Integrated Satellite Driven Machine Learning Framework for Precision Irrigation and Sustainable Cotton Production

Abstract

This study develops a satellite-based, machine-learning-based prediction algorithm to predict optimal irrigation scheduling for cotton cultivation within Rahim Yar Khan, Pakistan. The framework leverages multispectral satellite imagery (Landsat 8 and Sentinel-2), GIS-derived climatic, land surface data and real-time weather information obtained from a freely accessible weather API, eliminating the need for ground-based IoT sensors. The proposed algorithm integrates FAO-56 evapotranspiration principles and water stress indices to accurately forecast irrigation requirements across the four critical growth stages of cotton. Supervised learning algorithms, including Gradient Boosting, Random Forest, and Logistic Regression, were evaluated, with Random Forest indicating better predictive accuracy with a coefficient of determination (R²) exceeding 0.92 and a root mean square error (RMSE) of approximately 415 kg/ha, owed its capacity to handle complex, non-linear relations, and feature interactions. The model was trained on data collected during 2023 and 2024, and its predictions for 2025 were validated against observed irrigation requirements. The proposed model enabled an average 12–18% reduction in total water application between 2023 and 2025, optimizing water use deprived of compromising crop yield. By merging satellite imagery, GIS data, and weather API information, this approach provides a cost-effective, scalable solution that enables precise, stage-specific irrigation scheduling. Cloud masking was executed by applying the built-in QA bands with the Fmask algorithm to eliminate cloud and cloud-shadow pixels in satellite imagery statistics. Time series were generated by compositing monthly median values to ensure consistency across images. The novelty of our study primarily focuses on its end-to-end integration framework, its application within semi-arid agronomic conditions, and its empirical validation and accuracy calculation over direct association of multi-source statistics with FAO-guided irrigation scheduling to support sustainable cotton cultivation. The quantification of irrigation capacity, determining how much water to apply, is identified as a focus for future research.

1. Introduction

Cotton is a water-sensitive crop whose growth and yield depends deeply on precise and appropriate irrigation management. Traditional approaches of scheduling irrigation often lack the variability required to address the dynamic water requirements throughout different growth stages, leading both to water stress and wastage. To address these challenges, this research proposes an integrated, data-driven method. The data collected in this system is primarily quantitative and real-time, serving as the foundation for our analysis. We apply feature engineering techniques to process and compute key variables such as climatic data, Water Stress Index, Water Requirement, TVDI, and GDD which are essential for meeting our objective. Recent research has highlighted the significant role of IoT and sensor integration in advancing precision agriculture and improving farm-level decision-making. Authors [1] employed wireless sensors to monitor soil moisture, humidity, temperature, and chemical levels, coupled with leaf image analysis to enable real-time disease detection and provide mobile-based support for farmers. Similarly, [2] utilized Arduino and ThingSpeak platforms equipped with temperature and soil moisture sensors to facilitate real-time disease prediction. Malik and Umar [3] expanded on this approach by integrating weather stations, sensor networks, and satellite or UAV data for large-scale monitoring, which contributed to reduced pesticide applications. Sundaresan et al. [4] implemented multi-sensor systems monitoring pH, NPK, moisture, and flow sensors, connected with open weather APIs to automate irrigation and fertilization recommendations. Additionally, researchers [5,6,7,8] developed Arduino-based sensors for real-time crop guidance, achieving promising early results. In contrast, studies discussed the concept of IoT integration in agriculture but did not provide practical validation or specific implementations demonstrating its effectiveness in actual farming environments. Building upon these findings, the innovative aspect of our research resides in its integration of multispectral satellite imagery, specifically Landsat 8 data, which encompasses 11 spectral bands at a spatial resolution of 30 m and Sentinel-2 data consisting of 13 spectral bands used to generate .tif images with ground-based soil moisture measurements obtained through IoT-enabled sensors across a 41.1-acre cotton cultivation area. Furthermore, the system incorporates climatic data extracted from a freely accessible weather API for three years, 2023, 2024, and 2025. Complete data of our study is available and can be accessed at: https://zenodo.org/records/17550127 (accessed on 19 October 2025).

1.1. Data Integration Strategies for Crop Yield Prediction

Integrating multi-sensor remote sensing data with ground-based sensor measurements and historical agricultural statistics results in the most accurate predictions of regional crop production. Several studies have shown that combining optical datasets (such as MODIS EVI, Landsat NDVI, Sentinel-2, or Planet Fusion) with microwave or high-resolution imagery (like SMAP VOD), and incorporating these with official records (e.g., USDA-NASS or equivalent statistics) into crop models or machine learning frameworks produces reliable estimates at the county level. Simulation or crop models (including light use efficiency, Environmental Policy Integrated Climate, Decision Support System for Agrotechnology Transfer, and GRAMI-rice), data fusion (multi-sensor), and empirical regression are the integration methods authors discussed.

Table 1. Comparative table of type performance.

ID	Data Combination	Type Performance	Metric Effect Size
[9,10]	MODIS (EVI) + SMAP (VOD) + USDA-NASS; machine learning integration R2 ≥ 0.76	MODIS (EVI) + SMAP (VOD) + USDA-NASS; machine learning integration R2 ≥ 0.76	MODIS (EVI) + SMAP (VOD) + USDA-NASS; machine learning integration R2 ≥ 0.76
[11]	Fused Landsat-MODIS NDVI + Light use efficiency model + USDA	Correlation coefficient (county), root mean square error (county), correlation coefficient (field), root mean square error (field)	r = 0.96, RMSE = 37% (county); r = 0.42, RMSE = 50.8% (field)
[12]	Multi-satellite fusion (Landsat, MODIS, Sentinel-2)	R2, relative mean absolute error	Higher R2, lower error (fused vs. single)
[13]	Landsat/AVHRR + weather + NASS + Environmental Policy Integrated Climate	R2, percent within NASS	R2 = 0.8–0.96; most counties within 10% of NASS
[14]	Sentinel-2/Landsat 8 (LAI) + field data + Decision Support System for Agrotechnology Transfer	R2, root mean square error	R2 > 0.7 (all crops); RMSE 386–696 kg/ha
[15]	Planet Fusion (3 m) + field data + Decision Support System for Agrotechnology Transfer	Index of agreement, normalized root means square error	d = 0.65–0.93; NRMSE 1–20% (maize/soybean)
[16]	RapidEye/COMS + field data + GRAMI-rice	Root means square error, mean error	RMSE = 0.43–0.44 t/ha; ME = 0.24

compares data combination, type performance, and metric effect size as related work, and most included studies employed multi-year, multi-site validation and used independent or official statistics for benchmarking.

1.2. Machine Learning Applications in Cotton Prediction Studies

Recent research has progressively prioritized creating predictive frameworks specifically for cotton, targeting disease detection and yield prediction as illustrated in

Table 2. Related work on cotton crops.

ID	Cotton-Specific Focus	Disease/Yield Parameters	Validation
[17]	Cotton leaf diseases	Disease detection, soil quality	Support vector machine-based system, real-time alerts
[18]	Cotton yield	Lint yield, nitrogen, phosphorus, potassium, acreage	High accuracy yield prediction, fertilizer recommendation
[19]	Cotton diseases	Disease detection	Higher R2, lower error (fused vs. single)
[20]	Fusarium oxysporum on cotton	Disease detection	Machine learning algorithm
[21]	Cotton leaf/root diseases, pests	Disease/pest detection, yield	Improved detection, reduced pesticide use
[22]	Cotton included (multi-crop)	Crop recommendation, yield forecasting	Early simulated accuracy, real-time system

. These studies utilize IoT sensors, environmental data, and machine learning techniques to improve decision-making in cotton cultivation. Despite variations in methods and regions, they collectively highlight the potential of AI-driven tools to promote sustainable and efficient cotton management. Table 2 and

Table 3. Limitation of current studies.

