fix(uc3-uhi): unblock the Guatemala UHI use case on develop

configs/experiment/guatemala/heat_guatemala_full_fusion_avg_reg.yaml

@@ -0,0 +1,38 @@
+# @package _global_
+defaults:
+  - override /model: heat_full_fusion_avg_reg
+  - override /data: heat_guatemala
+  - override /metrics: guatemala_regression
+
+tags: ["heat_island", "guatemala", "full_fusion_avg", "tessera", "regression"]
+seed: 12345
+
+trainer:
+  min_epochs: 1
+  max_epochs: 50
+
+data:
+  batch_size: 64
+  dataset:
+    modalities:
+      coords: {}
+      tessera:
+        year: 2024
+        size: 10
+        format: npy
+
+callbacks:
+  model_checkpoint:
+    monitor: val_r2
+    mode: max
+  early_stopping:
+    monitor: val_r2
+    mode: max
+    patience: 20
+
+logger:
+  wandb:
+    tags: ${tags}
+    group: "heat_island"
+  aim:
+    experiment: "heat_island"

configs/experiment/guatemala/heat_guatemala_tessera_tabular_avg_best.yaml

@@ -0,0 +1,38 @@
+# @package _global_
+defaults:
+  - override /model: heat_tessera_tabular_avg_best
+  - override /data: heat_guatemala
+  - override /metrics: guatemala_regression
+
+tags: ["heat_island", "guatemala", "full_fusion_avg", "tessera", "regression"]
+seed: 12345
+
+trainer:
+  min_epochs: 1
+  max_epochs: 50
+
+data:
+  batch_size: 64
+  dataset:
+    modalities:
+      coords: {}
+      tessera:
+        year: 2024
+        size: 10
+        format: npy
+
+callbacks:
+  model_checkpoint:
+    monitor: val_r2
+    mode: max
+  early_stopping:
+    monitor: val_r2
+    mode: max
+    patience: 20
+
+logger:
+  wandb:
+    tags: ${tags}
+    group: "heat_island"
+  aim:
+    experiment: "heat_island"

configs/inference_heat_guatemala_concept_similarity.yaml

@@ -0,0 +1,29 @@
+# Does text similarity agree with measured LST?
+# Run: python src/inference/inference_heat_guatemala_concept_similarity.py
+defaults:
+  - paths: ${oc.env:STORAGE_MODE,local}
+  - extras: default
+  - hydra: default
+  - data: heat_guatemala_tessera_text
+  - _self_
+
+task_name: "inference"
+tags: ["inference", "heat_island", "guatemala", "concept_similarity"]
+seed: 12345
+
+# Aligned embeddings come from the alignment run, LST predictions from the
+# predictive run. Both consume the same pooled Tessera vector.
+alignment_ckpt_path: ${paths.checkpoint_dir}/epoch_035.ckpt
+predictive_ckpt_path: ${paths.checkpoint_dir}/epoch_093.ckpt
+
+concepts_file: ${paths.root_dir}/outputs/demo_concepts.json
+split: test
+device: cuda
+
+output_csv: ${paths.root_dir}/outputs/heat_guatemala_concept_similarity_per_location.csv
+output_summary_csv: ${paths.root_dir}/outputs/heat_guatemala_concept_similarity_summary.csv
+
+data:
+  dataset:
+    use_target_data: true       # the measured LST is the thing being validated against
+  saved_split_file_name: "split_indices_heat_guatemala_2026-08-27-clean.pth"

