The dry land is now being asked to map its own retreat
the dry land is now being asked to map its own retreat: a swin transformer and capsule network classify desertification across pakistan's arid south at 99.8 percent accuracy, a gradient boosted model on 40 years of landsat tracks degradation in oasis zones at 87.2 percent, and a u net reads 3 meter planetscope to watch shrub encroachment creep across inner mongolian steppe.
the sand now files its own boundary report.
Context
SwinCapT combines a Swin Transformer with a Capsule Network for multispectral land cover classification. The PeerJ Computer Science paper tests it on a dataset from Thatta and Badin in Pakistan covering 2010 to 2023, and reports 99.8% accuracy, 0.984 precision, 0.983 recall and 0.982 F1, ahead of CNN, LSTM, ResNet, VGG16 and ViT baselines.
The Scientific Reports study of the Ternata Oasis in southeastern Morocco uses the full Landsat archive from 1984 to 2024 in Google Earth Engine. A Gradient Tree Boosting classifier reaches 87.2% overall accuracy, ahead of Random Forest at 85.0% and CART at 82.0%. Farmer interviews and hydrogeological fieldwork report salinity stress, deeper wells and a drop in groundwater levels.
The PlanetScope study maps shrub cover in Xilinhot, Inner Mongolia, with a U-Net on 3 metre imagery in 15-day composites from May to October. Within-plot R2 is 0.649 and RMSE 0.055; leave-one-plot-out validation gives a mean R2 of 0.411. At a common 10 metre scale it reports R2 0.825 against 0.702 for Sentinel-2 with XGBoost.
The three accuracy figures measure different things. The 99.8% is classification accuracy on land cover classes in two districts of southern Pakistan, so it is not a measure of how well desertification is tracked. The 87.2% is overall accuracy for one oasis in Morocco, not for oasis zones in general. The shrub result is a regression of shrub cover with an R2 of 0.649 within plots and 0.411 on held-out plots, a long way from the first number, and the paper lists native-scale smoothing and limited independent validation as constraints. The note puts all three in one sentence, and the paper's own caveats separate them.
Related work
- Optimizing swin transformer encoder for spatiotemporal desertification classification in an arid environment (PeerJ Computer Science) ↗The SwinCapT paper with the Thatta and Badin dataset and the 99.8% figure.
- Desertification monitoring in arid oasis environment using Google Earth Engine, machine learning, and field-based hydrogeological assessment (Scientific Reports) ↗The Ternata Oasis study behind the 87.2% figure and the 40-year Landsat record.
- Integrating PlanetScope time-series and deep learning for monitoring of shrub encroachment in semi-arid steppes ↗The Xilinhot U-Net study with the R2 and leave-one-plot-out numbers.
Watch next
- Look for spatially independent validation of SwinCapT in other arid regions. For degradation tracking, compare classification accuracy with change detection across dates. For shrub encroachment, look for a multi-year change map from the same method.
Sources
- Optimizing swin transformer encoder for spatiotemporal desertification classification in an arid environment (PeerJ Computer Science)peerj.com
- Desertification monitoring in arid oasis environment using Google Earth Engine, machine learning, and field-based hydrogeological assessment (Scientific Reports)nature.com
- Integrating PlanetScope time-series and deep learning for monitoring of shrub encroachment in semi-arid steppes (Planet Publication)publications.planet.com
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 03:32 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →