The practice of growing cereal/legume intercrops among subsistence farmers in semi-arid parts of South is well documented. It is worth noting that the reason behind the continued practice of such a farming method cannot be attributed merely to culture. In TamilNadu’s arid areas, with an annual average rainfall of less than 350 mm and irregular rainfall within the monsoon season, intercropping sorghum with lentils has advantages that cannot be achieved by growing them individually, such as interception of light by two layers, nutrient absorption from various depths in the soil and two harvest times. When one component struggles in a dry year, the other frequently compensates and that yield-stability property is worth considerably more to a resource-constrained household than any marginal gain from sole-crop intensification.
       
Repeated over several years, this pattern pushes chemical inputs upward, weakens root nodule activity, reduces the soil’s native population of arbuscular mycorrhizal fungi and in calcareous semi-arid profiles gradually shifts pH in a direction that further penalises microbial communities. Integrated nutrient management shorthand for the deliberate combination of chemical fertilisers at reduced rates with organic manures and microbial inoculants has attracted attention as a way out of this cycle, but published evidence for the lentil-sorghum intercrop specifically, tracked across multiple seasons with soil biological measurements taken alongside yield, remains thin. Field studies on legume-cereal crop rotations have provided empirical data that proves that the combination of organic fertilizers with biofertilizers within INM provides better yield parameters and soil quality parameters compared to the use of individual sources of nutrients. According to Sodavadiya et al., (2023), treatment with 50% RDF along with vermicompost, Rhizobium and PSB resulted in the best seed yield and dry root biomass in a sequence of chickpea and forage sorghum crops by Gaba et al., (2015).
       
A second gap sits at the intersection of agronomy and data science. The interaction between nutrient input, soil organic matter carry-over, microbial activity, crop growth and seasonal weather does not unfold in straight lines and ordinary regression cannot capture it faithfully. Threshold responses, lagged effects of organic matter decomposition and feedback loops between soil biology and crop nitrogen uptake all require modelling approaches that can handle non-linearity and temporal sequence simultaneously (Bedoussac et al., 2021). Used together, they form a hybrid architecture suited to exactly the kind of multi-dimensional, season-length prediction problem that INM optimisation in dryland intercrops represents in Fig 1.


       
The relevance of the proposed machine learning approach to the problem domain is not an issue. Random Forest models have successfully identified patterns of crop diseases in leguminous systems with highly accurate classifications (Cho, 2024), whereas intercropping systems utilizing sorghum under rainfed vertisols reveal that yield improvement according to Land Equivalent Ratio (LER) is contingent upon treatment-level data that cannot be captured by linear models during different seasons (Sujathamma and Nedunchezhiyan, 2024). In other words, a combination of machine learning algorithms for the processing of multi-feature time series agricultural data is not only computationally convenient; it is a necessity, as the output variable-in this case, lentil-equivalent yields-is influenced by nutrients, soil biology and seasonality.
       
While all these developments have been made, there is still a major issue: no such research exists in which an ensemble approach using RF-LSTM has been applied to predict INM treatments in the context of an intercropping system involving lentil-sorghum crops. Machine learning models that have been proposed for similar agronomic settings either use monocrop datasets, thus not capturing the competition between the species (Prodhan et al., 2022; Corrales et al., 2022), or miss out on soil biological factors, including dehydrogenase activity, microbial biomass carbon content and AMF colonization, which have been proven sensitive to organic fertilizers (Sharma et al., 2021). Also, there have been no attempts by any machine learning models to consider the season-dependent impact of organic fertilizer application on soil organic carbon and subsequent crop yields, which can be successfully accounted for using LSTM models. This study bridges those gaps, employing the knowledge from two kharif seasons of agronomic, soil biological and meteorological data through a valid RF-LSTM ensemble.
       
The present investigation was accordingly structured around three objectives: first, to characterise the agronomic response of lentil-sorghum intercropping to seven INM treatments across two kharif seasons under field conditions at Pattukottai; second, to build and validate a hybrid RF-LSTM model capable of predicting lentil equivalent yield and soil organic carbon dynamics from field-collected soil, crop and weather data and third, to extract from that model a ranked account of the agronomic and environmental variables most responsible for yield variability, so that the findings carry practical relevance beyond the single experimental site.
 
Literature review and research gap analysis
 
Summary of representative studies (2020-2025)
 
Table 1 synthesises twelve peer-reviewed studies published between 2020 and 2025 that addressed, individually or in combination, INM in legume-cereal systems, machine learning applications in nutrient and yield prediction and soil biological response to organic inputs.


 
Identified research gaps
 
As shown in Table 1, there exist four major gaps that the current study intends to address, based on the systematic review of literature. Gap 1: To date, there are no published studies where the hybrid approach of RF-LSTM is used to predict INM treatment data in an intercrop of lentils and sorghum. Machine learning methods, either those used to analyze monoculture datasets (Prodhan et al., 2022; Corrales et al., 2022) or excluding organic predictors (Droutsas et al., 2022), were applied to solve agroecological problems. Gap 2: Despite the high responsiveness of soil biological parameters (dehydrogenase activity, microbial biomass carbon, AMF colonisation) to the impact of organic fertilizers (Sharma et al., 2021), none of the intercrop models described in the literature have considered them as independent or dependent variables. Gap 3: The impact of organic manure application in one season on the SOC level and yield in subsequent seasons has not yet been captured by any ML model except for the traditional regression-based approach which does not have memory of consecutive soil state changes. Gap 4: Decision-support outputs (Feature ranking leading to management recommendations) have never appeared in the literature among agronomic INM models.