configs/inference_heat_guatemala_embeddings.yaml

@@ -0,0 +1,28 @@
+# Aligned Tessera embeddings for the Guatemala UHI use case.
+# Run: python src/inference/inference_heat_guatemala_embeddings.py
+defaults:
+  - paths: ${oc.env:STORAGE_MODE,local}
+  - extras: default
+  - hydra: default
+  - _self_
+
+task_name: "inference"
+tags: ["inference", "heat_island", "guatemala", "tessera"]
+seed: 12345
+
+alignment_ckpt_path: ${paths.checkpoint_dir}/epoch_035.ckpt
+
+points_csv: ${paths.data_dir}/heat_guatemala/model_ready_heat_guatemala.csv
+limit: null                     # process only the first N rows, for a quick check
+
+tile_dir: ${paths.data_dir}/heat_guatemala/eo/tessera
+year: 2024
+tile_size: 10
+# v1.1 covers very little of Guatemala City (11 tiles for 2024 in the 1x1 degree
+# box); every tile in this use case came from v1.
+tessera_version: v1
+
+output_csv: ${paths.root_dir}/outputs/heat_guatemala_aligned_embeddings.csv
+wide_columns: false             # true: one column per dimension instead of one cell per vector
+normalize: true                 # the contrastive loss normalises, so keep this on for cosine
+device: cuda

configs/model/heat_full_fusion_cnn_best.yaml

@@ -13,8 +13,6 @@ geo_encoder:
         geo_data_name: tessera
         resnet_version: 18
         pretrained_cnn: imagenet
-        freezing_strategy: all
-        output_dim: 256
   fusion_strategy: concat
 prediction_head:
   _target_: src.models.components.pred_heads.mlp_regression_head.MLPRegressionPredictionHead

configs/model/heat_tessera_cnn_best.yaml

@@ -4,8 +4,6 @@ geo_encoder:
   geo_data_name: tessera
   resnet_version: 34
   pretrained_cnn: imagenet
-  freezing_strategy: all
-  output_dim: 256
 prediction_head:
   _target_: src.models.components.pred_heads.mlp_regression_head.MLPRegressionPredictionHead
   nn_layers: 2

outputs/demo_concepts.json

@@ -0,0 +1,45 @@
+{
+  "_comment": "Caption vocabulary for the Guatemala UHI demo. Each concept is scored by encoding every caption with the alignment model's text branch, averaging, L2-normalising, and taking the dot product with a location's aligned embedding. Only concepts with a positive skill index on the test split are included - see heat_guatemala_concept_evaluation.csv.",
+  "_checkpoint": "data/checkpoints/other_ckpt/epoch_035.ckpt",
+  "concepts": {
+    "vegetation_greenness": {
+      "label": "Vegetation greenness",
+      "recommended": true,
+      "skill_index": 0.7476,
+      "captions": [
+        "an area with high vegetation greenness",
+        "a location with lush green vegetation",
+        "a green area with healthy dense vegetation"
+      ]
+    },
+    "forest_cover": {
+      "label": "Forest cover",
+      "recommended": true,
+      "skill_index": 0.5409,
+      "captions": [
+        "an area with dense forest cover",
+        "a location covered by forest",
+        "a heavily forested area with many trees"
+      ]
+    },
+    "builtup_age": {
+      "label": "Established urban fabric",
+      "recommended": true,
+      "skill_index": 0.478,
+      "captions": [
+        "a long-established built-up area",
+        "an old urban area built many decades ago",
+        "a mature consolidated urban neighbourhood"
+      ]
+    }
+  },
+  "_excluded": {
+    "_note": "These scored at or below baseline on the test split. They will still return confident-looking numbers, which is exactly why they must stay out of the demo vocabulary.",
+    "tree_cover": 0.2178,
+    "lst_max_hottest": -0.1336,
+    "ndwi_water": -0.3533,
+    "slope": -0.0006,
+    "cool_area_r_vs_measured_lst": 0.217,
+    "_cool_area_note": "Removed after validation: similarity to 'coolest areas' correlates POSITIVELY with measured LST (r=+0.217), i.e. it points at hotter places. It tracks established urban fabric, not coolness. Captions kept here for reference only: one of the coolest areas in the city | a location with low land surface temperature | a cool shaded part of the urban area"
+  }
+}