Meet the IFS officer using AI to tell farmers what trees to plant and where

His initiatives, WhatToPlant and WhereToPlant, are Telegram chatbots that help farmers and forest rangers make informed decisions about what trees to plant and where to plant them.

Rana holds a doctorate in geography from the University of Illinois at Urbana-Champaign and has also worked as a postdoctoral researcher at the University of North Carolina at Chapel Hill.

Pushpendra Rana spoke to indianexpress.com on the need for tech to improve forest governance, help farmers and forest officials reduce wastage, and determine where to plant and what saplings to plant using tech. Edited excerpts:

Venkatesh Kannaiah: Tell us about your research interests and your journey towards the IFS.

Dr Pushpendra Rana: My research lies at the intersection of ecology, public policy, geospatial science, artificial intelligence and environmental governance.

I am interested in how governments can make better decisions in complex social-ecological systems, where ecological conditions, human behaviour and institutions are continuously interacting and changing.

I joined the Indian Forest Service because I wanted to work on questions of forests, landscapes and communities. Field experience, however, showed me that many environmental programmes operate with limited evidence. We know how many seedlings were planted or how much money was spent, but we do not know whether the right intervention was undertaken at the right place, whether it survived, and whether it produced lasting ecological and social benefits. This motivated me to pursue doctoral research in geography and GIScience.

Since then, I have tried to combine the perspectives of a forest officer, a researcher, and a policymaker.

My broader research interest is in how science, institutions and technology can be combined to make environmental governance more evidence-based and accountable.

Venkatesh Kannaiah: Tell us about the WhereToPlant initiative and its impact.

Dr Pushpendra Rana: WhereToPlant began with a simple question: where should we undertake ecological restoration?

Every year, substantial public resources are invested in plantations and targets are set in terms of hectares covered or seedlings planted, while the ecological suitability of the selected land receives less attention. Planting at an unsuitable site can result in low survival, wasted investment and, in some cases, ecological harm.

WhereToPlant was developed as an ecological decision-support system to address this problem. It combines information on climate, soil, terrain, existing vegetation, accessibility, disturbance and other landscape characteristics. Machine-learning models use these data to assess whether a location is suitable for restoration.

The innovation is not only the prediction model. We placed the analysis behind a simple conversational interface, as a Telegram chatbot, so that a forest ranger, a field officer, or a youth club can access the guidance without needing expertise in GIS.

We are asked why not a mobile app? People are overwhelmed with apps, and every department has its own app. So we decided not to build another one because we felt this was the best approach. We wanted to do it on WhatsApp, but WhatsApp charges for this kind of service, so we chose Telegram.

During the 2025 plantation season, the Himachal Pradesh Forest Department undertook plantations over approximately 3,000 hectares, and WhereToPlant informed site selection for more than 500 hectares at over 250 locations.

Venkatesh Kannaiah: Tell us about your learnings from WhereToPlant?

Dr Pushpendra Rana: First, the quality of the question is more important than the sophistication of the algorithm. A highly accurate model answering the wrong question is of little value. We therefore began with the operational decisions faced by field staff rather than beginning with a particular AI/ML technique.

To explain further, four years ago, working with IIT, we developed a plantation monitoring system. The idea was to map all the plantations in the state and build a database that could assist forest rangers. We also wanted to use AI or machine learning on that database to suggest possible locations for future plantations.

What we learned from that effort was that our question was wrong. We were trying to work from the top down, using historical data and existing records. It was more about generating recommendations from the database. But later we realised that we needed to approach it from the bottom up.

The real question should have been: What guidance can a forest guard receive on the ground, at the point of decision-making, using AI?

Second, ecological data are fundamentally contextual. A model trained in one ecological zone cannot automatically be transferred to another. Himalayan landscapes change sharply across elevation, rainfall, slope, aspect and human use.

Third, even a good prediction will not be used unless the recommendation fits departmental workflows, budgets, planting schedules and community priorities.

Fourth, field officers are more likely to engage with a recommendation when the system explains the decision. This is particularly important when machine recommendations differ from field experience.

Fifth, AI/ML should be treated as a participant in an adaptive learning process. Communities and field officers may accept, modify, or reject a recommendation. Those responses should be documented and fed back into the system. In ecological governance, disagreement with the model can itself be valuable data.

Sixth, identifying which tree species should be planted at a given location is just as important as identifying where to plant. Without species recommendations, AI/ML-based site suitability assessments remain incomplete. This is why we are now integrating a WhatToPlant module into WhereToPlant, and we expect this functionality to be available soon.

Venkatesh Kannaiah: Tell us about WhatToPlant and how it would work.

Dr Pushpendra Rana: WhereToPlant identifies where ecological restoration is most appropriate, while WhatToPlant identifies the native tree species or combinations of species most likely to thrive at a given location. The beta version of WhatToPlant is ready and will be released soon.

The database underpinning WhatToPlant has been developed from more than 226,000 geo-referenced field observations of naturally occurring trees collected across various regions of Himachal Pradesh.

At each location, surveys record tree species alongside 29 environmental predictors, including elevation, slope, terrain attributes, and 19 bioclimatic variables describing long-term temperature and rainfall patterns. This work has been undertaken in collaboration with partners like the University of Minnesota, the Indian School of Business, REST NGO, and scientists associated with CSIR-Institute of Himalayan Bioresource Technology.