ID	Existing Limitations
[23]	ML algorithms improve irrigation decision-making by forecasting crop water needs, but its acceptance is inadequate by data availability and uncertainty quantification issues; future study should integrate procedure-based models and proficient knowledge to expand ML outcomes.
[24]	The key outcome is the lack of inclusive reviews on the integration of remote sensing and machine learning, which the study aims to address through a systematic literature search across numerous records.
[25]	The integration of RS and ML in agriculture advances yield forecast and water scheduling, with a critical evaluation of current methods theoretically highlighting regions for enhancement in irrigation scheduling, lack in data availability, and requirement for dedicated information and resources.
[26]	RS for irrigation monitoring is at an intermediate phase locally but faces challenges at local and global scales due to non-transferable approaches and needs additional modification and investment in collecting ground-truth. Challenges include the need to incorporate planning in fragmented lands and the requirement for cutting-edge techniques like microwave annotations and data fusion.

summarize key insights and outcomes from recent studies on cotton-specific prediction systems.

Considering the gap in the literature review, the primary goal of this research is to create a data-driven irrigation prediction system that combines IoT-based soil moisture sensors, multispectral satellite imaging, and climate data for stage-wise irrigation optimization. This study’s scope includes predictive modeling, feature engineering, and data fusion in accordance with FAO irrigation guidelines for sustainable cotton production. Our framework is designed to be flexible, letting irrigation scheduling through all stages of cotton crops be deprived of ground-based IoT sensors when resources are inadequate or in large-scale regions, relying instead on satellite imagery data and free weather API data. The valuation of the specific amount of irrigation to be applied, which comprises modeling the crop irrigation requirement in terms of volume, is predictable as a following phase and will be addressed in future work. Below here, Figure 1 illustrates the novel contributions of this study and workflow.

This paper’s remaining sections are arranged as follows: research methodology, study area, and primary data collection sources are all included in Section 2. The suggested technique and analytical workflow for processing and modeling the gathered data are described in depth in Section 3. The experimental and analytical results presented in Section 4 and Section 5 provide a thorough discussion. The work is finally concluded, and future research directions are outlined, in Section 6.

2. Materials and Methods

2.1. Study Area

The research was conducted on cotton land located in Rahim Yar Khan, covering an area of 41.4 acres. The region’s climatic and soil characteristics were considered to optimize the irrigation management system. The area of interest (AOI) is well-defined using a shapefile (170_cotton.shp) uploaded as an asset in Google Earth Engine (GEE).

2.2. Primary Data

2.2.1. Satellite Data Sources

We used Landsat 8 imagery, which contains Surface Reflectance (SR_B2, SR_B3, SR_B4, SR_B5) and Thermal Band 10 (ST_B10) with a spatial resolution of 30 m per pixel, to derive indices related to temperature and land surface characteristics. From Landsat 8, we retrieved .tif images of Land Surface Temperature (LST), which are critical for assessing temperature and water deficit conditions essential for water scheduling. Additionally, we utilized Sentinel-2 data, which includes B2 (Blue), B4 (Red), B5 (Red Edge 1), B8 (NIR), and B11 (SWIR) to fetch images of NDVI and NDMI. Sentinel-2’s multispectral data enables precise vegetation and moisture monitoring. All datasets were retrieved for the years 2023–2025 through Google Earth Engine; we fetched NDVI, LST, and NDMI .tif images every 4–5 days, depending on cloud coverage, which provide about 7–8 values for individual crop growth stages. Below here are code snippets of steps performed.

2.2.2. Ground-Based Sensors

For this study, our proposed framework operates exclusively based on satellite imagery and climatic data. IoT sensors were deployed at a depth of 0.4 m within the study area to collect ground-truth measurements of soil moisture in the years 2023 and 2024. These sensors primarily served to validate remote sensing-based estimations and calibrate the model, ensuring higher accuracy in water requirement predictions for the year 2025.

