Severe decline in large farmland trees in India over the past decade

The study addresses the under-monitored but vital component of India’s landscapes: trees outside forests embedded within agricultural systems. Although India’s farmlands cover ~56% of the country while forests comprise ~20%, a substantial number of trees scattered in croplands, hedgerows, and urban areas are excluded from conventional forest monitoring. Existing satellite-based deforestation and forest dieback systems do not capture the dynamics of these non-forest trees. Reports of diseases affecting mature on-farm trees (for example, fungal infections in neem) and the potential impacts of climate change and changing agricultural practices raise concern, yet robust, repeated inventories at national scale have been lacking. The research question is whether large farmland trees in India have declined over the past decade and, if so, by how much and where, and what factors might explain observed losses. The purpose is to provide a national, tree-level monitoring approach and quantify the disappearance of large, mature farmland trees, thereby informing climate, biodiversity, and livelihood strategies that depend on agroforestry.

The paper situates its work within literature documenting widespread tree mortality in forests driven by drought, heat, wildfires, insects, pathogens, and management, with operational satellite systems enabling timely monitoring of forests. In contrast, trees outside forests—especially agroforestry trees—are poorly quantified at scale, despite recognized roles in climate mitigation, adaptation, biodiversity, and livelihoods. Prior mapping efforts often capture block plantations rather than individual on-farm trees. India is a global leader in agroforestry implementation, yet systematic, repeated, national-scale inventories of individual farmland trees are scarce. Neem (Azadirachta indica) has been highlighted in the literature and media for disease-related diebacks (for example, Phomopsis azadirachtae). Advances in high-resolution satellite constellations (RapidEye, PlanetScope) and deep learning have enabled tree-level detection, counting, and segmentation across large areas, but time-series analyses at national scale for non-forest trees have remained limited. The study builds on previous deep-learning and heatmap-based detection methods to overcome these gaps.

Study area and scope: India. The analysis targets individual non-forest trees within farmlands (and for 2018–2022 also urban and bare classes), excluding dense block plantations and closed canopy groups typically classified as forest in baseline land-cover maps. Imagery and periods: RapidEye mosaics at 5 m resolution for 2010 and 2011; PlanetScope mosaics at ~3–4 m resolution for 2018, 2019, 2020, 2021, and 2022. The 2010/2011 epoch enables detection of mainly larger trees (>~100 m² crown for clear visibility at 5 m pixels), while 2018–2022 allows more consistent, annual, wall-to-wall detection of larger trees (though small/young trees remain under-detected). Image preparation: Custom 1°×1° mosaics were built using strict quality filters. Phenology-based acquisition windows (MODIS phenology) targeted periods when herbaceous vegetation senesces but trees remain green (for deciduous areas: senescence to mid-greendown; for evergreen areas: mid-greendown to dormancy). Scenes had low sun elevation (<50°), were cloud-free (RapidEye clouds filtered via blue-band standard deviation; Planet cloud-free selection), and met sharpness constraints using a blur kernel threshold of 0.23; unsharp scenes were discarded and replaced. Histogram matching to Landsat/Sentinel-2 harmonized colors. PlanetScope images retained had GSD close to 3.1 m; scenes >4 m GSD were excluded. Final mosaics underwent CLAHE and normalization. Training data and models: Approximately 100,000 point labels for RapidEye and 130,000 for PlanetScope were created at tree crown centers, verified against high-resolution Google Earth/Bing imagery, iteratively improved to cover diverse landscapes and image conditions. A heatmap-based detection method using a UNet with a ResNet-50 encoder was trained to predict (1) a heatmap of tree-center confidence and (2) a scale map enabling adaptive Gaussian kernel resizing. Modifications included exact on-the-fly Gaussian rendering, minimum Gaussian σ (4.5 m PlanetScope, 5 m RapidEye), base σ values (6 m PlanetScope, 6.7 m RapidEye), and scale factors >1 to improve smoothness and separation of adjacent crowns. Training used 2,500 epochs, LR 1e-5 until epoch 2000 with linear decay, five-model ensembles with balanced 80/20 train/validation splits minimizing distributional differences (Wasserstein distance). Heatmap F1 scores ranged ~0.63–0.65 (20 m radius) and ~0.67–0.69 (50 m radius; precision 0.83, recall 0.61). Local maxima in ensemble-averaged heatmaps were extracted as points with associated detection confidence. Detection thresholds and segmentation: Points with confidence ≥0.35 were considered detections. A crown segmentation model (updated with ~52,000 Indian labels) was applied to 2021 imagery to derive crown-size distributions (used for characterization, not change detection). Segmentation indicates under-detection below 20 m² and identifies ~22.7 million trees >100 m², matching RapidEye high-confidence large-tree counts (~22 million). Sampled crown areas: in 2019, average crown size was ~55 m² (PlanetScope) and ~62 m² (RapidEye); high-confidence thresholds correlated with larger crowns (RapidEye >0.8 → ~96 m²; PlanetScope >0.7 → ~67 m²). Land cover masking: For 2010–2018 comparisons, cropland from ESA WorldCover 2020; for 2018–2022, cropland, urban, and bare classes. This excludes dense plantations and forests; trees in fallow fields, hedgerows, and along roads/rivers are included. Change analysis: Trees detected in 2010/2011 (RapidEye) were tracked to 2018–2022 (PlanetScope) using a 15 m buffer around crown centers to account for geolocation differences; absence in all later years classified as disappeared; presence in any later year classified as remained. For 2018–2022, a change-confidence metric integrated multi-year detection confidences: a tree detected in 2018 and/or 2019 with high confidence and absent in three consecutive years (2020–2022) was labeled disappeared. Low-confidence trees (detected sporadically with low/medium confidence) were excluded from reported losses. Only high change-confidence (>0.7) losses were reported to be conservative. Quality controls and uncertainty: Scene-wise RapidEye geolocation shifts were detected by footprint-level disappearance rates; scenes with >40% disappearances or visible shifts were masked. Image-quality temporal degradation was controlled by computing per-year average detection confidence within 1×1 km cells; cells with strong negative slopes (<−0.05) between 2018–2019 and 2020–2022 were flagged and excluded from loss counts. Evaluation via manual checks: 1,000 random 2010/2011 detections (false positives ~3%); 1,000 random 2010–2018 disappeared trees (false loss ~2%); 1,000 random 2018–2022 disappeared trees (false loss ~21%). Interviews (n=12; multiple states) provided qualitative context on drivers of change. Scope and caveats: The framework targets detection of the disappearance of large trees and does not provide wall-to-wall mapping of all trees for 2010/2011 or robust detection of small/young trees; thus, gains and net change are not estimated. Reliance on WorldCover 2020 introduces land-cover classification uncertainties, and some large farmland trees may be misclassified as forest or plantations.