These field observations are integrated with environmental data to model the ecological niche of each species. The system predicts habitat suitability for 98 native tree species across every 265-metre grid cell in Himachal Pradesh. What we’ve done is create a wall-to-wall map covering the entire state of Himachal. It’s divided into about 795,000 grid tiles, and each tile is seven hectares. It is a wall-to-wall prediction system for the entire state of Himachal.

When a user opens WhatToPlant anywhere in the state, the system instantly recommends the tree species and species combinations most likely to thrive under local ecological conditions.

Venkatesh Kannaiah: How much can AI/ML be used for predicting ecosystem outcomes?

Dr Pushpendra Rana: AI and machine learning can be very useful, but their predictive boundaries must be clearly understood.

These models work best when there is reliable data and when environmental conditions follow consistent patterns. They can help identify suitable habitats for different species, find the best places for restoration, detect forest loss, fires, and environmental damage, and predict short-term risks such as wildfires. It can also classify vegetation using satellite or drone images, estimate the chances of newly planted trees surviving, and identify ecological patterns across large landscapes.

However, ecosystem outcomes are not generated by climate and soil alone. They are also affected by grazing, fire, invasive species, community participation and extreme weather.

AI/ML may tell us that a site is biophysically suitable. It cannot guarantee that seedlings will be protected, watered, or maintained, nor can it fully predict unprecedented climate conditions.

Venkatesh Kannaiah: Tell us about technologies that are changing the way we think in the field of ecosystem restoration.

Dr Pushpendra Rana: There is satellite imagery that is transforming ecological restoration from periodic field assessments to continuous landscape monitoring. We can detect changes in canopy cover, vegetation health, fire, soil moisture, land degradation, and restoration progress.

Platforms such as Global Nature Watch integrate earth observation data to provide high-resolution information for restoration planning and monitoring. We can not only monitor how many trees were planted, but also assess whether ecosystems are actually recovering.

There are integrated digital platforms that are bringing together ecological data, remote sensing, and project monitoring into a single decision-support system.

Drones are increasingly used for high-resolution mapping, seed dispersal, wildfire and mangrove monitoring, wildlife surveys, and post-restoration assessment.

Venkatesh Kannaiah: As AI enters the conservation space, how do we ensure that local knowledge complements machine intelligence rather than being replaced by it?

Dr Pushpendra Rana: First, tech should address a problem identified jointly with local users. Communities should not be brought in only after the tool has already been designed.

Second, recommendations should be explainable. A system should communicate why it considers a site or species suitable. This allows local users to compare the recommendation with observations of frost, grazing, water availability, wildlife, or soil conditions that may not appear in the dataset.

Third, communities must retain the ability to modify or reject a recommendation, and local responses should become part of the learning system. When a community rejects a recommendation, the reason should be recorded. Perhaps the site experiences seasonal waterlogging, or a recommended species is vulnerable to livestock, or the land has a cultural use invisible to satellite data.

The distribution of power and benefits must be considered. Who owns the data? Who decides the objective of restoration? Who is accountable when the recommendation fails? Who receives the benefits? These are governance questions that cannot be solved technically.

Venkatesh Kannaiah: Tell us about tech interventions that have worked in the ecosystem restoration, climate adaptation and biodiversity space in India.

Dr Pushpendra Rana: First, the Forest Survey of India’s satellite-based forest fire alert system has enabled state forest departments to receive near-real-time information on potential forest fires.

Second, India’s advances in cyclone forecasting, early warning, communication, and evacuation represent one of the strongest examples of technology embedded within effective institutions.

Third, camera traps have revolutionised wildlife monitoring, particularly for wide-ranging and elusive species. Combined with spatial capture-recapture methods and AI-based image classification, they enable estimates of species abundance, movement, occupancy, and habitat use.

Fourth, remote sensing has become an indispensable tool for monitoring changes in forests, mangroves, wetlands, and coastal ecosystems.

Finally, a growing range of digital decision-support tools is translating complex scientific information into practical guidance for farmers, forest managers, and local communities.

Venkatesh Kannaiah: Tell us about some tech interventions in the same space that have failed.

Dr Pushpendra Rana: I would not describe these tech interventions as failures. They show how promising technologies fail to realise their potential.

Drone-based tree planting without adequate ecological follow-up is unlikely to deliver meaningful restoration. Dispersing seeds is only one step in the restoration process. Long-term success depends on appropriate species selection, suitable soil and climate conditions, grazing management, fire control, invasive species management, and sustained institutional and community support.

Many digital dashboards focus on administrative indicators such as seedlings distributed, funds spent, or geo-tagged photographs uploaded. While these systems improve programme monitoring, they can create an illusion of success if they do not measure ecological outcomes.

Globally trained species distribution or carbon models are valuable for broad-scale planning and initial screening. However, they may perform poorly when used as precise local prescriptions because they often cannot capture fine-scale variation in soils or local ecological knowledge.

AI, machine learning, and remote sensing applications are technically sophisticated but see limited long-term adoption once pilot funding ends.

Finally, technology should support better ecological decisions rather than simply increase administrative outputs.