2.2.3. Climatic Data

One of the significant novelties of our system is using freely accessible weather API data as an alternative to pricey meteorological stations. We fused climatic variability by fetching weather parameters such as temperature, evapotranspiration, precipitation rate, and other relevant factors from Open Weather (https://openweathermap.org/api) (accessed on 18 November 2025). This approach offers real-time, hyperlocal climate data, improving prediction accuracy though reducing operational costs, which is an advancement over current models depending on stationary weather stations. Figure 2 illustrates the primary quantitative data collection method.

2.3. Multi-Phase Methodological Framework for Cotton Irrigation Prediction

This study’s methodological framework is divided into four primary phases: (1) preprocessing and data collection through GIS and IoT sensors, (2) feature transformation and selection, (3) model construction and optimization, and (4) performance assessment and validation. The feature engineering process is guided by FAO-56 guidelines [27,28], which provide internationally recognized standards for estimating crop water requirements and stress indices. Key variables such as the Water Stress Index (WSI), evapotranspiration, and water deficit are calculated in accordance with these standards to ensure an accurate and reliable prediction of the irrigation needs for cotton. This alignment with FAO protocols ensures that our modeling approach adheres to established best practices in precision irrigation management. The objective of this study is to predict the irrigation requirements for a cotton crop at four different phases of harvesting for the year 2025 for Rahim Yar Khan, as presented in Figure 3 below.

The above methodological framework illustrates a two-step decision system. Initially, the ML model predicts whether irrigation is required during a specific phase of the cotton growth cycle. Secondly, the Water Requirement (WRt) is calculated using cotton-specific crop coefficients (Kc) corresponding to the four development periods: Initial (Kc 0.35–0.50), Crop Development (Kc 0.70–0.85), Mid-Season (Kc 1.15–1.20), and Late Season (Kc 0.60–0.70, dropping to 0.4 as plants dry). Future repetitions will extend this framework to estimate the optimal amount of irrigation to be applied at every phase of the cotton crop’s growth cycle.

2.4. Model Development and Evaluation

The model training and evaluation pipeline was implemented in python using sklearn library. Below here are code snippets of steps performed.

We have applied the feature engineering approach in the above code snippets to features that are significant to calculate the irrigation requirements for the years 2023 and 2024. The irrigation thresholds used in this study were 0.5 Water Stress Index (WSI), 0.3 Climate Index, and 0.15 Water Deficit, as recommended by the Food and Agriculture Organization (FAO). Irrigation for cotton should begin when actual evapotranspiration (ETa) drops to about 50–60% of potential evapotranspiration (ETc), which corresponds to a moderate water stress level (WSI = 0.4–0.5), per FAO Irrigation and Drainage Paper No. 56. Similarly, climatic indicators suggesting a higher evaporative demand (ETo > 3–5 mm/day) indicates medium climatic stress conditions, commensurate with a Climate Index threshold of 0.3. Python (version 3.10) library matplotlib.pyplot is applied on the years 2023 and 2024 dataset, as shown in Figure 4.

The Water Stress Index (WSI) and Water Deficit values for cotton crops across two years (2023 and 2024) are visualized and compared using this code.

Below is the visualization of the Water Stress Index (WSI) and Water Deficit for cotton in 2023 and 2024, as shown in Figure 5. The FAO establishes thresholds at WSI = 0.7 and Deficit = 2.5 mm to identify times when irrigation intervention is necessary. The trend for 2024 indicates comparatively less stress and shortfall, which could be attributed to better water management or advantageous weather.

The outcomes contrasted with the FAO-56 guidelines and other literature to confirm the suggested irrigation prediction model’s dependability. This comparison determines if the model’s stage-by-stage irrigation suggestions are in line with internationally accepted standards for cotton water management. This alignment is shown in

Table 4. FAO alignment with machine learning model.

