BEGIN:VCALENDAR VERSION:2.0 PRODID:-//pretalx//talks.staging.osgeo.org//foss4g-2024-academic-track//spe aker//MP98AQ BEGIN:VTIMEZONE TZID:-03 BEGIN:STANDARD DTSTART:20000101T000000 RRULE:FREQ=YEARLY;BYMONTH=1 TZNAME:-03 TZOFFSETFROM:-0300 TZOFFSETTO:-0300 END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:pretalx-foss4g-2024-academic-track-NKNMCD@talks.staging.osgeo.org DTSTART;TZID=-03:20241204T161500 DTEND;TZID=-03:20241204T164500 DESCRIPTION:Advances in spatial and spectral resolution in private sector s atellite imagery\, together with geography aware algorithms\, have created new venues for the use of Artificial Intelligence (AI) in geospatial appl ications\, sometimes referred to as AI4Geo. However\, these advancements a re accompanied by significant costs in the procurement of data\, computing resources\, communication infrastructure and human expertise. We describe a case study in central Bali in which we developed multiple AI4Geo approa ches to assist the WISNU foundation\, a Non-Governmental Organization in B ali\, Indonesia\, in their ongoing efforts to manage community resources a nd to perform land mapping across small villages in Bali. \n\nConcepts \nT he concept we explore here is multipath AI4Geo that seeks to find the “b est” approach to AI4Geo for resource constrained environments. The assum ption that larger models are always better does not hold where AI4Geo\, tr ained on data from dominant western institutions\, is applied in the major ity world. Some of the most ambitious AI4Geo models are trained for land c over categories that are mostly of interest to the Northern Hemisphere. Gi ven this imbalance\, we ask how participants from low-resourced environmen ts can best make use of AI4Geo.\n\nMethodology\nBased on field data from a study site in Bali\, Indonesia\, we have developed multiple open source A I4Geo land cover approaches to find the best way to represent agroforestry \, a key indicator of sustainable and robust food production. We compare t he image segmentation results from small models such as Random Forests (RF ) and Support Vector Machines (SVM) with large models such as U-Net and ResNet152 not only along established model performance metrics such as f-s core\, but also in terms of their suitability for use in low-resource cond itions. This generally includes limited ability to collect large data sets \, limited computational infrastructure\, limited AI expertise and limited internet connectivity. We then describe a mixed-method multi-pathway appr oach to produce good AI4Geo results while building capacity for the NGO to continue the integration of AI4Geo into its operations while planning fo r an even more challenging AI4Geo future dominated by large homogenizing A I models. \n\nHere are links to code experiments and instructions on gener ating the required input data for the U-Net model from geospatial shapefil es.\n\nSmall models (RF\, SVM based on the Orfeo library)\nhttps://github. com/realtechsupport/cocktail/tree/main/code \n\nLarge models (Custom desig ned U-Net and SATLAS based ResNet models)\nhttps://github.com/realtechsupp ort/cocktail/tree/main/satlas_test\nhttps://github.com/realtechsupport/coc ktail/blob/main/sandbox/working_model/working_model_inference.ipynb \n\nRe sults\nWhile RF\, SVM and U-Net approaches were all able to detect agrofor estry in 8-band\, 3-meter spatial resolution datasets provided by Planet L abs\, we found that the SVM algorithm was most responsive to the limitatio ns of our dataset while producing useful results that we could verify in t he field. SVM was furthermore painless to update with additional field dat a. Figure 1 summarizes the results from the image segmentation after model training.\n\nWhile U-Net’s f1 accuracy for agroforestry exceeds that of RF and SVM\, it is likely an overestimate of the actual extent of agrofor estry. We believe this to be the case because the U-Net architecture inges ted patches of 16 x 16 pixels\, and these dimensions exceed the size of t he smaller agroforestry plots detected in the field. The choice of the inp ut patches was in turn a function of the dimensions of the U-Net architect ure selected for its ability to minimize loss during training across all l and cover categories. \n\nAs opposed to the three other models listed abov e\, the large ResNet152 model was not trained on data Planet Lab satellite imagery but on Sentinel-2 imagery. Because Sentinel-2 only has a maximum spatial resolution of 10m/pixel it is not able to distinguish small scale landscape features\, agroforestry that typically utilizes small plots in r andom arrangements. While the ResNet algorithm was trained on the largest dataset\, with over 300 distinct labels across 137 classes represented acr oss 64 million images\, the class labels are not tuned to the spectral sig natures of agroforestry and deliver only crude results in our selected stu dy area\, as Figure 2 shows. Moreover\, The ResNet152 model that supports multi spectral Sentinel-2 input has over 80 million trainable parameters\, exceeding our bespoke U-Net model by more than an order of magnitude\, th us making its use more costly.\n\nWhile we have not fine-tuned the Resnet 152 model with our own highest resolution Planet Lab data due to spatial r esolution mismatches\, it seems clear that the effort would exceed the cap acities of our partner organization WISNU. Our dilemma is that the most pr omising large models are unwieldy and not adapted to our land cover condit ions while the smaller models we have end to end control over can be tuned with smaller dataset but run the risk of becoming obsolete in the AI arms race over time\, where larger and more powerful models become standard-be arers. While the agroforestry specific results we observe are characterist ic of our study area and the constraints our project operates under\, the homogenizing forces of large models pose a condition all AI4Geo operations are faced with. For that reason\, the territory of this project is signif icant beyond the immediate results we produce.\n\nOur solution to this dil emma is two-fold. We deploy multi-pathway AI4Geo across various technical complexity levels while retaining agency for local stakeholders. The Wisnu foundation does preliminary studies of Sentinel-2 satellite imagery throu gh the QGIS environment to survey sites and build simple datasets. They th en use QGIS integrated small model approaches such as Random Forests to bu ild baseline segmentation maps of a given area. The research team will the n collect Planet Labs based higher resolution data and use the cocktail su ite of models\, including U-Net\, to deepen the study results. Parallel to this approach\, we together use the SATLAS ResNet models to find synergie s in those results. Across the approaches\, we build land cover analysis r esults that optimize limited resources while producing solid analytical re sults. DTSTAMP:20260513T083944Z LOCATION:Room III SUMMARY:GeoAI in resource-constrained environments. - marc böhlen URL:https://talks.staging.osgeo.org/foss4g-2024-academic-track/talk/NKNMCD/ END:VEVENT END:VCALENDAR