The research team, led by Ge Li and Weibo Nie of Xi’an University of Technology, focused on the forage belt of the Yellow River Basin within the Inner Mongolia Autonomous Region. The area stretches from 106°36′E to 112°04′E, spans elevations of 800 to 2400 meters, and receives a meager 120 to 420 millimeters of annual rainfall against potential evapotranspiration exceeding 2000 millimeters. Agriculture there consumes roughly 4.5 billion cubic meters of irrigation water each year, mostly through traditional surface irrigation that frequently drives water and dissolved nitrogen deep beyond the root zone. Excess nitrogen application, meanwhile, has contributed to eutrophication of surface waters and nitrate contamination of groundwater. The stakes for smarter scheduling could hardly be higher.

At the heart of the framework lies a classic problem in crop modeling: parameters calibrated at a few experimental sites rarely transfer well to a heterogeneous landscape. The researchers tackled this by assimilating remotely sensed observations into the DSSAT-CERES-Maize model. They chose two state variables with complementary physiological meaning. Leaf area index, or LAI, tracks canopy expansion and light interception, while aboveground biomass, AGB, reflects the cumulative outcome of photosynthesis and dry matter partitioning. Because different cultivar parameters govern these distinct processes, constraining both simultaneously reduces parameter equifinality, the situation in which multiple parameter combinations produce similar model outputs and leave calibration ambiguous.

Getting reliable LAI and AGB estimates across a fragmented agricultural landscape posed its own challenge. Optical sensors such as Sentinel-2 suffer from cloud contamination and saturate when LAI exceeds about 3, so the team turned to Sentinel-1 synthetic aperture radar, whose C-band microwave backscatter penetrates cloud cover and offers a six-day revisit cycle at roughly 10-meter resolution. Using a random forest algorithm trained on field measurements from 2024 and 2025, they first mapped silage maize planting areas with an overall accuracy of 90.34 percent and a Kappa coefficient of 0.803, then estimated LAI with a testing-set coefficient of determination of 0.923 and AGB with 0.691. A bootstrap ensemble of thirty random forest models quantified retrieval uncertainty, which proved modest for LAI, at a coefficient of variation of 5.2 percent, but larger for AGB, at 18.2 percent, reflecting the more complex nonlinear relationship between radar signals and biomass.

The assimilation engine was the Shuffled Complex Evolution algorithm, SCE-UA, a population-based global optimizer that iteratively improves the worst-performing members of a parameter population. Six cultivar parameters, P1, P2, P5, G2, G3, and PHINT, which together control phenology, kernel number, kernel filling rate, and leaf appearance intervals, were optimized against the satellite-derived observations. Because a single parameter set cannot represent an entire basin, the researchers partitioned the study area into three sub-regions using k-means clustering on 51 years of growing-season precipitation records from 32 meteorological stations, with the average silhouette coefficient confirming that three clusters best balanced internal compactness against separation. Three assimilation schemes were tested: LAI only, AGB only, and the joint assimilation of both variables.

The results were decisive. When field-calibrated parameters were applied directly across the region, the model’s agreement with 90 independent validation sites was mediocre, with R² values of 0.53 for LAI and 0.672 for AGB. Joint LAI-AGB assimilation lifted those figures to 0.883 and 0.872, cutting the mean absolute error for LAI from 0.81 to 0.34 and for AGB from 2443 to 1322 kilograms per hectare. Statistical tests confirmed the improvement was significant rather than an artifact of the metrics, with p-values below 0.05 against both single-variable schemes. Monte Carlo perturbation of the satellite observations showed that the optimized parameters remained stable, with coefficients of variation mostly below 20 percent, and the ensemble mean parameter sets were adopted for regional simulation.

With a trustworthy regional model in hand, the team turned to scheduling. They classified typical rainfall years using empirical frequency analysis of 1975–2025 precipitation, designating years at the 25th, 50th, and 75th percentiles as wet, normal, and dry. The NSGA-III multi-objective genetic algorithm then searched for irrigation and nitrogen schedules that simultaneously maximized yield, water use efficiency, and nitrogen fertilizer partial productivity, with constraints preventing either efficiency metric from falling below 75 percent of its achievable maximum. Decision variables included the number of irrigation events, their timing and amounts, and the timing and rate of nitrogen applications, encoded as fifteen optimization variables that MATLAB scripts translated into DSSAT-compatible experiment files. The TOPSIS decision method then selected the best compromise schedule from each Pareto front.

