Immune checkpoint blockade (ICB) has transformed melanoma treatment, yet a large proportion of patients fail to respond, and resistance signatures remain poorly defined. Here, we utilised a 480-gene single-cell spatial transcriptomics panel to profile 287 pre-treatment melanoma specimens from two independent cohorts, including patients treated in the advanced (n = 91) and adjuvant (n = 58) settings. We developed NIMO, a highly predictive deep learning-based framework to capture the spatial-molecular landscape of the tumour microenvironment. We reveal multiple resistant niches exhibiting distinct phenotypes, and melanoma cells showing proliferative and survival programs, with elevated endothelial signalling. We further reveal distinct molecular profiles of lymphoid aggregates in non-responders, highlighting a subniche rich in B and CD4⁺ T cells exhibiting a molecular signature failing to support effective antitumour immunity, including elevated immunoregulation and diminished chemokine expression. This study provides novel insights into the spatial heterogeneity of ICB resistance to inform patient stratification and personalised therapy.