AA
Andika Hadi Andika Hadi Hutama
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Historical maps are a rich source of information about past urban landscapes, but their contents remain largely inaccessible for computational analysis. This thesis investigates and compares two deep learning approaches for automated urban extent extraction from Dutch historical topographic maps, and applies the results to urban change analysis.
Method 1 applies patch-level image classification using the MapReader framework on the Bonnebladen map series (1:25,000) covering the Province of South Holland. Map sheets are divided into 200-meter square patches and classified into three classes using a fine-tuned ResNet-101 model. A weak classifier trained on 1,768 manually annotated patches achieved a macro-averaged F1-score of 0.900. Inference was applied to 74 sheets across two temporal editions (T1: 1869–1916, T2: 1910–1943), producing georeferenced classified patch datasets at provincial scale. Post-classification comparison between T1 and T2 detected approximately 65 km² of total urban growth in South Holland, with new building blocks as the dominant growth pattern. The results are validated against the HGN-1900 land use dataset, showing broad spatial agreement in urban areas.
Method 2 applies object-level instance segmentation using Mask R-CNN on Rotterdam 1897 municipality maps (1:5,000). To avoid manual annotation, BGT vector data is style-transferred to approximate the historical cartographic style and translated using CycleGAN to generate a synthetic training dataset with automatically aligned annotation masks. The epoch 6 CycleGAN checkpoint is selected based on FID evaluation and visual inspection of building boundary preservation. The Mask R-CNN model trained on this synthetic data achieved a validation mAP50 of 0.8784. After post-processing and shape regularisation, 12,299 building polygon predictions are produced from sheet 3. External evaluation against BGT buildings confirmed to predate 1897 yielded a recall of 41.4%, reflecting both model detection limitations and the substantial post-war urban renewal of Rotterdam following the 1940 bombing.
The two methods differ substantially in their capabilities and requirements. Method 1 scales to provincial level with limited annotation effort and supports statistical-level change analysis, but is sensitive to cartographic style shifts across temporal editions and produces coarse 200-meter raster output. Method 2 produces building-level vector polygons enabling urban persistence analysis, but requires a complex multi-stage pipeline, is constrained by georeferencing quality, and depends on synthetic training data that underrepresents dense historic urban configurations. Both methods are fully operational on a consumer-grade GPU laptop, confirming their accessibility without HPC infrastructure. Together, they demonstrate that IIIF-based map access infrastructure and modern deep learning techniques can be combined to support reproducible urban change analysis from historical map collections at multiple scales of spatial granularity. ...
Method 1 applies patch-level image classification using the MapReader framework on the Bonnebladen map series (1:25,000) covering the Province of South Holland. Map sheets are divided into 200-meter square patches and classified into three classes using a fine-tuned ResNet-101 model. A weak classifier trained on 1,768 manually annotated patches achieved a macro-averaged F1-score of 0.900. Inference was applied to 74 sheets across two temporal editions (T1: 1869–1916, T2: 1910–1943), producing georeferenced classified patch datasets at provincial scale. Post-classification comparison between T1 and T2 detected approximately 65 km² of total urban growth in South Holland, with new building blocks as the dominant growth pattern. The results are validated against the HGN-1900 land use dataset, showing broad spatial agreement in urban areas.
Method 2 applies object-level instance segmentation using Mask R-CNN on Rotterdam 1897 municipality maps (1:5,000). To avoid manual annotation, BGT vector data is style-transferred to approximate the historical cartographic style and translated using CycleGAN to generate a synthetic training dataset with automatically aligned annotation masks. The epoch 6 CycleGAN checkpoint is selected based on FID evaluation and visual inspection of building boundary preservation. The Mask R-CNN model trained on this synthetic data achieved a validation mAP50 of 0.8784. After post-processing and shape regularisation, 12,299 building polygon predictions are produced from sheet 3. External evaluation against BGT buildings confirmed to predate 1897 yielded a recall of 41.4%, reflecting both model detection limitations and the substantial post-war urban renewal of Rotterdam following the 1940 bombing.
The two methods differ substantially in their capabilities and requirements. Method 1 scales to provincial level with limited annotation effort and supports statistical-level change analysis, but is sensitive to cartographic style shifts across temporal editions and produces coarse 200-meter raster output. Method 2 produces building-level vector polygons enabling urban persistence analysis, but requires a complex multi-stage pipeline, is constrained by georeferencing quality, and depends on synthetic training data that underrepresents dense historic urban configurations. Both methods are fully operational on a consumer-grade GPU laptop, confirming their accessibility without HPC infrastructure. Together, they demonstrate that IIIF-based map access infrastructure and modern deep learning techniques can be combined to support reproducible urban change analysis from historical map collections at multiple scales of spatial granularity. ...
