Coconut: Agronomy, Sensing, and Applications

• Coconut is a tropical crop vital for rural livelihoods, characterized by high disease and pest risks that threaten yields and economies.
• Recent studies employ machine vision, transfer learning, and acoustic sensing to achieve rapid, accurate disease diagnosis and fruit maturity assessment.
• Coconut-derived materials show promise for hydrogen storage while public discourse on coconut oil highlights discrepancies from sustainability science.

Coconut, identified in the recent technical literature as Cocos nucifera, is a tropical crop system whose contemporary research profile extends well beyond conventional agronomy. Current work treats coconut simultaneously as a plantation species exposed to severe disease and pest pressure, a target for computer-vision and acoustic sensing, a source of biomass-derived carbon from its solid endosperm, and a commodity whose oil is discussed online through frames that often diverge from sustainability science [(Daga et al., 15 May 2025); (Vidhanaarachchi et al., 31 Jan 2025); (Dixit et al., 2014); (Candellone et al., 2023)].

1. Agronomic importance and exposure to loss

Coconut cultivation is described as underpinning rural livelihoods across the tropics, but the literature emphasizes that this dependence is coupled to substantial biological risk. Reported cases include destruction of 350 hectares in Côte d’Ivoire with an annual loss of 12,000 tons of copra, a cadang-cadang trajectory in the Philippines that could threaten 3 million palms by 2032 with losses of approximately \$300M, and yield reductions in India ranging from 10% to 80% due to Root Wilt Disease, Leaf Blight, Bud Rot, and Basal Stem Rot (Daga et al., 15 May 2025).

Sri Lankan studies add a second regional cluster of high-impact threats. Weligama Coconut Leaf Wilt Disease (WCWLD) and Coconut Caterpillar Infestation (CCI) are presented as major causes of production loss in Sri Lanka and nearby coconut-producing countries. In the WCWLD setting, contaminated palms in Matara are removed and burned, and research access is conducted under strict observation, indicating the operational seriousness attached to early diagnosis (Vidhanaarachchi et al., 31 Jan 2025).

A recurring theme across these studies is that coconut management remains constrained by diagnosis pipelines that are slow, expertise-dependent, and difficult to scale. Manual field inspection is repeatedly described as labor-intensive, subjective, and poorly suited to early intervention, particularly where age-old practices persist or expert labor is scarce (Daga et al., 15 May 2025, Vidhanaarachchi et al., 31 Jan 2025). This has made coconut a prominent case study for precision agriculture workflows based on transfer learning, instance segmentation, acoustic classification, and aerial detection.

2. Disease complexes and diagnostic ambiguity

Recent coconut pathology work distinguishes several distinct but operationally confusable visual syndromes. In the Indian image dataset used by DeepSeqCoco, five expert-validated classes were annotated: Gray Leaf Spot, Leaf Rot, Stem Bleeding, Bud Root Dropping, and Bud Rot, from a total of 5,798 images captured in daylight conditions in Kendur, Pune district (Daga et al., 15 May 2025). In Sri Lanka, WCWLD severity is modeled through three early-stage symptom classes—flaccidity, uneven yellowing, and tip browning—while CCI is characterized by dried brown patches and caterpillar galleries on leaflets (Vidhanaarachchi et al., 31 Jan 2025).

The literature stresses that diagnostic ambiguity is not incidental but structural. WCWLD uneven yellowing is reported as difficult to distinguish from magnesium deficiency yellowing even for trained personnel, while CCI can resemble Leaf Scorch Decline, a physiological disorder (Vidhanaarachchi et al., 31 Jan 2025). In the DeepSeqCoco experiments, the rare errors in the confusion matrices cluster in Leaf Rot and Gray Leaf Spot, which the paper treats as visually similar leaf-lesion classes (Daga et al., 15 May 2025). This suggests that coconut disease recognition is shaped not only by class imbalance and domain shift, but also by genuine symptom overlap at the phenotype level.

That overlap has methodological consequences. Because field symptoms can be subtle, early-stage, or lighting-dependent, coconut datasets increasingly emphasize expert validation, controlled preprocessing, and explicit augmentation. The DeepSeqCoco dataset was resized to 300×300 RGB and normalized to $[0,1]$, with random horizontal flips and $\pm 20\%$ rotations and zooms; the Sri Lankan datasets used resizing to 300×300, normalization by division by 255, and augmentations including rotation, filling, shearing, flipping, and zooming (Daga et al., 15 May 2025, Vidhanaarachchi et al., 31 Jan 2025).

3. Machine vision for disease detection and severity assessment

The most detailed recent image-classification study introduces DeepSeqCoco, an EfficientNet-B3-based transfer-learning pipeline for automatic coconut disease identification. Its end-to-end architecture is specified as input layer $(300\times300\times3)$, data augmentation, normalization, EfficientNet-B3 feature extraction, global average pooling, and a dense 5-way softmax classifier. The model contains 10,791,213 parameters, of which only 7,685 are trainable, reflecting a frozen backbone and a compact fine-tuning head (Daga et al., 15 May 2025).