The optimized schedules delivered substantial savings. Across wet, normal, and dry years, irrigation water use fell by 9.4 to 21.4 percent, equivalent to 37 to 101 millimeters, while nitrogen inputs dropped by 8.4 to 24.7 percent, or 32 to 90 kilograms per hectare, relative to current farmer practices derived from regional water quotas and field surveys. The savings followed a clear spatial logic: the driest sub-region, SR1, required the most irrigation, receiving 40 to 150 millimeters more than SR2 and SR3 depending on the rainfall scenario, while nitrogen rates showed the opposite pattern. Simulated yields under the optimized schedules reached roughly 22,000 to 26,000 kilograms per hectare across sub-regions, and the gains were most pronounced in dry years, when the area experiencing yield increases expanded to about 22,725 square kilometers while the yield-decrease area shrank to roughly 700 square kilometers.

The study is candid about its limits. The CERES-Maize nitrogen module, built on one-dimensional empirical parameterizations of mineralization, nitrification, and leaching, reproduced soil ammonium and nitrate dynamics poorly, and the tipping-bucket water scheme only moderately captured soil moisture. The framework also assumes uniform sowing dates and densities and cannot reliably distinguish silage maize from grain maize in satellite imagery, since both share similar spectral signatures. The authors propose coupling the model with physically based tools such as HYDRUS and DNDC, integrating multi-source remote sensing including UAV imagery, and adopting Bayesian calibration to better characterize parameter uncertainty in future work.

Even so, the significance of the approach extends well beyond one forage belt. By demonstrating that radar-based remote sensing can regionalize crop model parameters, that joint assimilation of canopy and biomass variables outperforms either alone, and that multi-objective optimization can reconcile productivity with resource conservation, CropINSs offers a template for precision agriculture at scales where field experiments are impractical. As water scarcity intensifies across the Yellow River Basin and nitrogen regulations tighten worldwide, frameworks of this kind may become the difference between sustainable forage production and continued depletion of the resources on which it depends.

Subject of Research: Remote sensing data assimilation with crop models to optimize irrigation and nitrogen scheduling for silage maize

Article Title: CropINSs: A multi-objective irrigation and nitrogen scheduling optimization framework for silage maize based on coupling remote sensing and crop models

Article References: Li, G., Nie, W., Wang, Q., Ji, M., Liu, Y., & Ma, X. (2026). CropINSs: A multi-objective irrigation and nitrogen scheduling optimization framework for silage maize based on coupling remote sensing and crop models. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.10.001

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.10.001

Keywords: silage maize, remote sensing, data assimilation, DSSAT, CERES-Maize, irrigation scheduling, nitrogen management, NSGA-III, random forest, water use efficiency, Yellow River Basin, precision agriculture

The context is urgent. Global installed solar capacity reached approximately 2.39 terawatts in 2025 following a record annual addition of 511 gigawatts, and the COP28 commitment to triple renewable capacity by 2030 will demand even more land. Agrivoltaic systems, in which photovoltaic panels share ground with crops, offer an elegant answer, with partial shading reducing heat stress, improving soil-moisture retention and enhancing water-use efficiency, while dual income streams strengthen farm economics. Yet despite rapid deployment in Germany, France, Italy, Japan and China, Türkiye, despite averaging 2,640 sunshine hours and 1,527 kilowatt-hours per square metre of annual solar irradiation, has seen few practical implementations. The Ayaş pilot was designed to fill that gap under semi-arid Central Anatolian conditions.

What distinguishes the new work from previous digital agrivoltaics efforts is integration. Earlier projects have typically addressed only fragments of the problem: conceptual IoT frameworks without field validation, digital twins focused on panel control but blind to water, low-cost sensors with no decision layer, or fertigation platforms with no photovoltaic management or financial module. The Ayaş platform, by contrast, weaves environmental monitoring, crop observation, irrigation scheduling, photovoltaic operational modules and techno-economic evaluation into a single decision-support environment. The authors state that, to their knowledge, no previous study has reported a field-deployed platform combining all of these functions within one architecture.

The hardware foundation is a layered sensor network. Two meteorological stations track air temperature, humidity, rainfall, pressure, wind and solar irradiance every 15 minutes. Pyranometers measure incoming radiation on both the front and rear faces of the bifacial modules, while photosynthetically active radiation sensors quantify the light actually available to crops under the panels and in the open field. Twelve soil-moisture probes and twelve soil-temperature sensors, distributed between 10 and 60 centimetres deep in both shaded and unshaded plots, feed the irrigation logic, alongside flow meters and solenoid valves on the irrigation lines. An ESP32-based gateway and remote terminal units timestamp every record, and an edge-processing layer validates the data before it reaches the cloud, flagging physically implausible values, duplicates and isolated spikes, and buffering measurements during communication outages rather than fabricating replacements.