- Large-tree losses 2010–2018: Approximately 11 ± 2% of high-confidence large trees detected in 2010/2011 (average crown ~96 m²) were absent in 2018–2022. Loss hotspots include central India (notably Telangana and Maharashtra), with local losses up to ~50% and up to ~22 trees per km² disappearing. - National inventory 2018–2022: About 597,638,431 trees were mapped across cropland, urban, and bare classes. Mean density ~0.6 trees per hectare (s.d. 1.6), with highest densities up to ~22 trees per hectare in regions such as Rajasthan and Chhattisgarh. - Losses 2018–2022: An estimated 5.3 million trees (2.7 trees per km²) that were present in 2018/2019 were not detected in 2020–2022 at high change confidence (>0.7). Figure 3 captions also report an estimate of 5.6 million disappeared trees. Some regions lost >50 trees per km². Disappearing trees in this class have average crown sizes ~67 m², indicating mature trees. - Temporal concentration: A majority of the 2018–2022 losses likely occurred between 2018 and 2020. - Drivers: Climatic analyses show temperature increases and variable rainfall/drought trends, but the decade’s precipitation was above long-term mean in most areas, offering limited support for climate as the primary driver. Qualitative interviews across several states consistently reported removal of mature on-farm trees due to intensification and expansion of paddy cultivation facilitated by new boreholes/irrigation; shade from large canopies was perceived to reduce crop yields. Interviewees did not attribute recent losses to fungi or climate. - Methodological contribution: A scalable, deep-learning, heatmap-based approach enables annual monitoring of individual large trees at sub-continental scale with open data outputs and quantified change confidence.

By creating annual, high-resolution, tree-level inventories, the study directly addresses the lack of monitoring for trees outside forests in India. The findings show a substantial decline in large, mature farmland trees in multiple regions over a decade and accelerated losses after 2018, indicating that agroforestry components critical to climate mitigation, adaptation, biodiversity, and livelihoods are being reduced. The evidence suggests agricultural intensification—particularly expansion of irrigated paddy—has driven the removal of large trees within fields. Climatic trends alone do not explain the observed patterns, and interviews corroborate management-driven removal as a dominant factor. These results contextualize apparent discrepancies with official reports of net increases in tree cover: the study reports gross losses of large on-farm trees, excludes block plantations, and does not assess gains, so it does not contradict net-increase findings. Nonetheless, the loss of large trees likely diminishes ecological functions and long-term socioecological benefits. The provided dataset and methodology can inform policy and management by identifying hotspots of loss, guiding protection of mature on-farm trees, and supporting monitoring of restoration, illegal logging, and tree health at scale.

The study introduces a robust, scalable deep-learning framework that maps and tracks individual large farmland trees annually across India, generating an open dataset with per-year detection confidences and a change-confidence metric. Applying this tool reveals significant gross losses of mature on-farm trees: ~11% decline from 2010/2011 to 2018–2022 among large trees, and 5.3–5.6 million additional losses between 2018/2019 and 2020–2022, with pronounced hotspots. The results underscore the vulnerability of mature agroforestry trees to shifts in cultivation practices and agricultural intensification, raising concerns for climate, biodiversity, and livelihoods. Future research should: integrate spectral/temporal health indicators; refine land-cover masks and incorporate dynamic land-use histories; expand to global scales; assess net change by combining gains and losses; improve small-tree detection; and conduct field validation to quantify species-specific drivers and socioeconomic impacts. Policymakers and practitioners can leverage these data to prioritize conservation of mature on-farm trees and design incentives compatible with yield goals.

- Detection bias toward large trees: Small/young trees (<~10–20 m² crown; <~4 m height) are under-detected, especially pre-2018, limiting comprehensive counts and gains assessment. - No net change: The study emphasizes gross losses; gains and regeneration are not quantified; block plantations are masked out. - 2010/2011 mapping not wall-to-wall: RapidEye heterogeneity and conservative labeling reduce false positives but miss some trees, especially smaller ones. - Image quality and geolocation: Variable sharpness across PlanetScope scenes and RapidEye geolocation shifts can cause missed detections or false losses; mitigated but not eliminated by quality filters and scene masking. - Land-cover dependence: Reliance on ESA WorldCover 2020 may misclassify some large farmland trees as forest or plantations and does not capture land-use changes prior to 2020. - Uncertainty in change detection: Despite the change-confidence metric and additional quality checks, false loss rates remain (estimated ~2% for 2010–2018 and ~21% for 2018–2022 at chosen thresholds). - Qualitative drivers: Interview-based attribution to agricultural practices is informative but not statistically representative or causally conclusive.