On the five-class Indian dataset, DeepSeqCoco reaches 99.3% validation accuracy under both Adam and SGD single-optimizer runs, and the sequential hybrid schedule SGD→Adam yields the lowest reported validation loss at 2.81%, with training accuracy 99.5% and validation accuracy 99.3% (Daga et al., 15 May 2025). The same study reports prediction times of 1.61–1.72 s per image on Colab TPU, versus 13.47 s for ResNet50, 4.52 s for VGG16, and 8.62 s for MobileNetV2, and characterizes the system as achieving up to 18% reduction in training time and up to 85% reduction in prediction time relative to baselines (Daga et al., 15 May 2025). The paper’s interpretation is that EfficientNet-B3’s compound scaling and MBConv+SE blocks capture fine-grained lesions and stem exudates efficiently at 300×300 resolution.

A complementary Sri Lankan line of work decomposes the problem into disease detection, severity staging, infestation segmentation, and object counting. For WCWLD detection, DenseNet121 achieved 90% test accuracy with precision 71%, recall 100%, and F1-score 83%; for WCWLD severity, InceptionResNetV2 achieved 97% test accuracy with precision 99%, recall 99%, and F1-score 98% (Vidhanaarachchi et al., 31 Jan 2025). For CCI, Mask R-CNN with a ResNet101+FPN backbone achieved 92% precision, 89% recall, and mAP 95.26%; YOLOv5x achieved approximately 96.87% mAP and 97% correct counts against CRISL ground truth on a 100-image caterpillar-counting evaluation (Vidhanaarachchi et al., 31 Jan 2025).

The Sri Lankan study is notable because it moves beyond label assignment to progression quantification. CCI progression is computed after leaflet masking and cropping, followed by HSV color segmentation, replacement by canonical RGB values for healthy and necrotic regions, and $K=3$ K-means clustering; the progression level is then expressed as the ratio of brown pixels to total leaf pixels (Vidhanaarachchi et al., 31 Jan 2025). This places coconut pathology within a broader shift from classification to structured agronomic measurement.

Both image-based lines of work also identify generalization limits. DeepSeqCoco is trained on a single Indian region under daylight conditions, with no reported calibration, domain adaptation, or on-device TFLite deployment; the Sri Lankan datasets were collected in Matara, Puttalam, and Makandura, and future work is explicitly directed toward broader regional testing, including Kerala (Daga et al., 15 May 2025, Vidhanaarachchi et al., 31 Jan 2025). A plausible implication is that coconut disease models are already highly performant within curated domains, but cross-region transfer remains an open problem.

4. Post-harvest and landscape-scale sensing

Coconut sensing research also operates outside pathology. For post-harvest maturity assessment, one study reuses a dataset of 129 tall-variety coconuts—8 premature, 36 mature, and 85 overmature—from which three acoustic knocks per fruit yielded 387 original signals. After audiomentation and procedural audio generation, the class counts rose to 4,050 premature, 4,050 mature, and 5,850 overmature signals (Caladcad et al., 2024). The decisive preprocessing choice was “Summing + MFCC”; alternative combining methods and feature representations did not separate classes as effectively (Caladcad et al., 2024).

The resulting architecture applies two Conv1d layers, AvgPool1d(1), then either an RNN or an LSTM with 64 hidden units, followed by dropout 0.5 and three-class prediction. Both recurrent models achieved 97.42% accuracy, with average precision approximately 97.2%, average recall 97.32%, and no statistically significant difference between them ($p=0.9930$) (Caladcad et al., 2024). Misclassifications were concentrated in the mature class, while premature and overmature classes were recognized at very high rates. The study presents this as a route toward non-destructive, mechanized sorting.

At aerial scale, coconut has become a test case for low-annotation visual learning. The FLIM study on Tonga imagery uses 13,587 aerial regions of 90×90 pixels, with 10,268 labeled “coconut” and 3,319 “non-coconut,” and learns convolutional filters directly from user-drawn markers on only four representative images selected via t-SNE projection (Souza et al., 2020). With a single 60-filter, 7×7×3 convolutional layer followed by ReLU, max pooling, batch normalization, and an MLP classifier, FLIM+MLP achieved precision 0.863, recall 0.849, and F-score 0.854, slightly exceeding fine-tuned VGG-16 on F-score (0.851) while using only 0.07% of the parameters of the VGG feature extractor (Souza et al., 2020). The paper frames this as a gain in interpretability and annotation efficiency rather than sheer architectural depth.

Drone-based plantation inventory introduces a still broader spatial scale. A Ghanaian study used 73 DJI Mini 2 drone frames at 4000×2250 pixels, captured at 10 m, 25 m, 45 m, 70 m, and 85 m, and trained YOLOv7 primarily on synthetic images assembled from real palm cutouts placed on generated backgrounds (Rohe et al., 2024). By matching synthetic backgrounds to the real domain and adding negative classes—okra, weeds/grass patches, and tree trunks—the study increased [email protected] from an initial 0.65 to 0.88 (Rohe et al., 2024). The authors describe 70 m as the “most promising” altitude in their data, but only three images were available at that height, so the result is treated cautiously.