The pilot installation itself pairs a 122-kilowatt-peak array of bifacial crystalline-silicon modules, elevated 3.72 metres above ground with a 36 percent ground coverage ratio, against an adjacent 1,613-square-metre open-field reference plot of identical size. Four horticultural crops, table tomato, processing tomato, capia pepper and bell pepper, were transplanted into both zones on 31 May 2025 at identical densities, with more than 5,800 seedlings established in total. Because the two plots shared soil characteristics, irrigation practice and climatic exposure, differences in performance could be attributed to the shading environment rather than to other variables.

The crop results tell a strikingly species-dependent story. Capia pepper proved remarkably tolerant of the agrivoltaic microclimate, yielding 3,064 kilograms per decare against a reference yield of 3,104, a reduction of just 1.3 percent. Bell pepper lost 15.6 percent, while table tomato and processing tomato fell by 23.5 and 34.7 percent respectively, revealing the greater sensitivity of tomatoes to reduced incident radiation. Just as intriguing, the shading extended the harvest window: open-field table tomato production ended on 17 September 2025, whereas plants beneath the panels kept producing until 23 October. Processing tomato and capia pepper showed the same pattern, suggesting that agrivoltaic conditions can stagger and prolong harvests in ways that may carry real market value for growers.

On the water side, the irrigation module fuses root-zone soil moisture, soil temperature and rainfall observations with the FAO-56 Readily Available Water methodology. Crop-specific thresholds were derived from field capacity and permanent wilting point estimates, the latter calculated for Ayaş’s heavy clayey soils using Saxton–Rawls pedotransfer functions. A recommendation to irrigate is issued only when the soil-moisture index drops below the threshold, root-zone temperature sits within the 15 to 30 degree Celsius window suitable for crop development, and no significant rainfall has occurred in the preceding 12 hours, with confirmation required across two consecutive 15-minute measurement intervals. Across five monthly monitoring periods the module generated 15 recommendation days: seven for processing tomato, five for bell pepper, three for capia pepper and none for table tomato, whose soil moisture stayed marginally above threshold. Notably, average soil-water depletion remained around 11 percent of available capacity, well below the FAO-56 allowable limits of 30 to 40 percent, indicating that the site’s conventional fixed schedule was probably over-irrigating, and that demand-based scheduling could deliver substantial water savings.

The economic innovation is arguably the paper’s boldest contribution. Conventional levelized cost of electricity metrics capture only power generation, ignoring the crops that agrivoltaic land also produces. The team therefore proposes the Levelized Cost of the Agrivoltaic System, or LCOAS, which adds agricultural operating costs to the numerator and converts crop revenue into electricity-equivalent energy by dividing it by the prevailing electricity price. Applied to the 2025 season, with a unit capital cost of 1,230 euros per kilowatt-peak totalling 150,060 euros and seasonally allocated through a capital recovery factor, the pilot returned an LCOAS of 0.0846 euros per kilowatt-hour equivalent. Total seasonal income reached 8,093.20 euros, of which 6,080 euros came from electricity and 2,013.20 euros from crops, against 6,850.07 euros in expenses, leaving a seasonal net benefit of 1,243.13 euros and a positive net agricultural profit of 301.98 euros. For a system in its first season, the framework offers operators a single comparable number spanning both enterprises.

Two further modules, a panel-cleaning advisor that recommends washing only when accumulated soiling losses exceed cleaning costs, and a multi-objective panel-angle manager balancing energy yield against crop light needs, are presented as proof-of-concept components with integrated interfaces but were not executed or validated in the field. The authors are candid about this and about the study’s other limits: one site, one season, four crops. Still, the demonstration matters. It shows that a working farm-scale agrivoltaic array can be run through a single digital nervous system, converting streams of raw sensor data into irrigation advice, crop comparisons and integrated economics, and it offers a practical template for the multi-site, multi-season validation that will determine whether agrivoltaics can scale from promising niche to mainstream land-use strategy.

