GeoDIM: Dissimilarity-Priced Federated Edge Learning over Geographically Heterogeneous Environmental Observation Networks

Environmental observation networks are federating: weather and hydrology stations, air-quality sub-stations and satellite ground receiving stations increasingly train downstream Earth-science models locally and exchange model updates instead of raw archives, because of data sovereignty, archive size and response deadlines. Their heterogeneity is not random label skew but a geographically structured property of climate zones, soils, phenological phases and sensors; their data volumes differ by an order of magnitude across sites; and their operating preferences flip with the seasons while connectivity is intermittent. We present GeoDIM, a geoscience dissimilarity measure that decomposes site differences into identifiable covariate, label-prior and concept components on a shared unlabelled probe set, weights them by site descriptors, and drives both dissimilarity-discounted aggregation and a spectral clustering into climate–phenology clusters; GAPO, a preference-conditioned hypernetwork policy whose low-rank parameter delta covers the whole Pareto front of accuracy, communication and energy, whose state is conditioned on site descriptors for zero-shot transfer, and whose gradients are projected onto the feasible set of latency, energy and uplink budgets; and SPARC, which restricts sharing to a low-rank backbone, personalizes through a descriptor-retrieved bank of recalibration controllers, and aggregates asynchronously under a staleness bound. On crop-type mapping, Sentinel-1 flood mapping, crop-yield regression and cross-continent land-cover classification, the framework reaches 0.914 mean AUC, 93.4 mIoU, 5.02 t/ha RMSE and 0.36 kappa, retains 85.2–91.3% of source-domain performance under new climate zones, continents and sensors, and cuts the upload required to reach the target metric from 9.66GB to 0.54GB.