Infectious-disease risk models often rely on occurrence records that are incomplete and spatially biased by surveillance effort, diagnostic access, and outbreak history. Ecological niche modelling (ENM) can identify areas where disease occurrence is environmentally plausible, yet most approaches represent locations using only pointwise covariate values and therefore overlook the surrounding spatial context and rely on presence-only data. Here, we present a deep-learning framework for presence-only data that estimates relative disease suitability by comparing the environmental conditions surrounding reported occurrences with those sampled across the wider study area. The model processes gridded environmental patches at local, neighbourhood, and broader landscape scales, learns the contribution of each scale, and accommodates missing raster values. Using dengue virus as a global case study, we evaluate whether multiscale spatial representation improves upon point-based ENM baselines including random forest and maximum entropy (MaxEnt) under a spatially disjoint train-test design. The model achieved a Boyce index of 0.971 and an AUC of 0.976 on the held-out test set. Learned scale weights and ablation experiments indicated that neighbourhood context contributed most strongly, while local and broader-scale information provided complementary predictive signals. Compared with point-based baselines, the model identified 6-18% more environmentally suitable area across South Asia, Southeast Asia, and South America, encompassing tens of millions of residents. These findings demonstrate that multiscale spatial context can improve estimates of relative dengue suitability. More broadly, mask-aware convolutional density-ratio estimation provides a flexible framework for mapping environmentally structured pathogens from incomplete, presence-only occurrence data.