Subject of Research: An IoT-based decision-support platform for managing agrivoltaic solar and crop production

Article Title: An integrated IoT-based decision-support platform for agrivoltaic systems: Design, implementation and pilot-scale field evaluation

Article References: Gümüş, H., Çubukçu, M., Kuzyaka, D., Çolak, N., Yüksel, M., Altındişli, A., Zafer, C., & Özden, T. (2026). An integrated IoT-based decision-support platform for agrivoltaic systems: Design, implementation and pilot-scale field evaluation. Smart Agricultural Technology, 15, Article 102579. https://doi.org/10.1016/j.atech.2026.102579

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102579

Keywords: agrivoltaics, Internet of Things, precision agriculture, photovoltaics, irrigation scheduling, decision-support platform, smart farming, LCOAS, soil moisture sensors, renewable energy, food security, Türkiye

The experiment, conducted across three growing seasons from 2021 to 2024, was designed as a split-plot field trial with nine treatment combinations. Three drip irrigation levels were tested: D1 at 65 percent of field capacity, D2 at 75 percent, and D3 at 85 percent, where field capacity represents the amount of water a soil retains after gravitational drainage. Each irrigation level was crossed with three nitrogen application rates: N1 at 150 kilograms per hectare, N2 at 210 kilograms per hectare, and N3 at 270 kilograms per hectare. Field capacity is a critical anchor in irrigation science, because targeting a percentage of it determines how closely soil moisture is held near the ideal range where roots can extract water with minimal effort while excess drainage and deep percolation of nitrate are minimized.

The headline result is that the middle combination, D2N2, outperformed the other eight treatments on the metrics farmers care about most. Grain yield rose by 1.88 to 17.9 percent relative to the alternatives, and water use efficiency climbed by 7.65 to 14.5 percent. That dual gain is significant because yield and efficiency often pull in opposite directions: pushing more water and fertilizer onto a crop can inflate yields while quietly eroding the return on every additional unit of input. The D2N2 treatment appears to thread the needle, keeping the crop near its yield potential without wasting the resources used to get there.

There was, however, a trade-off the researchers were careful to document. The D2N2 combination reduced nitrogen partial factor productivity, a measure of how many kilograms of grain are produced per kilogram of nitrogen applied. Because D2N2 applies a moderate nitrogen rate while achieving high yields, the ratio shifts in ways that depend on how the metric is calculated relative to other treatments. This kind of tension is precisely why the team framed the work as a multi-objective optimization problem rather than a search for a single best number. In real farming systems, no single metric captures sustainability; the goal is a combination of inputs that performs acceptably across yield, efficiency, and the slower-moving variables that determine whether the soil remains productive for decades.

Stability emerged as another dimension of the story. Across the three seasons, the D2N2, D3N2, and D3N3 treatments promoted higher yield stability and sustainability than the drier or more nitrogen-starved combinations. Yield stability matters enormously in a region where interannual variability in rainfall and temperature can swing harvests dramatically. A treatment that produces a spectacular yield in one favorable year but collapses in a dry one is less valuable than one that delivers consistently strong results. The fact that the moderately wetter irrigation regimes with moderate to high nitrogen supported stability suggests that maintaining adequate soil moisture buffers the crop against the stresses that otherwise amplify year-to-year variation.

Perhaps the most novel contribution of the study is its treatment of soil quality as a first-class objective rather than an afterthought. The researchers constructed a soil quality index, or SQI, that integrates physical, chemical, and biological indicators of soil health. Under D2N2, soil physical quality improved by 3.23 to 18.7 percent and soil biological quality improved by 4.53 to 36.6 percent compared with the other eight treatments. Comprehensively, the SQI under D2N2 was enhanced by 3.98 to 28.9 percent. The mechanisms behind these gains are interconnected: better soil physical structure improves water retention and aeration, which in turn supports nutrient availability and the microbial communities that cycle nitrogen and carbon through the root zone. Drip fertigation, by avoiding the waterlogging and crusting associated with surface flooding, appears to create a soil environment in which those beneficial feedbacks can accumulate over multiple seasons.

To move from individual treatment comparisons to a generalizable prescription, the team applied response surface methodology, a statistical technique that models how a response variable changes across a continuous space of input settings rather than only at the discrete points actually tested. Fitting response surfaces to the yield, efficiency, and soil quality data allowed the researchers to interpolate between treatments and identify the ranges where all objectives were simultaneously satisfied. The result was strikingly specific: an irrigation level of 77 to 85 percent of field capacity, coupled with a nitrogen application rate of 181 to 194 kilograms per hectare. That nitrogen window sits below the highest rate tested, reinforcing a recurring theme in modern agronomy: more fertilizer is not better fertilizer, and the marginal kilogram of nitrogen beyond a crop’s capacity to use it tends to leach away, acidify soil, or escape as greenhouse gas rather than become grain.

