Building an agentic-ready planetary scale distributed geodata structure
The CoRE stack has taken a novel approach to geospatial programming by providing ready-to-use pre-computed data of various landscape entities – micro-watersheds, waterbodies, forests, agroforestry plantations – organized in nested and connected spatial units, and populated with tons of datapoints about these entities to build a comprehensive place-based social-ecological understanding. Ecologists and landscape planners therefore do not need to worry about running complex geospatial workflows to generate all this data – they can rather just focus on asking the right questions from the data. These datapoints include changes over the years in cropping intensity, water-table levels, health of waterbodies, forests and plantations, welfare fund allocation, with many more under development, and can power very interesting landscape analysis of interactions between different entities. Working out the underlying data structures that can support querying this graph at scale and the programming model for data analysis can transform how we work with geospatial data. With much of this querying slated to be performed by AI-agents, it further opens up new problems in distributed systems to run such a planetary scale data structure off a hybrid federated network of cloud platforms, on-prem supernodes, and local personal nodes.
Data flow and algorithm management for geospatial data processing
Building geospatial datasets and indicators can involve complex data flow pipelines that start with base datasets, on which algorithms including data-driven algorithms like ML models are applied, to produce new downstream datasets, on which further algorithms are applied, and so on. This results in a directed acyclic graph on which data and algorithm version control is critical to implement to manage the datasets. If a base dataset changes or an intermediate algorithm is updated, it should trigger a re-computation in impacted downstream paths. Similarly, some of these datasets are temporal in nature and need to be updated on a regular frequency. A thesis that builds out these data and algorithm standards, generalizes it to operate over the web so that the data flow and algorithm graph can span a distributed system of multiple data hosting and computation nodes, allows scaling to novel data flow and algorithm graphs built by AI agents, and implements scalable versions of many algorithms that can leverage GPUs, will be a very relevant contribution.
Strategic land-use management to cope with a changing climate
Recent research in Earth systems outlines the effect of land-use changes on local climate such as the role that cropping cycles, mosaics of cropping and agroforestry fields, forest restoration of hills on windward vs leeward side, etc. can play in alleviating climate extremes and making adaptation easier. However, being able to run km-scale simulations of different land-use scenarios for all localities is not feasible. A reinforcement-learning based framework that discovers the relationship between land-use and local climate by sampling states optimally can help scale such planning for villages and cities. Further, enriching the simulations with ML-based models trained on historical remote-sensing data, such as several projects on hydrology, forest, and cropping patterns outlined in ongoing work can be generalized to produce outputs in future climate scenarios. Put together, such a framework can help rural communities to plan for land-use changes that enables them to cope better with especially extreme weather events of high temperature, heavy rainfall, and increased drought frequencies and intensities.
Multi-resolution self-supervised learning
Every little piece of data can potentially contribute new information. Hi-res RGB map imagery can help with very good but current land-use classification. Multi-spectral time-series from satellites is at lower spatial resolution but can provide useful information on land-use changes. Other factors like terrain, rainfall, temperature, soil, etc. contribute to unique vegetation classes. Self-surpervised learning through transformers or masked auto-encoder architectures of such multi-resolution data can thus embed both historical and current information about patches of land, which can be helpful in many ways. For example, to understand the characteristics of restoration sites which have been shaped by various historical events to appear as what they are. Or, to build a cropping history profile of a farm plot that can explain the soil health of the plot. Many such applications can be conceived by training models on long time-series of historical data.
Computationalizing a relational model for ecological management
Recent research in especially forest ecological systems has highlighted the complex relationships that exist between different trees, tree species, and other flora and fauna. Many of these relationships have been published in research papers but many are uncovered, however, some may be predictable based on observed species behavior. First, we can use the advances made in LLMs to extract these relationships automatically into a structured database. Second, we can use remote sensing and other data to track forest ecosystems and use these observations to predict new relationships, much like how new chemical discovery is done. Third, these relationships can be used to build better forest restoration plans and go beyond some of the ongoing work in our lab.
Understanding agroecology through extensive instrumentation
Imagine if you could do repeated drone runs over farming plots to observe, right from the time of field preparation, the extent of tilling that was done and ridge and furrow spacing that was maintained, to monitoring different growth stages of the exact time when flowers appeared, fruits and seeds appeared, and finally when harvesting was done, followed by whether post-harvest field treatment like mulching was done. Additionally, if you had sensors that regularly reported the soil moisture, air temperature, soil temperature, etc. plus soil health in terms of acidity, NPK, soil organic carbon, etc. If such data could be collected and matched with remote sensed satellite data then detailed crop growth models can be created and simulated to predict cropping yield from satellite data. This can be invaluable to provide precision advisory to smallholder farmers and help them make crop insurance claims without having to deploy expensive sensors in their fields.