Experimental site and soil
 
The current research was held out at Pattukottai, Tamil Nadu state, India, during the kharif season of the years 2022 and 2023 with coordinates of 10°25'N latitude, 79°19' E longitude at 6 m above the mean sea level. The weather falls under the category of semiarid tropics (BSh) based on the Koppen classification, receiving an average annual rainfall of 920 mm, northeast monsoon month of October to December around 60% occurs. The experimental soil is characterized by sandy loam texture, pH 7.8, EC 0.31 dS/m, OC 0.38%, available nitrogen 138 kg/ha, available phosphorus 13.4 kg/ha and available potassium 186 kg/ha in the upper 15 cm layer at the start of the first experiment.
 
Field treatments
 
Seven treatments in a randomised block design (Three replications; gross plot 5.0 m × 4.0 m) compared sole crops with intercropped lentil (var. IPL 316)-sorghum (var. CSV 23) in a 2:1 row arrangement at 30 cm inter-row spacing. Treatment combinations ranged from 100% RDF with no organic input (T3) to reduced-chemical plus vermicompost plus biofertiliser (T7). Full treatment details are presented in Table 2.


 
Data collection and feature engineering
 
Over both field seasons, 2,184 structured observations were recorded at 14-day intervals per plot. Each observation comprised 18 variables: treatment code, days after sowing (DAS), plant height (cm), leaf area index (LAI), SPAD chlorophyll index, soil temperature at 10 cm depth (°C), gravimetric soil moisture (%), soil pH, electrical conductivity (dS m^-1), available nitrogen (kg ha-1), available phosphorus (kg ha^-1), cumulative rainfall (mm), maximum air temperature (°C), minimum air temperature (°C), cumulative growing-degree days (GDD, base 10°C), dehydrogenase activity (DHA, µg TPF g^-1 24 h^-1), microbial biomass carbon (MBC, µg C g^-1) and AMF colonisation percentage. Final grain yield and LEY at harvest served as primary response variables; post-harvest SOC was a secondary response variable. Categorical treatment codes were one-hot encoded. Missing values from sensor faults (≈1.3% of records) were imputed using the predictive mean matching method implemented in the R package mice (version 3.16).
 
Machine learning framework: RF-LSTM hybrid model
 
Architecture overview
 
A two-stage hybrid architecture was designed to exploit the complementary strengths of random forest (RF) and long short-term memory (LSTM) networks. In stage 1, an RF ensemble performed feature selection and generated plot-level intermediate predictions (RF¯predictions) at each observation time-step. In stage 2, sequences of RF¯predictions combined with raw temporal features were fed as input vectors to an LSTM network that modelled season-length temporal dependencies and soil carry-over effects. The final LEY prediction was produced at season end from the LSTM output layer. This architecture is consistent with the hybrid approach validated by Kuradusenge et al., (2023) for potato and maize yield prediction (R² = 0.91) but extends that framework by incorporating soil biological parameters and an explicit intercropping competition index as input features are shown in Fig 2.


 
Implementation details
 
The RF component was implemented in python 3.11 using scikit-learn 1.4.2. The LSTM component was built in TensorFlow 2.15 with the Keras functional API. SHAP explanations were computed using the shap library (version 0.45). All experiments were conducted on a workstation with an NVIDIA RTX 3060 GPU (12 GB VRAM) and 32 GB RAM running Ubuntu 22.04. Training time for the complete pipeline was 47 minutes per season fold. Model weights and the pre-processed dataset (anonymised plot codes) have been deposited in a public repository for reproducibility.
 
Algorithm: RF-LSTM hybrid for INM yield prediction
 
Algorithm 1 presents the complete pseudocode for the RF-LSTM training and prediction pipeline.


 
Conventional agronomic observations
 
Yield attributes (pods plant^-1, seeds pod^-1, 100-seed weight), grain yield, LEY, LER and area-time equivalent ratio (ATER) were recorded following standard protocols. Post-harvest soil samples (0-15 cm) were analysed for SOC (Walkley-black), available N (Alkaline permanganate), available P (Olsen), dehydrogenase activity (TTC reduction), MBC (fumigation-extraction) and AMF colonisation (trypan blue gridline intersection). Economic analysis was conducted using prevailing market prices. All statistical comparisons employed ANOVA with LSD at P≤0.05.

ML model performance
 
Table 3 summarises the predictive accuracy of the RF-LSTM hybrid model against baseline multiple linear regression (MLR) and a standalone RF model on the held-out test set. The hybrid model achieved R² = 0.93 and RMSE = 112 kg ha^-1, representing a 19.2% reduction in RMSE over MLR and a 7.4% reduction over standalone RF. The MAPE of 4.8% falls within the ≤5% threshold recommended for field-crop yield forecasting systems (Ennaji et al., 2023). Validation loss stabilised at epoch 88 on average across five folds, with no evidence of over-fitting beyond epoch 60 for any fold.