The implications for the North China Plain are considerable. The region produces a large share of China’s wheat but has paid for that productivity with some of the fastest-declining aquifers in the world, and with nitrate contamination that traces directly back to over-application of synthetic nitrogen. Drip fertigation has been promoted as a remedy, but adoption has been hampered by uncertainty about how to operate the systems: too little water stresses the crop, too much negates the savings, and the interaction with nitrogen rate is nonlinear. By quantifying those interactions across three full seasons and folding soil health into the optimization, the study gives extension services and growers a defensible target rather than a rule of thumb. The finding that the optimal irrigation range extends up to 85 percent of field capacity is particularly useful, because it suggests farmers can err slightly toward wetter soils without sacrificing efficiency, provided nitrogen is held in the moderate range.

The work also fills a methodological gap that the authors explicitly identify. Previous studies of drip-fertigated winter wheat in the region typically optimized for yield and water or nitrogen use efficiency alone, leaving soil quality out of the objective function. Incorporating the SQI changes the answer. A combination that maximizes short-term yield might degrade soil structure or suppress microbial activity in ways that only become apparent years later, so including soil indicators in the optimization shifts the recommended operating point toward regimes that sustain the resource base. This is a quiet but consequential reframing: it treats soil not as an inert growing medium but as a living system whose condition is part of the farm’s output.

Caveats remain, as they do in any field study. The results come from a specific soil, climate, and management context in the North China Plain, and the optimal ranges may shift elsewhere or under different cultivars and residue practices. The three-season span, while substantial for a field trial, is still short relative to the timescales over which soil quality evolves, so longer-term monitoring will be needed to confirm that the SQI gains persist. The study was supported by China’s National Key R&D Program, the S&T Program of Hebei, and the State Key Laboratory of North China Crop Improvement and Regulation, and it is published open access, making the detailed data available to researchers worldwide. For a region where every cubic meter of water and every kilogram of nitrogen is contested, the message is clear and actionable: irrigate near the upper-middle of the soil moisture range, feed the crop moderately rather than generously, and the payoff arrives not just in this season’s grain but in the ground beneath it.

Subject of Research: Multi-objective optimization of subsurface drip fertigation for winter wheat in the North China Plain

Article Title: Multi-objective optimization of subsurface drip fertigation for winter wheat in the North China Plain: balancing grain yield, resource use efficiency, and soil quality

Article References: Wang, B., Wang, P., Zhang, Y., Yang, J., Li, L., He, F., & Wang, G. (2026). Multi-objective optimization of subsurface drip fertigation for winter wheat in the North China Plain: balancing grain yield, resource use efficiency, and soil quality. BMC Plant Biology. https://doi.org/10.1186/s12870-026-10067-x

Image Credits: AI Generated

DOI: 10.1186/s12870-026-10067-x

Keywords: drip fertigation, winter wheat, North China Plain, soil quality index, water use efficiency, nitrogen management, multi-objective optimization, response surface methodology, sustainable agriculture, irrigation scheduling, soil microbial activity, grain yield

The research tackles a problem that has quietly plagued evapotranspiration modeling for decades. The gold standard for computing ETo, the FAO-56 Penman-Monteith equation, requires a full suite of meteorological measurements including air temperature, solar radiation, wind speed, and humidity. Many weather stations around the world, particularly in developing regions, simply do not record all of these variables reliably. Machine learning models have emerged as powerful alternatives because they can learn the complex, non-linear relationships between whatever meteorological inputs are available and the evapotranspiration rates observed at a site. Yet feeding a model every possible variable is not always wise: redundant or irrelevant inputs inflate model complexity, slow computation, and can degrade accuracy through noise. The question of which variables to keep, and which to discard, is known as feature selection, and it is precisely where the new study makes its mark.

The researchers built a hybrid framework that couples seven distinct feature selection methods with an Extreme Gradient Boosting model, better known as XGBoost. XGBoost was chosen deliberately for its proven track record with tabular meteorological datasets, its built-in regularization that helps prevent overfitting, its native ability to capture non-linear interactions between variables, and its computational efficiency on moderately sized station records. The seven selection techniques span the two major families of feature selection. Filter-based methods, which evaluate variables independently of any predictive model, included the Pearson correlation coefficient, the minimum redundancy-maximum relevance criterion known as mRMR, the ReliefF algorithm, the Fisher score, and the Stability method. Embedded methods, which perform selection as part of model training, included LASSO regression and SHAP, the SHapley Additive exPlanations approach borrowed from the field of interpretable machine learning.

