proportional asymptotic regime (dimension and sample size diverging proportionally). The key methodological advance is the deliberate avoidance of a pre-whitening step, which is the primary bottleneck that renders competing GoF tests unreliable in high dimensions. The test is validated both theoretically (size control) and empirically (power across varied conditions), with demonstrated applicability to real-world gene-expression data.