Coati-optimized Google Earth Engine framework improves crop prediction in India

In the paddy-rich heart of Krishna district on India’s southeastern coast, an interdisciplinary team of researchers has built a satellite-driven forecasting system that can predict crops with nearly 98 percent accuracy, while simultaneously mapping the agricultural zones most vulnerable to climate stress. The study, led by S. Rohini of Annamacharya Institute of Technology and Sciences together with S. Narayana Reddy and D. Vivekananda Reddy of Sri Venkateswara University College of Engineering in Tirupati, has been published in Theoretical and Applied Climatology. Its central innovation is the Coati-based Recurrent Crop Prediction (CbRCP) model, a machine learning framework that fuses more than a decade of Landsat imagery with ground-based climatic records and then fine-tunes itself through a bio-inspired optimization algorithm modeled on the foraging behavior of the coati, a nimble mammal native to the Americas.

The research addresses one of the most stubborn problems in precision agriculture: how to make reliable, district-scale crop predictions in regions where fields are fragmented, weather is erratic, and ground surveys are slow and expensive. Krishna district, an agriculturally significant delta region in Andhra Pradesh, exemplifies the challenge. Its rice paddies, cotton fields and mixed croplands shift from season to season, and increasingly erratic monsoon rainfall and heat stress have made traditional planning methods risky. The researchers wanted a system that could not only identify what is being grown and where, but also flag which areas face the greatest climatic threats, giving planners a quantitative basis for climate-resilient decision-making.

At the foundation of the work is Google Earth Engine, Google’s cloud-based geospatial processing platform, which allows scientists to run computations over enormous archives of satellite imagery without downloading a single pixel to a local machine. The team used Landsat data covering the period from 2011 to 2022, and paired it with climatic records supplied by the Andhra Pradesh Development Planning Society (APSDPS), including temperature, rainfall and humidity measurements. By processing the two data streams together, the researchers could track spatio-temporal variations in vegetation across the district, capturing both the seasonal rhythm of cropping cycles and the longer-term effects of a changing climate.

The spectral heart of the framework lies in three vegetation indices, each computed directly from satellite-measured reflectance. The Normalised Difference Vegetation Index, or NDVI, exploits the fact that healthy chlorophyll-rich vegetation strongly absorbs red light while reflecting near-infrared radiation, producing a value that rises and falls with biomass and plant vigor. The Soil Adjusted Vegetation Index, SAVI, adds a correction factor that reduces the distorting influence of exposed soil brightness, a critical refinement in semi-arid and mixed agricultural landscapes where bare earth contaminates the signal. The Visible Atmospherically Resistant Index, VARI, works exclusively within the visible spectrum and applies an atmospheric correction, making it robust against haze and aerosol scattering. By combining these indices with the climatic variables, the model gains a multidimensional fingerprint for every field: how green it is, how that greenness evolves through the season, and what weather it endured along the way.

What elevates the study above routine crop classification is the optimization layer. The CbRCP model is a recurrent architecture, meaning it maintains memory across time steps, an essential property when classifying crops whose spectral signatures overlap at certain growth stages but diverge across a full phenological sequence. Yet even the best recurrent networks depend on correctly tuned hyperparameters and feature weightings, and naive tuning often traps models in suboptimal configurations. The researchers turned to the Coati Optimization Algorithm, a nature-inspired metaheuristic that mimics the hunting and foraging strategies of coatis, which sweep through terrain in coordinated patterns, balancing exploration of new areas with exploitation of known food sources. In computational terms, this translates to a search process that iteratively refines the model’s parameters, escaping local optima that simpler gradient-based or grid-search approaches might miss.

The results are striking. Evaluated against a battery of established methods, including the Honey Badger Algorithm (HBA), the 3D-UNet segmentation network, AGLM, ASM and Mobile UNet, the CbRCP model achieved a prediction accuracy of 97.87 percent, precision of 97.96 percent, recall of 97.87 percent, and a Kappa coefficient of 0.9575. The Kappa statistic is particularly telling: it measures agreement between predicted and actual classifications while correcting for chance agreement, and values above 0.8 are conventionally considered to indicate almost perfect concordance. A Kappa of 0.9575 means the model’s classifications are far better than random, and the margins over the comparative methods suggest that the coati-driven optimization genuinely improved the recurrent model’s ability to separate spectrally similar crop types.

Beyond raw accuracy, the framework delivered something arguably more valuable for policymakers: a vulnerability map. When the team conducted spatial analysis across the district’s administrative units, they identified Gudivada, Machilipatnam and Vuyyuru as high-risk agricultural zones. These areas, the study found, were more susceptible to rainfall irregularities and temperature stress than neighboring regions. Machilipatnam, a coastal town exposed to the vagaries of cyclonic weather and saline intrusion, and the inland agricultural centers of Gudivada and Vuyyuru, both historically significant rice-producing areas, now carry a data-backed warning label. For district agricultural officers deciding where to prioritize irrigation infrastructure, drought-tolerant seed varieties or crop insurance outreach, such spatially explicit risk information transforms abstract climate anxiety into actionable geography.

The technical pipeline also illustrates how modern cloud computing has democratized large-scale remote sensing. A decade ago, processing twelve years of Landsat scenes over an entire district would have required substantial local storage, significant computing power and considerable expertise in image preprocessing, including cloud masking, atmospheric correction and mosaic assembly. Google Earth Engine handles much of that automatically, and the study demonstrates that feature extraction and time-series analysis that once demanded dedicated supercomputing can now be scripted and scaled. The authors argue that this scalability is precisely what makes the framework suitable for replication across diverse agro-ecological environments, from the deltas of Andhra Pradesh to other monsoon-dependent agricultural regions facing similar climatic volatility.

The implications extend well beyond a single district. Accurate, timely crop prediction underpins food security planning, market forecasting, insurance design and the allocation of subsidies in a country where agriculture remains the livelihood of hundreds of millions of people. As climate change intensifies the frequency of droughts, unseasonal rains and heat waves, the gap between planting decisions made on tradition and those informed by data becomes a matter of economic survival. A framework that integrates remote sensing, climatic information and optimization-based learning, the authors note, provides a reliable and scalable pathway toward climate-adaptive decision-making and sustainable agricultural management. In Krishna district, the coati, an animal that survives by foraging intelligently across uncertain terrain, has lent its name to a tool designed to help farmers do much the same.