Each of these techniques interrogates the data in a fundamentally different way. Pearson correlation measures simple linear association between each meteorological variable and ETo. The mRMR criterion, rooted in information theory, seeks variables that are highly relevant to the target while being minimally redundant with one another, avoiding the trap of selecting several measurements that all encode the same information. ReliefF evaluates how well a variable distinguishes between neighboring instances that are similar or dissimilar in the target value, making it sensitive to feature interactions. The Fisher score assesses class separability in the feature space, while the Stability method quantifies how consistently a selection algorithm picks the same variables across resampled datasets, a property that matters enormously when models must be trusted on new data. LASSO, by contrast, shrinks regression coefficients toward zero and effectively eliminates weak predictors through regularization, and SHAP derives feature importance from game-theoretic principles by computing each variable’s marginal contribution to model predictions.

To test the framework under genuinely different atmospheric regimes, the team used data from two contrasting climatic conditions, an arid region and a humid region, with records obtained from the Islamic Republic of Iran Meteorological Organization. The study was structured in two modeling stages. In the first stage, the three most important features identified by each selection technique were fed into XGBoost as inputs. In the second, more aggressive stage, only the two most important features were used. This staged design functions as a feature-based domain adaptation procedure, systematically evaluating the relevance and redundancy of key meteorological variables while minimizing the input space, which in turn reduces model complexity and computational cost. The practical payoff is a set of modeling tools suited to data-scarce conditions, where reliable measurements are limited and every sensor counts.

The results are striking. In the arid region, the best-performing combinations achieved a coefficient of determination of at least 0.980, a root mean square error no greater than 0.344 millimeters per day, and a mean absolute error of at most 0.270 millimeters per day. The humid region performed even better, with R-squared values of at least 0.982, RMSE as low as 0.188 millimeters per day, and MAE down to 0.142 millimeters per day. For a variable that typically ranges across several millimeters per day seasonally, sub-0.2 millimeter daily errors represent an extraordinary level of precision, comparable to the uncertainty of the reference equation itself.

Perhaps the most actionable finding is that the optimal feature selection method depends on climate. In the arid region, ReliefF and LASSO delivered the lowest error indices when only the two most important features were used. In the humid region, the Pearson, SHAP, Fisher, and Stability methods produced the best results with the same two-feature approach. This divergence makes physical sense. In arid environments, where atmospheric demand is driven largely by energy availability and vapor pressure gradients, instance-based and regularization-based selectors appear to isolate the dominant drivers most effectively. In humid conditions, where radiation and temperature interplay differently and air moisture is less limiting, correlation-based and interpretability-driven methods capture the signal more cleanly. The lesson for practitioners is that feature selection is not a one-size-fits-all exercise; the choice of selector should be matched to the climatic context of the station being modeled.

The authors did not stop at raw accuracy figures. They subjected their models to bootstrap analysis, effect size assessment, and decision boundary analysis to evaluate robustness and statistical significance rather than relying on single-point error metrics alone. Bootstrap resampling tests whether performance holds across repeated random draws from the data, guarding against lucky splits. Effect size analysis quantifies the magnitude of differences between competing models, ensuring that reported improvements are meaningful rather than trivially small. Decision boundary analysis probes how models behave across the full range of input conditions, revealing whether accuracy is uniform or concentrated in particular regimes. Together, these diagnostics lend the headline numbers a level of statistical credibility that many applied machine learning studies lack.

As a final benchmark, the XGBoost models were compared against Random Forest and a feed-forward Artificial Neural Network, two of the most widely used alternatives in hydrological machine learning. XGBoost came out on top, confirming that its gradient-boosted tree architecture, with its regularization and efficient handling of non-linear interactions, is particularly well suited to the structure of daily meteorological records. This finding adds to a growing body of evidence that gradient boosting is a formidable default choice for tabular environmental data, often outperforming deep networks on datasets of modest size.

The implications reach well beyond Iranian weather stations. As climate change intensifies pressure on freshwater supplies, water managers in arid and semi-arid regions worldwide face the twin challenges of rising crop water demand and sparse monitoring infrastructure. A modeling recipe that achieves R-squared values above 0.98 using only two meteorological inputs means that stations lacking full instrumentation can still produce trustworthy evapotranspiration estimates for irrigation scheduling and drought assessment. The framework’s emphasis on stability and interpretability, through methods like SHAP and the Stability criterion, also aligns with the broader push for transparent artificial intelligence in environmental science, where black-box predictions alone are no longer sufficient for policy decisions. By demonstrating that the marriage of thoughtful feature selection and gradient boosting can squeeze remarkable accuracy from minimal data, the study offers a practical template for the next generation of evapotranspiration models, one that could help stretch every drop of information as far as every drop of water.