Historical maps are a rich source of information about past urban landscapes, but their contents remain largely inaccessible for computational analysis. This thesis investigates and compares two deep learning approaches for automated urban extent extraction from Dutch historical topographic maps, and applies the results to urban change analysis.
Method 1 applies patch-level image classification using the MapReader framework on the Bonnebladen map series (1:25,000) covering the Province of South Holland. Map sheets are divided into 200-meter square patches and classified into three classes using a fine-tuned ResNet-101 model. A weak classifier trained on 1,768 manually annotated patches achieved a macro-averaged F1-score of 0.900. Inference was applied to 74 sheets across two temporal editions (T1: 1869–1916, T2: 1910–1943), producing georeferenced classified patch datasets at provincial scale. Post-classification comparison between T1 and T2 detected approximately 65 km² of total urban growth in South Holland, with new building blocks as the dominant growth pattern. The results are validated against the HGN-1900 land use dataset, showing broad spatial agreement in urban areas.
Method 2 applies object-level instance segmentation using Mask R-CNN on Rotterdam 1897 municipality maps (1:5,000). To avoid manual annotation, BGT vector data is style-transferred to approximate the historical cartographic style and translated using CycleGAN to generate a synthetic training dataset with automatically aligned annotation masks. The epoch 6 CycleGAN checkpoint is selected based on FID evaluation and visual inspection of building boundary preservation. The Mask R-CNN model trained on this synthetic data achieved a validation mAP50 of 0.8784. After post-processing and shape regularisation, 12,299 building polygon predictions are produced from sheet 3. External evaluation against BGT buildings confirmed to predate 1897 yielded a recall of 41.4%, reflecting both model detection limitations and the substantial post-war urban renewal of Rotterdam following the 1940 bombing.
The two methods differ substantially in their capabilities and requirements. Method 1 scales to provincial level with limited annotation effort and supports statistical-level change analysis, but is sensitive to cartographic style shifts across temporal editions and produces coarse 200-meter raster output. Method 2 produces building-level vector polygons enabling urban persistence analysis, but requires a complex multi-stage pipeline, is constrained by georeferencing quality, and depends on synthetic training data that underrepresents dense historic urban configurations. Both methods are fully operational on a consumer-grade GPU laptop, confirming their accessibility without HPC infrastructure. Together, they demonstrate that IIIF-based map access infrastructure and modern deep learning techniques can be combined to support reproducible urban change analysis from historical map collections at multiple scales of spatial granularity.
Method 1 applies patch-level image classification using the MapReader framework on the Bonnebladen map series (1:25,000) covering the Province of South Holland. Map sheets are divided into 200-meter square patches and classified into three classes using a fine-tuned ResNet-101 model. A weak classifier trained on 1,768 manually annotated patches achieved a macro-averaged F1-score of 0.900. Inference was applied to 74 sheets across two temporal editions (T1: 1869–1916, T2: 1910–1943), producing georeferenced classified patch datasets at provincial scale. Post-classification comparison between T1 and T2 detected approximately 65 km² of total urban growth in South Holland, with new building blocks as the dominant growth pattern. The results are validated against the HGN-1900 land use dataset, showing broad spatial agreement in urban areas.
Method 2 applies object-level instance segmentation using Mask R-CNN on Rotterdam 1897 municipality maps (1:5,000). To avoid manual annotation, BGT vector data is style-transferred to approximate the historical cartographic style and translated using CycleGAN to generate a synthetic training dataset with automatically aligned annotation masks. The epoch 6 CycleGAN checkpoint is selected based on FID evaluation and visual inspection of building boundary preservation. The Mask R-CNN model trained on this synthetic data achieved a validation mAP50 of 0.8784. After post-processing and shape regularisation, 12,299 building polygon predictions are produced from sheet 3. External evaluation against BGT buildings confirmed to predate 1897 yielded a recall of 41.4%, reflecting both model detection limitations and the substantial post-war urban renewal of Rotterdam following the 1940 bombing.
The two methods differ substantially in their capabilities and requirements. Method 1 scales to provincial level with limited annotation effort and supports statistical-level change analysis, but is sensitive to cartographic style shifts across temporal editions and produces coarse 200-meter raster output. Method 2 produces building-level vector polygons enabling urban persistence analysis, but requires a complex multi-stage pipeline, is constrained by georeferencing quality, and depends on synthetic training data that underrepresents dense historic urban configurations. Both methods are fully operational on a consumer-grade GPU laptop, confirming their accessibility without HPC infrastructure. Together, they demonstrate that IIIF-based map access infrastructure and modern deep learning techniques can be combined to support reproducible urban change analysis from historical map collections at multiple scales of spatial granularity.