 
Feature importance (SHAP analysis)
 
The results obtained by using the SHAP method mentioned in Table 4, for evaluating the significance of individual features showed that FYM rate, SOC at sowing and GDD are the top three features with the greatest mean absolute SHAP values (0.312, 0.284 and 0.251, respectively). The fourth place belongs to the DHA feature, whose mean |SHAP| value equals 0.198. This shows the contribution of the soil biota to the prediction model as an additional factor in addition to the traditionally considered chemical fertility indicators. Chemical N rate ranked seventh (0.148), substantially below FYM rate, which underscores a shift in the feature-importance space away from the primary input used in sole-crop management recommendations.


       
The dominance of SOC at sowing as the second most influential predictor aligns with the meta-analytical finding of Bargaz et al., (2018) that soil carbon stock at the start of a growing season is a stronger predictor of subsequent crop response to nutrient inputs than the inputs themselves, because SOC governs the soil’s buffering capacity for water, nutrients and microbial inoculants. This finding has direct implications for recommendation systems: farmers with below-threshold SOC (≤0.35%) should prioritise multi-season organic matter build-up before expecting full yield response to biofertiliser inoculation.
       
Among all input the most influential predictor of lentil equivalent yield (mean |SHAP| = 0.312), ranking more than twice the nitrogen chemical (mean |SHAP| = 0.148, ranked seventh). This dominance is agronomically justified: Farmyard Manure improves sandy loam water retention during pod filling, supplies slow-release nutrients and sustains rhizobium-PSB inoculant populations, as reflected in T6 recording the highest DHA (54.2 µg TPF g^-1 24 h^-1) and MBC (318 µg C g^-1) -60.7% and 48.6% above sole-chemical T3 respectively (Table 5). The LSTM component successfully encoded this as a temporal carry-over effect across growth stages, a dependency that standalone regression cannot capture. SOC at sowing (mean |SHAP| = 0.284, ranked second) further confirms application rate, not chemical N, should be the primary management decision in semi-arid lentil-sorghum systems.


 
Soil chemical and biological fertility
 
Post-harvest soil analysis (Table 5) confirmed significant improvements in all measured soil health parameters under INM treatments relative to T3. Soil organic carbon reached 0.59% under T6, a 44% increase over initial values (0.41%) and a 26% gain over T3 (0.47%). Dehydrogenase activity was highest in T6 (54.2 µg TPF g^-1 24 h^-1), 60.7% above T3 and 90.8% above the initial measurement. Microbial biomass carbon under T6 (318 µg C g^-1) was 48.6% higher than T3 (214 µg C g^-1). AMF colonisation in lentil roots was 52.4% under T6, compared with 34.6% under T3, with a significant negative correlation between P fertiliser rate and AMF colonisation (r = -0.77, P≤0.01).

Four lines of evidence emerged from this investigation. In the field, Treatment T6-75% recommended fertiliser dose paired with farmyard manure at 5 t ha^-1 and rhizobium-PSB seed inoculation-consistently delivered the strongest lentil equivalent yield (3,842 kg ha^-1), land equivalent ratio (1.48), area-time equivalent ratio (1.31) and soil biological indicators across both kharif seasons at Pattukottai. The hybrid RF-LSTM model predicted LEY with R² = 0.93 and MAPE = 4.8% from 17 agronomic and weather features, outpacing linear regression (R² = 0.71) and standalone random forest (R² = 0.88). The LSTM component was particularly valuable for picking up the delayed soil carbon-yield relationship that plays out gradually within a growing season. SHAP attribution placed FYM rate, SOC at sowing and cumulative growing-degree days as the top three yield drivers, together accounting for roughly 85% of the variance explained by the six leading features. Chemical nitrogen landed seventh-a signal that organic matter build-up, not synthetic N loading, is the more durable lever in these systems. A negative SHAP interaction between chemical P rate and AMF colonisation (Pearson r = -0.77) showed the model had independently recovered a well-established soil biology mechanism, lending credibility to its outputs. The limitations include data for training that is limited to one location only and a smaller plot scale sampling. It is recommended that the future research work include data related to satellite-based vegetation indices, genomics of microbial communities and a full rabi-kharif season.

The authors gratefully acknowledge the editorial board and reviewers of Legume Research-An International Journal for their valuable time, constructive comments and scholarly guidance, which substantially improved the quality of this manuscript. The authors also extend their sincere thanks to the field staff and laboratory personnel who assisted in data collection across both kharif seasons and to the farming community of the study site whose cooperation made the on-farm trials possible. The institutional support extended by the respective affiliations of all co-authors is gratefully acknowledged.

Disclaimers
 
The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided, but do not accept any liability for any direct or indirect losses resulting from the use of this content.
 
Informed consent
 
No animal is harmed.

The authors declare that there are no conflicts of interest regarding the publication of this article. No funding or sponsorship influenced the design of the study, data collection, analysis, decision to publish, or preparation of the manuscript.