Subject of Research: Machine learning-based feature selection for estimating daily reference evapotranspiration from meteorological data

Article Title: Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data

Article References: Sadeghzadeh, M., Karimi, S., Shiri, J., & Sivakumar, B. (2026). Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data. Theoretical and Applied Climatology, 157(10), Article 629. https://doi.org/10.1007/s00704-026-06568-2

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06568-2

Keywords: reference evapotranspiration, feature selection, XGBoost, machine learning, meteorological data, ReliefF, LASSO, SHAP, arid climate, humid climate, irrigation scheduling, Theoretical and Applied Climatology

The study, published in the open-access journal Discover Soil, addresses a fundamental quantity in agricultural hydrology known as soil moisture retention. This property describes the relationship between the volume of water held in soil and the suction with which that water is bound to soil particles. Two reference points matter most. Field capacity marks the maximum water a soil retains after gravity has drained away the excess, conventionally measured at a pressure of −0.33 bar. The permanent wilting point marks the dryness at which plant roots can no longer extract water fast enough to survive, measured at −15 bar. The difference between the two is available water, the share of soil moisture genuinely accessible to crops, and it is the number on which irrigation schedules, hydrological models, and crop choices ultimately depend.

Measuring these constants traditionally requires a pressure plate apparatus, a device in which saturated soil samples sit on porous ceramic plates inside an airtight chamber. Increasing air pressure forces pore water through the plate until equilibrium is reached, a process that takes roughly a week across the full range of suction levels. The method is reliable but slow, costly, and demanding of technical expertise, which is precisely why characterisations of field capacity and permanent wilting point remain unavailable for many of the world’s agricultural regions. Punjab Agricultural University’s soil testing laboratory receives samples from farmers across the state, yet the sheer scale of demand far outstrips what pressure plate analysis can deliver.

The researchers, led by Swati Kashyap together with Bharat Bhushan Vashisht, Harsimran Kaur and Mohit Arora, took the modelling route instead. They assembled nine well-known pedotransfer functions, equations that translate easily measured soil properties such as sand, silt, clay content, soil organic carbon and bulk density into estimates of water retention. Pedotransfer functions were first proposed in the late 1980s as a way of adding value to routine soil survey data, and they have since been tested from the Congo basin to the Mekong Delta. The catch is that an equation calibrated on Ugandan ferrallitic soils or Brazilian Amazonian clays does not necessarily perform well on the alluvial sandy loams of north-western India, so local evaluation is essential.

To provide that evaluation, the team collected around 200 surface soil samples from 0 to 15 centimetres depth across Punjab’s different agroclimatic zones, drawing on farmer submissions held by the university’s Soil Testing Laboratory and on research fields. Seventy-eight samples were used to calibrate the candidate equations and forty independent samples were reserved for validation. Each sample was analysed for particle size distribution by the pipette method, organic carbon by wet oxidation, and water retention by the pressure plate apparatus itself, giving the researchers ground truth against which every model prediction could be scored. Performance was judged with three statistical indicators: root mean square error, which penalises large deviations; the index of agreement, which ranges from zero to one; and mean absolute error, which measures average prediction offset.

The results were strikingly clear. For field capacity, an equation published by J. D. Pidgeon in 1972 for ferrallitic soils in Uganda outperformed the field, achieving a root mean square error of 0.05 cubic centimetres of water per cubic centimetre of soil on validation, an index of agreement of 0.72 and a mean absolute error of 0.049. For the permanent wilting point, the 1979 equation of S. Gupta and W. E. Larson, built on particle size distribution, organic matter and bulk density, proved best, with an RMSE of 0.048, an index of agreement of 0.77 and a mean absolute error of 0.041. Critically, the researchers found that raw application of these imported equations systematically over- or under-estimated water contents. By adding a simple bias correction factor derived from the calibration data, −0.018 for the Pidgeon model at field capacity and +0.008 for the Gupta–Larson model at the wilting point, prediction accuracy improved markedly, shifting predicted values visibly closer to the one-to-one line when plotted against observations.