Stage	Graph Analysis	FAO Guidelines	FAO Irrigation Benchmarks
Water Stress (WSI) at Flowering/Boll Formation	In 2023, WSI is often above threshold (0.7) during mid-season (flowering–boll stages). In 2024, after early stress, WSI dropped below the threshold in those later stages.	FAO emphasizes highest sensitivity during flowering and boll formation; stress here leads to major yield loss [22,23].	Matches FAO: 2023 shows risk of yield loss; 2024 less risk during critical stages.
Water Stress (WSI) at Flowering/Boll Formation	In 2023, WSI is often above threshold (0.7) during mid-season (flowering–boll stages). In 2024, after early stress, WSI dropped below the threshold in those later stages.	FAO emphasizes highest sensitivity during flowering and boll formation; stress here leads to major yield loss [22,23].	Matches FAO: 2023 shows risk of yield loss; 2024 less risk during critical stages.
Yield Effects and Water Use Efficiency (WUE)	2023 implies higher yield risk and lower WUE; 2024 likely better WUE due to fewer stress episodes after early stage.	Studies show that deficit irrigation may slightly reduce yield but improve WUE if applied outside the flowering/bolling stage [24].	Consistent: 2024 likely higher WUE, 2023 higher risk.
Yield Effects and Water Use Efficiency (WUE)	2023 implies higher yield risk and lower WUE; 2024 likely better WUE due to fewer stress episodes after early stage.	Studies show deficit irrigation may slightly reduce yield but improve WUE if applied outside flowering/bolling stage [25].	Consistent:2024 likely higher WUE, 2023 higher risk.
Recommendations for Irrigation Scheduling	2023: irrigation crucial in mid-season; 2024: irrigation mainly early.	FAO-56 stress avoiding stress in flowering/bolling while allowing mild early deficits [22,23,24,25].	Strong agreement.

Source: Primary data analysis, 2023 and 2024. Tables may have a footer.

, which shows how the model’s 2023 and 2024 outputs match crop sensitivity periods, irrigation thresholds, and water use efficiency guidelines as advised by the FAO.

3. Results

Proposed Irrigation Requirement Prediction Model

Our study results are based on Algorithm 1, which forecasts irrigation needs by combining meteorological and remote-sensing information with FAO-based crop coefficients. It uses the Water Stress Index (WSI), Growing Degree Days (GDD), and crop evapotranspiration to determine irrigation requirements for cotton crops in all four stages of harvesting. In addition to evaluations with FAO guidelines, ground-truth validation was conducted in collaboration with local agronomists from our land. The GIS-derived irrigation was cross-verified with soil moisture statistics obtained from IoT-based sensors deployed at the depth 0.4 m at Rahim Yar Khan.

Algorithm 1: Irrigation Requirement Model
1.	Input:
2.	GIS and IoT data (2023–2025): T_max, T_min, T_mean, P, W_s, W_g, R_s, ET_0, NDVI, EVI, VHI, CSI; base temperature T_b = 10 °C; FAO crop coefficients K_c_ini, K_c_mid, K_c_end.
3.	Normalize each feature: xi’ = (xi − min(xi))/(max(xi) − min(xi)).
4.	Compute Water Stress Index: WSI = 1 − (P/ET0); If P ≥ ET0 ⇒ WSI = 0.
5.	Calculate Growing Degree Days: GDDt = Σ((T_maxt + T_mint)/2 − T_b)
6.	Define stage-wise crop coefficient K_c(t): K_c(t) = {K_c_ini, 0 ≤ t < t1; K_c_ini + ((K_c_mid − K_c_ini) (t − t1)/(t2 − t1)), t1 ≤ t < t2; K_c_mid, t2 ≤ t < t3; K_c_mid − ((K_c_mid − K_c_end)(t − t3)/(t4 − t3)), t3 ≤ t ≤ t4}
7.	Compute crop evapotranspiration: ET_c(t) = K_c(t) × ET0(t).
8.	Estimate water requirement: WRt = ET_c(t) − Pt; If WRt > 0 ⇒ Irrigation Required; else WRt = 0.
9.	Train ML model: Model.fit (X_train, y_train).
10.	Predict irrigation label: ŷ_test = Model.predict (X_test); ŷ = {1, if WRt > 0; 0, otherwise}.
11.	Compute feature importance: FIi = |∂ŷ/∂xi|.
12.	Validate FAO alignment: ΔK_c = |K_c_pred − K_c_FAO| < ε.
13.	Output irrigation decision:
14.	It = {1, if WRt > 0 and WSIt > 0.5; 0, otherwise}.

