Satellite-based Monitoring of Monsoon Crops in Smallholder Diversified Agriculture Systems of India: Emphasis on Soybean

Abstract

Timely and accurate monitoring of monsoon crops is essential for India’s food and nutritional security, as the Kharif season contributes more than 50% of the country’s annual food grain production and includes major staple crops such as rice, maize, pulses, and oilseeds. Yet, monitoring remains challenging in smallholder-dominated agricultural systems where persistent cloud cover, fragmented fields, and diverse cropping patterns limit traditional optical remote sensing approaches. This dissertation provides a comprehensive assessment of multi-source satellite observations, particularly Sentinel-1 SAR and Sentinel-2 optical data, for monitoring monsoon crops with focused attention on soybean, India's most important oilseed crop. The research progresses from understanding historical production patterns and current monitoring needs to developing and evaluating remote sensing methodologies suitable for operational deployment in India's complex agricultural landscapes. The research integrates spatio-temporal analysis with methodological advances for operational crop monitoring. Historical statistical analysis revealed major shifts in soybean production hotspots moving approximately 93 km southward and 24 km westward, with Maharashtra emerging alongside Madhya Pradesh as a co-dominant producer state. Despite significant area expansion since the 1970s, national average yield has stagnated at approximately 1 t/ha, well below global averages of 2 t/ha and demonstrated potential of 3 t/ha, while correlation analysis demonstrated strong linkages between soybean consumption and GDP per capita (r=0.96) and population growth (r=0.90), projecting rising future demand to approximately 5.7 MT soybean oil by 2050 that underscores the need for enhanced domestic soybean production. Building on these production dynamics, the dissertation develops a multi-layered remote sensing framework starting with broad agricultural landscape characterization before focusing on specific crop discrimination challenges. Multi-seasonal cropland mapping using a two-stage Random Forest classification of Sentinel-1 and Sentinel-2 time series achieved strong performance with producer's and user's accuracies exceeding 95% for Rabi and 90% for Zaid, with balanced accuracies of 75-80% for Kharif despite monsoon-season spectral confusion. The analysis revealed spatial patterns of agricultural intensification with cropping intensity of 1.6, where Kharif dominates (54% of gross cropped area) reflecting monsoon dependence; double-cropping prevails in central/western districts; and localized triple-cropping occurs in irrigation command areas. Novel time-integrated SAR phenological metrics were developed to capture seasonal crop growth patterns independent of fixed calendar dates. Comprehensive evaluation across five major monsoon crops demonstrated that intensity-based structural metrics during peak biomass periods outperform timing-based features, with cotton achieving strong separability (UA 0.62, PA 0.93, F1-score 0.75) with rice achieved strong cross-district transferability(mean OA: 74%), but systematic poor discrimination among cereal-legume crops (soybean, urad, maize) due to shared monsoon-aligned phenological patterns. These multi-crop phenological insights informed targeted soybean mapping experiments that systematically compared classification approaches and assessed operational scalability across different agro-climatic zones. Systematic comparison of classification approaches trained and tested with field data revealed that optical data significantly improved accuracy compared to SAR-only methods (optical-enhanced: OA 77-82% vs SAR-only: OA 62-66%), with October senescence-phase features most discriminative. Critically, computationally intensive dual-polarimetric decomposition and time-integrated phenological metrics provided no meaningful improvements over traditional time series approaches. State-level soybean area estimation demonstrated reasonable agreement with official statistics (6.48 Mha vs 5.96 Mha 5-year average, +9%), but district-level performance varied substantially (OA: 57-71%) due to spectral-temporal confusion with morphologically similar crops. Severe degradation in model performance when transferred between agro-climatic zones (cross-district accuracy: ~50-55%) exposed fundamental limitations of considered satellite systems and machine learning approaches in accommodating smallholder system variability. This research makes several key contributions: providing the first systematic evaluation of SAR and optical features for monsoon crop monitoring in Indian smallholder systems; introducing interpretable time-integrated phenological metrics; empirically quantifying cross-district transferability limits. By transparently documenting both successes and failures, this work provides essential guidance for developing next-generation monitoring systems. The findings demonstrate that while current satellite capabilities support approximate state-level assessments, robust operational systems may require integration of next-generation sensors, region-specific calibration, advanced machine learning with domain adaptation capabilities, and systematic ground reference datasets. The methodological framework developed provides an important foundation for scalable agricultural monitoring in India, contributing to improved food security planning and enhanced resilience of agricultural systems to climate variability.