The physics behind the correlations is instructive. Silt, clay and organic carbon all correlated positively with water content at field capacity, while sand content correlated negatively with both constants. Clay governed retention at the wilting point more strongly than at field capacity, whereas organic carbon mattered more at field capacity. This makes sense because water held at low suction depends on the architecture of pore spaces, which organic matter helps build, while water held near the dry end is governed by adsorption forces on particle surfaces, a function of texture. Bulk density, meanwhile, increased retention at −15 bar, echoing earlier Indian studies on the influence of compaction on dry-end moisture.

With validated equations in hand, the team scaled up. Using soil maps covering 520 pedons, the basic mapping units of soil classification, compiled by the Department of Soil Science at Punjab Agricultural University, they extracted sand, silt, clay and organic carbon values for every pedon and predicted field capacity and permanent wilting point state-wide. Texture-specific bulk density values, ranging from 1.70 grams per cubic centimetre for sandy soils to 1.30 for clay loams, completed the input set. The predictions were then loaded into QGIS, the open-source geographic information system, and symbolised in graduated classes across four agroclimatic zones: the sub-mountain undulating region, the undulating alluvial plain, the central plain and the western alluvial plain.

The maps reveal a state with substantial but uneven water-holding wealth. Field capacity across Punjab soils ranges from 0.131 to 0.387 cubic centimetres per cubic centimetre, with roughly two-thirds of the land falling in a good band of 0.200 to 0.300. Permanent wilting point values span 0.009 to 0.228, with about 65 percent of soils in the 0.050 to 0.150 interval. Available water ranges from 0.113 to 0.183, and fully 93 percent of the state sits in the 0.120 to 0.160 band, a limited-to-good status in which ideal conditions are notably absent. The driest retention profiles appear in the arid western zone, where sandy loam and loamy sand textures combine with low organic carbon and clay. Intriguingly, the finest-textured clay loams, despite holding the most total water, show reduced availability, because water molecules bond tightly to negatively charged clay surfaces and resist extraction by roots.

For a state where rice and wheat consume some 61 percent of total water demand and unregulated groundwater extraction has created genuine scarcity, the practical implications are considerable. A farmer or irrigation planner equipped with these maps and a basic soil test can now estimate plant-available water for a specific field without ever touching a pressure plate, and schedule irrigation to match what the soil can actually store. The authors suggest the calibrated equations could be extended under different management systems for crop-specific water budgeting under a changing climate. More broadly, the study is a demonstration of a quiet but powerful idea in soil science: that the right simple model, rigorously calibrated and validated against local ground truth, can democratise information that expensive instruments have long reserved for a privileged few. In Punjab’s water-stressed fields, that democratisation may arrive just in time.

It is worth noting that the predictive skill reported in the study, while respectable, still leaves room for uncertainty. An index of agreement near 0.75 indicates that the calibrated equations capture the broad pattern of retention across Punjab’s soils but not every local deviation, so the mapped values are best treated as planning-grade estimates rather than substitutes for direct measurement where high-stakes decisions depend on precise water budgets.

The regional context also matters. Punjab’s soils are dominated by Inceptisols and Entisols developed on alluvial plains under a hyperthermic temperature regime, with annual rainfall between 400 and 1300 millimetres concentrated in the July-to-September monsoon. In such settings, sandy loam textures prevail, and coarse particles paired with low organic carbon naturally depress both field capacity and wilting point, which is consistent with the drier retention profiles the maps show in the arid western zone.

The approach also fits a wider trend in soil science toward digital soil mapping, where sparse laboratory measurements are extrapolated through pedotransfer functions and geographic information systems to produce continuous property surfaces. Because the underlying inputs, particle size distribution and organic carbon, are already collected routinely by soil testing laboratories, the framework could be updated cheaply as management practices change, and adapted to neighbouring alluvial regions facing similar groundwater stress.

Subject of Research: Modelling and mapping of soil moisture retention characteristics of Punjab soils using pedotransfer functions

Article Title: Modelling and mapping of soil moisture characteristics of the Punjab soils

Article References: Kashyap, S., Vashisht, B. B., Kaur, H., & Arora, M. (2026). Modelling and mapping of soil moisture characteristics of the Punjab soils. Discover Soil, 3(1), Article 153. https://doi.org/10.1007/s44378-026-00307-9

Image Credits: AI Generated

DOI: 10.1007/s44378-026-00307-9

Keywords: soil moisture retention, field capacity, permanent wilting point, available water, pedotransfer functions, Punjab, QGIS mapping, soil organic carbon, bulk density, irrigation scheduling, agricultural water management, soil texture