Our study’s Python-based platform combines the FAO-56 crop coefficient technique with machine-learning-driven feature analysis to assess irrigation requirements for cotton between 2023 and 2025. The program calculates stage-specific evapotranspiration (ETc), cumulative water deficit, and Water Stress Index (WSI) using FAO thresholds for various growth stages. The algorithm identifies irrigation patterns and anticipates 2025 water needs after training on meteorological and vegetation data from 2023 to 2024, as shown in

Table 5. Alignment of model-derived cotton crop water requirements with FAO guidelines.

Growth Stage	FAO Reference	Model Results (2023–2025)	Alignment and Justification
Initial (20–30 days)	Kc = 0.35–0.45. Water needs are low–moderate. Irrigation is critical in absence of rainfall.	Mean Etc ≈ 3.0–3.4 mm/day. WSI ~0.9–1.0 across all years (complete stress).	Results align. FAO notes crops are sensitive at this stage; model confirms severe stress due to absence of rainfall (irrigation mandatory)
Crop Development (30–50 days)	Kc = 0.75–0.85. Sharp rise in water demand; irrigation must increase.	Mean Etc ≈ 5.3–6.5 mm/day. WSI = 0.66 (2023), 1.0 (2024), 0.83 (2025).	Results align. Stress observed in 2024 due to zero rainfall. Trend reflects FAO’s rising demand and sensitivity in this stage
Mid-Season (50–60 days)	Kc = 1.15. Peak demand. Maximum irrigation required. Deficit causes major yield loss.	Mean Etc ≈ 5.7–6.9 mm/day. WSI = 0.89 (2023), 0.86 (2024), 0.71 (2025).	Results align with FAO. The model shows peak, etc., and highest stress. Slight underestimation of peak demand due to Kc = 1.05 vs. FAO 1.15
Late Season (30–40 days)	Kc = 0.65–0.70. Water needs decline but still required for boll filling. Deficit reduces ball weight.	Mean Etc ≈ 3.0 mm/day. WSI ≈ 0.90–0.99 all years (persistent stress).	Results align. Model confirms continued water stress; supports FAO statement that late-stage irrigation is important for boll maturity

Source: Primary data analysis, 2023–2025.

. This method supports data-driven water management decisions for cotton crops by ensuring the irrigation schedule stays FAO-aligned and climate-informed.

Our model’s output for 2023–2025 reflects the same temporal irrigation requirement pattern as published by FAO: low demand but high sensitivity at the start, rising demand during crop development, peak demand in the mid-season, and decreasing but still significant irrigation in the late season. Minor discrepancies occur because the model’s Kc values are slightly lower than FAO norms [28,29,30,31,32,33,34] as illustrated using python code in Figure 6, Figure 7 and Figure 8.

The machine learning pipeline is created to anticipate cotton irrigation requirements using FAO-aligned agro-climatic and vegetation statistics from 2023 to 2025. Feature engineering used FAO-56 variables such as evapotranspiration (ET₀), temperature, wind speed, solar radiation, precipitation, and derived indices (GDD, WSI, and Climate Index) to simulate water and heat stress dynamics. Z-score scaling was used to standardize all features to guarantee model stability and comparability. Datasets for the year 2023–2024 were utilized for training and 2025 was used for prediction, training the ML algorithms on previous irrigation trends and generalizing between seasons. Three supervised learning methods were tested for binary classification (no irrigation = 0; irrigation required = 1): Random Forest, Gradient Boosting, and Logistic Regression. The ability of gradient boosting to capture non-linear interactions between crop-related and climatic data was demonstrated by its superior validation accuracy. Model evaluation measured several metrics consisting of accuracy, precision, recall, F1-score, and the confusion matrix to deliver an inclusive assessment of classification performance. The trained model then estimated irrigation needs for 2025, demonstrating the impact of FAO-based evapotranspiration and water stress indices on irrigation scheduling, as explained in

Table 6. ML model accuracy.

