notes:n23The Limits of AI: Geospatial Foundation Models Do Not Understand GeographyGeospatial foundation models can learn powerful spatial representations. Their harder test is geographic reasoning: scale, spatial dependence, regional heterogeneity, mechanism, and transfer beyond familiar places.2026.09.15
notes:n01A Model of Good Practice for Machine Learning in Remote SensingThe hard part is not fitting a model. It is proving that what the model learned survives sparse samples, spatial heterogeneity, weak signals, and geographic transfer.2026.07.04
notes:n02nature-reviewer-skills: 450 Recurring Critique Patterns from 23,000 Nature Review CommentsA review-assistance system built from public peer-review files: not a substitute for judgment, but a structured catalogue of the questions reviewers repeatedly ask.2026.07.03
notes:n03Distilling Nature Reviewers in Earth-System ScienceA domain-specific extension of reviewer-pattern mining across atmosphere, hydrology, ecology, and remote sensing.2026.07.02
notes:n04Distilling a Nature Remote-Sensing ReviewerA lightweight reviewer skill assembled from public review files, focused on the points where remote-sensing papers most often lose credibility.2026.07.01
notes:n05Opening the AlphaEarth Black Box: Physical Semantics in the EmbeddingA new study asks whether AlphaEarth’s latent dimensions correspond to interpretable land-surface properties rather than merely useful predictive features.2026.04.30
notes:n06Writing Causal Claims Without OverreachCausal language is earned by a question, a defensible causal structure, an identification strategy, and sensitivity analysis—not by significance alone.2026.04.29
notes:n07Greener Is Not Necessarily More ProductiveGreenness, vegetation cover, and productivity often move together—but global evidence shows that they diverge across large parts of the vegetated world.2026.04.27
notes:n08Chl-a Is Not Biomass: When the Proxy ReversesChlorophyll-a and phytoplankton biomass can decouple, vary for different physiological reasons, and in some regions even move in opposite directions.2026.04.25
notes:n09Global Cropland Dynamics, 2015–2024A new Landsat-based annual cropland dataset shows where farmland expanded, where it contracted, and how trade, conflict, drought, and land conversion left different regional signatures.2026.04.24
notes:n10What Remote Sensing Sees—and What It Does Not: The Proxy-of-a-Proxy ProblemHigh retrieval accuracy does not guarantee a valid environmental conclusion. Remote sensing often estimates a measurable variable first and interprets the target object only through a second proxy relation.2026.04.23
notes:n11Qiusheng Wu and the Open-Source Route into Geospatial ComputingA personal note on the educator whose open-source teaching made GIS and remote sensing feel learnable by building, not by memorising software menus.2026.04.23
notes:n12Retrieving Chl-a Is Not Retrieving Phytoplankton BiomassThe satellite product may be chlorophyll-a; the paper may discuss biomass. The missing step between them is an ecological representation relation, not part of the retrieval itself.2026.04.22
notes:n13Some Trends Are Artifacts of Ignored AutocorrelationLong time series contain memory. If that dependence is ignored, apparent accelerations, breakpoints, and regime shifts can look more certain than the evidence allows.2026.04.19
notes:n14Can Satellites and AI Reveal “Maxwell Equations” for the Earth System?A Wuhan University perspective proposes a planetary laboratory in which global satellite observations and AI are used to search for compact, interpretable laws of Earth-system behaviour.2026.04.15
notes:n15Long Remote-Sensing Records and the Problem of Temporal ConsistencyA decades-long satellite series is rarely one instrument watching continuously. Sensor changes, orbit drift, calibration, fusion, and processing can create trends that look environmental.2026.04.13
notes:n16When Nature Disagrees with Nature: The Observation Problem in Remote SensingTwo apparently conflicting studies of Amazon seasonality expose two different failures of remote sensing: seeing a signal created by geometry, and failing to sample enough of the phenomenon.2026.04.12
notes:n17Causal Shapley Is Not Causal IdentificationAdding a causal graph to Shapley attribution can change the decomposition, but it does not by itself identify a causal effect from observational data.2026.04.11
notes:n18Three Boundaries of SHAP: Attribution, Mechanism, and Causal IdentificationThree Nature-family papers show that Shapley values can play very different epistemic roles depending on the surrounding evidence.2026.04.10
notes:n19SHAP’s Interpretive Boundary: From Model Attribution to Causal IdentificationThe question is not whether SHAP is “misused” in the abstract, but which class of claim the evidence actually supports.2026.04.09
notes:n20Extrapolation Risk in Global-Scale StudiesA global map can look continuous while its evidence remains local. Three questions matter: the input data, the validation design, and the boundary of extrapolation.2026.03.29
notes:n21Reassessing Global Tree-Restoration Potential After a Wave of CritiquesThe influential 2019 Science estimate of global restoration potential became a case study in how basemaps, validation, and extrapolation shape policy-scale numbers.2026.03.27
notes:n22Stress-Testing AlphaEarth in AgricultureAlphaEarth is competitive under local training, but its advantage weakens as agricultural tasks demand cross-region transfer, fine temporal sensitivity, and physical interpretation.2026.03.26