Taken together, these studies place coconut within a multi-scale sensing continuum: individual fruit maturity, leaf-level pathology, tree-level aerial classification, and plantation-level counting. This suggests that coconut agriculture is becoming a useful integration point for signal processing, human-in-the-loop representation learning, and synthetic-data-driven object detection.

5. Coconut-derived materials and hydrogen storage

Coconut is also studied as a materials precursor through its solid endosperm. Carbon derived from coconut solid endosperm (CDSE) was produced primarily by carbonization at 350 °C for 4 hours under flowing $\mathrm{N_2}$, without additional activation, and was found to retain native K, Na, Mg, and especially KCl within the carbon matrix (Dixit et al., 2014). XRD showed a broad carbon-related peak at $2\theta \approx 19.9^\circ$ and additional peaks indexed to cubic KCl with lattice parameter $a = 6.293\ \text{Å}$; SEM showed layered, porous macroporous morphology with uniformly distributed secondary phases, and TEM/EDX corroborated Mg and KCl inclusions (Dixit et al., 2014).

Hydrogen-storage measurements were performed with an automated four-channel Sieverts-type PCI system. The reported adsorption values are 1.10 wt.% at 30 atm and 2.30 wt.% at 70 atm at room temperature, and 3.50 wt.% at 30 atm and 8.00 wt.% at 70 atm at liquid nitrogen temperature (Dixit et al., 2014). Adsorption and desorption curves overlap without hysteresis, only $\mathrm{H_2}$ was detected in desorbed gas, and kinetics are described as fast and fully reversible (Dixit et al., 2014).

The study treats these results as atypical for a carbon whose BET surface area appears very low and non-representative because of macroporosity and embedded salts. Its mechanistic interpretation is twofold: KCl-induced polarization may raise the physisorption binding energy of $\mathrm{H_2}$, and native Mg may facilitate spillover onto adjacent carbon surfaces (Dixit et al., 2014). These mechanisms are presented as hypotheses rather than directly quantified effects. What is firmly established is that coconut-derived carbon can function as a hydrogen adsorbent without elaborate activation, and that embedded light elements inherited from the biomass are central to the interpretation.

6. Coconut oil in public discourse and sustainability debate

A separate research thread analyzes coconut through the discourse surrounding coconut oil on Twitter. In a multilingual corpus exceeding 20 million tweets, coconut oil accounted for 5,165,202 tweets between 21 March 2006 and 31 December 2021, placing it second only to olive oil at 6,543,149 and ahead of palm oil at 3,650,834, despite coconut’s lower contribution to global vegetable-oil supply than several other oils (Candellone et al., 2023). The study concludes that coconut discussion is dominated by health, beauty, food, and marketing frames rather than by production volume or ecological footprint (Candellone et al., 2023).

The temporal dynamics are distinctive. Coconut oil became the most tweeted oil monthly between 2014 and 2018, peaked in 2016, and then declined; unlike olive and palm oil, its growth was not significantly Granger-caused by the baseline growth of common-word tweets ($\pm 20\%$0, $\pm 20\%$1), which the authors interpret as evidence of burst-like, fad-like dynamics rather than platform-wide growth (Candellone et al., 2023). Sentiment is relatively stable over time at roughly 50% neutral, 12% negative, and approximately 38% positive, with a notable positive spike in March 2019 generated by a single joke tweet that received more than 86,000 retweets (Candellone et al., 2023).

The same study identifies a strong science–social media gap. Coconut discourse maps poorly onto sustainability and environmental concerns compared with palm oil, even though the study notes that scientific work it cites indicates that coconut oil may threaten more species than other oils (Candellone et al., 2023). Instead, coconut’s dominant hashtags occupy virality regimes associated with promotional bursts and lifestyle content rather than the negative environmental framing that drives larger cascades in palm-oil debates (Candellone et al., 2023).

This directly addresses a common misconception: online salience around coconut oil does not track sustainability relevance. The available evidence instead indicates that coconut’s public narrative is health/beauty-first, while biodiversity and land-use concerns remain weakly represented in the dominant social-media frames (Candellone et al., 2023).

Coconut research therefore spans a notably wide technical range. In agriculture, it is a high-value tropical crop with documented exposure to disease, pest, and classification bottlenecks; in sensing, it supports transfer-learning, recurrent, remote-sensing, and synthetic-data pipelines; in materials science, its solid endosperm is a bio-precursor for hydrogen-storage carbons; and in public discourse, coconut oil functions as a case in which lifestyle and marketing narratives can dominate over ecological assessment [(Daga et al., 15 May 2025); (Caladcad et al., 2024); (Rohe et al., 2024); (Dixit et al., 2014); (Candellone et al., 2023)].