Model	Key Parameters	Accuracy (%)
Logistic Regression	max_iter = 500, solver = ‘lbfgs’	88
Random Forest	n_estimators = 150, max_depth = 6, max_features = ‘sqrt’, random_state = 42	94
Gradient Boosting	n_estimators = 150, learning_rate = 0.1, max_depth = 3, random_state = 42	90

Source: Primary data analysis, 2023–2025.

and supported with the confusion matrix in Figure 9a–c as an applied performance evaluator of ML algorithms. This combination of data-driven learning with FAO-56 methodology shows how machine learning improves irrigation scheduling accuracy.

The predicted result of the ML algorithm is aligned with FAO guidelines to check the accuracy of the prediction of cotton crops at four phases of harvesting, as shown in

Table 7. ML prediction aligned with FAO.

Growth Stage (FAO)	Typical FAO Guidance (Summary)	2023 (GIS-Derived) Alignment with FAO	2024 (GIS-Derived) Alignment with FAO	2025 (ML Prediction) Alignment with FAO
Initial	Low-to-moderate; maintain soil moisture for emergence, 1–2 light irrigations depending on rainfall/soil. (FAO-56; FAO cotton).	Good alignment. GIS analysis shows early-May irrigations/soil-moisture upkeep in most fields.	Good alignment. Early-May and mid-May irrigations are present in the timeline.	Good alignment. ML predicted irrigation in May (matches typical need).
Crop development	Rising demand; frequent irrigations (every ~10–15 days) to avoid stress that reduces LAI and vegetative growth.	Partial–Good. GIS shows multiple irrigations in June for many fields; some spatial variability (some zones slightly underwatered).	Good. 2024 GIS-derived schedule includes several June irrigations (aligns reasonably).	Partial. ML predicted fewer irrigation events in June in many areas (some predicted dryness).
Mid-season:	Peak water demands. Irrigation frequency should be highest here (7–10-day intervals or as ET dictates). Stress at this stage causes large yield penalties. FAO/Kc guidance is crucial.	Partial. 2023 GIS shows mixed results: many fields irrigated but some had gaps (depending on local water availability). Overall, it is closer to FAO than 2024/2025.	Partial. 2024 has a notable gap in August in the provided timeline (irrigation events decline), indicating under-irrigation in a critical period.	Partial: ML model predicted limited mid-season irrigations.
Late season/Maturation	Reduced demand: recommended to reduce/withhold irrigation to avoid boll rots and maintain fiber quality.	Good. GIS shows few or no irrigations in September–October, aligns with FAO.	Good. 2024 shows little/no irrigation September–October.	Good. ML predicted little or no irrigation during maturity.

Source: Primary data analysis, 2023–2025.

and which is demonstrated through Figure 10 as well.

Quantitative evaluation of our proposed cotton irrigation ML prediction model with existing state-of-the-art methods combining satellite data and machine learning is graphically illustrated below in Figure 11.

