Advances in AI-based forecasting

Over the past decade, research has demonstrated how deep learning architectures such as convolutional neural networks can effectively downscale coarse reanalysis datasets and correct biases introduced by terrain and land-cover heterogeneity [3]. These models fuse multispectral satellite observations (including soil-moisture and vegetation-health indices) with ground-station measurements and crowd-sourced sensor data to generate forecasts at resolutions finer than five kilometers [4]. Crucially, probabilistic approaches such as Bayesian neural networks and ensemble predictors provide a measure of uncertainty alongside point estimates, enabling farmers to make decisions under risk rather than facing overconfident, deterministic outputs [5].

From forecast to farm-level advisory

Raw probabilistic forecasts, however, must be reframed into clear action triggers. For example, a sub-20 percent probability of rainfall in the next three days might translate into a recommendation to irrigate or postpone planting [6]. Developing these rule-based overlays in collaboration with agronomists ensures that model outputs align with crop calendars, soil characteristics and local cultural practices. To reach diverse farmer populations, advisories should be delivered through multiple modalities: SMS for basic phones; interactive voice response for low-literacy communities; and smartphone apps for richer, graphical content. Fig 1 shows the end-to-end workflow of the proposed AI-driven advisory system, from raw data inputs through to participatory feedback loops. Establishing farmer focus groups to ground-truth forecasts not only builds trust but generates valuable feedback, which can be cycled back into model training to enhance both accuracy and relevance over time [7].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic overview of the AI-Driven Climate Advisory System, illustrating the end-to-end process data inputs → AI downscaling → probabilistic forecasts → rule-based advisories → delivery (SMS/IVR/app) → feedback loop.

https://doi.org/10.1371/journal.pclm.0000675.g001

Institutional and policy imperatives

Scaling AI-driven forecasts requires sustained public investment. We recommend governments and funders support a transparent, open-source forecasting platform including model code, APIs, and training modules to prevent vendor lock-in and ensure methodological rigor. Agricultural extension networks, trusted by rural communities, provide an ideal deployment channel. Integrating AI tools into extension curricula and training district-level “digital champions” will promote local ownership and sustainability. Equally critical is establishing clear data governance frameworks that uphold privacy, ownership, and interoperability, with adherence to FAIR principles to align with global standards and build trust among data providers [8].

Inclusive and affordable access to AI tools

Despite the promise of AI, there is a risk of exacerbating existing inequalities if data-desert regions lacking sufficient ground sensors or historical records are left unaddressed. To mitigate this, model training datasets must proactively sample diverse agroecological zones and cropping systems [9]. Open-access deployment, minimal or zero user fees and subsidized connectivity for the poorest farmers will help ensure affordability. Finally, investing in local capacity building through data-science fellowships, village-level digital ambassadors and participatory workshops will empower communities to adapt and iterate on AI tools, rather than remain passive consumers of external technologies.

Deploying AI-driven forecasting in resource-constrained extension systems

Agricultural extension systems in many low- and middle-income countries (LMICs) remain chronically underfunded and understaffed, often operating below 50% of recommended capacity, with budgets covering little more than travel and training. Extension agents frequently serve over 2,500 farmers, while spending typically accounts for less than 4% of national agricultural budgets [10]. Structural inequities such as social exclusion and limited resource endowments further restrict service access for marginalized groups; for instance, Krishna et al. (2019) report reduced access due to exclusion from training and meetings. These limitations pose significant barriers to adopting AI-driven forecasting via conventional extension system [11]. Therefore, we propose a multi-tiered deployment model to address these limitations:

1. 1. Centralized AI with Para-Extension Support: A national AI platform, managed by data scientists and agronomists (e.g., within the Ministry of Agriculture), hosts crop and weather models. Para-extension agents such as NGO staff or lead farmers receive advisories via SMS, IVR, or apps, and disseminate information through community platforms like “weather clinics” and self-help group meetings, enabling localized outreach without new infrastructure.
2. 2. Public–Private and NGO Partnerships: Collaborations between government, agritech startups, NGOs, and cooperatives can facilitate affordable platform deployment. Joint hosting through donor funds, CSR investments, or grants reduces reliance on public budgets.
3. 3. Train-the-Trainer Cascade Models: Extension delivery can be made more cost‐effective by adopting “training‐of‐trainers” (ToT) models where trained officers can cascade AI tool knowledge to local farmer leaders.
4. 4. Low-Bandwidth Digital Channels: Advisories delivered via SMS and IVR in local dialects ensure access in remote areas, compatible with feature phones and 2G networks.
5. 5. Cooperative Subscription and Cost-Sharing: Farmer Producer Organizations may co-subscribe to AI services, with minimal fees supported by local governments or grants to sustain maintenance. Lead farmers act as liaisons, easing demands on extension agents.

This approach enhances digital outreach, minimizes dependency on constrained public systems, and ensures inclusive access for smallholders.

Gender equity and inclusive design in AI advisory

Women contribute nearly 45% of the agricultural labor force in LMICs but face barriers to digital advisory services due to sociocultural norms, mobility constraints, lower literacy, and caregiving roles [10]. To ensure gender equity, AI systems must:

• Use gender-representative datasets, including female-headed households, labor roles, and cropping calendars [12].
• Ensure at least 40% women’s participation in governance to guide content relevance and interface design.
• Track sex-disaggregated indicators such as platform usage, cost-per-user, decision-to-action timeframes, and yield or income changes [13].
• Employ inclusive interfaces (e.g., voice-first IVR, local dialects, visual aids) and integrate complementary services such as nutrition, finance, and childcare advisories to align with women’s intersecting needs.

Data governance, literacy, and social inclusion

Digital platforms must ensure inclusive data practices. Cooperative data models, where farmers retain ownership and control over consent, access, and deletion, are recommended. Given weak legislative frameworks in many LMICs, clear policies on data rights and transparency are essential.

To foster adoption, digital literacy must improve. Short community-based workshops led by local youth can teach USSD navigation and IVR use. These sessions should prioritize marginalized castes and women, and be scheduled to accommodate caregiving responsibilities, enhancing both access and participation.

Financing and sustainability

Most LMICs allocate less than 2–5% of agricultural budgets to extension services (World Bank, 2024), making full public financing of AI services unlikely. We recommend a blended model:

1. 1. Government Sponsorship: Small analytic teams manage updates using reallocated extension budgets.
2. 2. Donor Grants: Secure pilot funding from climate resilience facilities.
3. 3. Private Sector Engagement: Collaborate with agritech firms or telecoms for subsidized services and co-branding options.
4. 4. Farmer Co-Payments: After demonstrating impact, introduce tiered fees free for the poorest, minimal fees for commercial users.

Farmer data sharing and ethical considerations

Platforms rely on farmer-supplied data (e.g., yields, planting dates). When benefits are clear, up to 80–90% of smallholders willingly share data [14]. A “data-for-service” model offering tailored advisories, discounts, or microcredit rewards can incentivize participation. Transparent feedback loops, limited reporting frequency, and strict privacy safeguards are essential to build trust.

Ethically, AI platforms must prioritize transparency, equity, and consent. Open-source models, explainable interfaces (e.g., infographics or audio rationales), and inclusive governance are key to preventing digital exclusion and enhancing smallholder empowerment [15].