Using Artificial Intelligence to Estimate Evapotranspiration

Evapotranspiration (ET) is the combined loss of water to the

atmosphere through evaporation from soil and water surfaces and

transpiration from plants. Because ET governs how water moves between

land and air, estimating it accurately is central to hydrology and

agriculture. It underpins water-balance studies, rainfall–runoff

modeling, ecosystem modeling, irrigation scheduling, and

water-resources planning. In water-limited regions, ET estimates are

especially important for deficit irrigation, where farmers

intentionally apply less water than full crop demand and must

carefully manage the resulting stress. ET also influences decisions

about crop selection and cropping patterns by helping evaluate whether

local weather conditions can support a particular crop.

Why ET is hard to estimate

ET is a complex, dynamic, and non-linear process. It depends on

interacting meteorological variables (such as radiation, temperature,

humidity, and wind) and land-surface conditions (such as crop type and

growth stage). A major challenge is that we often lack a complete

physical understanding of how these factors interact under real-world

conditions, which makes precise estimation difficult.

The traditional framework: ETo → ETc

A widely used “conventional” approach estimates crop

evapotranspiration (ETc) indirectly. First, one calculates reference

evapotranspiration (ETo)—the ET from a standardized reference surface

(a well-watered, actively growing grass-like crop). The Food and

Agriculture Organization (FAO) defines this reference crop with

specific characteristics (including height, albedo, and canopy

resistance) so ETo represents the atmosphere’s evaporative demand

rather than any specific crop’s biology. Because ETo depends mainly on

climate, it is considered independent of crop type. Then, ETc is

estimated by multiplying ETo by a crop coefficient (Kc) that adjusts

for the crop’s characteristics and growth stage. Kc values are

available in the literature for many crops, which makes the approach

practical for operational water management.

How ETo is measured and estimated

ETo can be measured directly using methods such as lysimeters, eddy

covariance, and energy-balance approaches. Lysimeters are among the

most accurate tools because they directly measure water loss from a

controlled soil–plant system. However, lysimeters and other advanced

instruments are often expensive, labor-intensive, and not widely

available, which limits their use in many regions.

As a result, ETo is commonly estimated indirectly using empirical

equations based on weather data. Many methods have been proposed over

decades (e.g., Thornthwaite, Blaney–Criddle, Priestley–Taylor,

Hargreaves–Samani), along with combination approaches that incorporate

both energy and aerodynamic factors. A persistent drawback of many

empirical methods is that their performance can vary considerably by

location and climate, often requiring local calibration, which reduces

their general applicability.

The standard method: FAO-56 Penman–Monteith

The FAO-56 Penman–Monteith (FAO-56 PM) equation is widely accepted as

the standard for estimating ETo because it combines physical

considerations of energy availability and aerodynamic transport, and

it has been tested and calibrated against lysimeter observations in

diverse climates. It is generally preferred because it can be applied

across regions without the same degree of local tuning needed by many

simpler empirical formulas. Its main limitation is practical: FAO-56

PM typically requires a relatively complete set of meteorological

inputs, which may be missing or unreliable in data-scarce areas,

particularly in parts of the developing world.

Over the last two decades, artificial intelligence (AI) has become a

prominent alternative for modeling ETo. AI methods can learn

relationships between inputs and outputs from data without explicitly

coding the underlying physics, making them well-suited to processes

like ET that are non-linear and governed by interacting factors.

Importantly, many studies find AI models can provide reasonable

accuracy even when weather data are limited, which addresses one of

the biggest constraints of FAO-56 PM in practice.

A broad set of AI approaches—such as neural networks (NNs), support

vector machines (SVM), and adaptive neuro-fuzzy inference systems

(ANFIS)—has been applied not only to ETo prediction but also to many

hydrologic tasks including streamflow forecasting, flood prediction,

rainfall–runoff modeling, drought forecasting, groundwater level

prediction, water quality modeling, and pan evaporation estimation.

Within the specific domain of ETo, numerous studies report that

AI-based models can compete with, and sometimes outperform,

traditional empirical methods and even serve as viable substitutes

when complete meteorological datasets are unavailable.

In summary, evapotranspiration is foundational to water management and

agricultural decision-making, yet difficult to estimate due to its

complex drivers. While FAO-56 Penman–Monteith remains the

gold-standard method, practical data limitations motivate the growing

use of AI approaches. The expanding body of AI research has created a

need for structured synthesis—both to guide practitioners in selecting

appropriate methods and to define the next steps for improving ETo

prediction in diverse climates and data conditions.