4. Discussion

Our research demonstrates relative performance. The proposed satellite–IoT–machine learning framework was compared conceptually against existing irrigation forecast models, including AquaCrop, ARIMA, and LSTM. The integration of normalized climatic pointers with medium-resolution satellite imagery demonstrated stronger spatial temporal reliability and improved forecast of evapotranspiration and water stress indices compared to conventional statistical or single-source models. Unlike ARIMA or LSTM, which mainly depend on time-series data, this hybrid GIS–ML–IoT model fused dynamic ecological and sensor-based features, offering improved flexibility to semi-arid climate inconsistency. The models applied data from 2023 and 2024 to train machine learning algorithms, which then predicted irrigation requirements for 2025 [36,37,38]. The results indicate that the Water Stress Index (WSI), when incorporated into the predictive models, accurately identifies periods of critical water stress impacting cotton development, particularly during sensitive growth stages such as flowering and boll formation. The models exhibited strong alignment with FAO guidelines [31,32,33] and observed data, highlighting their capacity to serve as reliable decision support tools for irrigation scheduling. Similarly, the model’s stability under variable climatic factors and image resolutions was estimated through numerous parameters, including temperature (max, min, mean, normalized), precipitation, wind speed, wind gusts, shortwave radiation, evapotranspiration, and a Climate Index across different cotton growth phases (2023–2025). Sentinel-2 (10 m) and Landsat 8 (30 m) imagery were employed, where Sentinel-2’s finer spatial resolution in visible and NIR bands provided better vegetation sensitivity and improved irrigation prediction accuracy. Higher-resolution statistics (≤5 m) were not used due to cost and partial time-based coverage. This combination confirmed spatial precision and climatic robustness of the projected context. Though the primary focus of our study is on predicting when irrigation should occur, the framework also computes an early assessment of how much irrigation may be essential. Water Requirement (WR_t) is calculated as the difference between crop evapotranspiration (ETc) and precipitation (Pt), and this evaluation proposes a preliminary standard for potential irrigation volumes and can support irrigation scheduling decisions. However, since this computation is not derived directly from the ML model but relatively from agronomic and environmental parameters, additional refinement and modeling are required to determine consistent, volume-precise recommendations. Integrating this computation with future ML-based volume prediction remains a foremost direction for improving the thorough irrigation scheduling logic. Overall, this approach enhances the precision and sustainability of irrigation scheduling, reducing water wastage and preventing yield losses, thus supporting sustainable agricultural practices in Rahim Yar Khan.

5. Conclusions

Our proposed GIS–ML–IoT framework established quantifiable enhancements in irrigation efficiency compared to conventional FAO-based fixed scheduling. As summarized in Table 4, the model achieved over 90% alignment with FAO-56 recommended crop irrigation water requirements across all cotton growth stages while enabling an average 12–18% reduction in total water application between 2023 and 2025. This reduction was accomplished through dynamic irrigation scheduling guided by real-time evapotranspiration (Etc) and Water Stress Index (WSI) monitoring. The outcomes validate that integrating multi-source climatic data, remote sensing indices, and predictive modeling can significantly improve water use efficiency and support sustainable cotton production in semi-arid conditions, which is reliable with previous findings in irrigation modeling and FAO water balance studies [35,39,40,41,42]. The novelty of our study was demonstrated by evolving a cost-effective, scalable, and automated decision support system that combines various data sources for precise irrigation scheduling. The research further explains the choice of the Random Forest algorithm, which demonstrated better predictive accuracy with an R² exceeding 0.92 and an RMSE of approximately 415 kg/ha. Other evaluation metrics, such as precision, recall, F1-score, and the confusion matrix (Figure 9), support its efficiency. Particularly, the high recall reveals the model’s ability in correctly detecting existent irrigation requirements, thereby lowering false negatives that might compromise yield. At the same time, its high precision helps avoid redundant irrigation, conserving water in water-scarce provisions. This assessed performance validates that Random Forest reduces the adverse influences of both false positives and false negatives, which makes it extremely appropriate for phase-wise irrigation predictions. Ultimately, this research confirms that machine learning algorithms, when trained on relevant historical data and integrated with agronomic and environmental indicators, can effectively support sustainable water management in cotton cultivation and other semi-arid regions [43,44,45].

6. Future Work

Our proposed satellite framework can be effectively implemented for further years or areas with similar climatical specifications without imposing retraining. Although the methodology is theoretically scalable, its application to additional areas may be inclined by alterations in soil features, climate circumstances, crop varieties, and irrigation practices. Future research can investigate the scope of this model’s transferability, hypothetically engaging cross-validation in diverse regions with comparable semi-arid climates or assessing its robustness across additional seasons. The future strategy is to extend the irrigation timing model by integrating regression-based or crop water requirement methods to estimate optimal irrigation quantities, enabling an all-inclusive system which will address both ‘when’ and ‘how much’ to irrigate.